# My EA Blog

> Personal essays, drafts, cross-posts, and bullet-point notes by Alejandro Acelas (https://myea.blog). This file contains the full text of every post, newest first.

---

# My Claude Code spiel to non-technical EAs

Source: https://myea.blog/claude-code-spiel

*Getting started with Claude Code when you're not technical*

[Alejo Acelas](mailto:alejoacelas@gmail.com). Jun 4, 2026\. [alejoacelas.com](http://alejoacelas.com)

\[What follows is mostly verbatim from an improvised call with Jeremiah, who had tried Claude Code exactly once and felt frustrated when it broke. It has a different format to my usual coaching calls which are more customized to a person's workflow, but hopefully this will be more widley useful as well.

Claude helped me cut the small talk and added section headers; the rest is what I said on the call.\]

## You don't need to learn how to code

**Jeremiah**
That's one of the things that put me off ever going to Code. I thought: Claude Code — I'm going to need to learn how to code, at least a bit. So it's quite a relief that you just speak to it in natural language and it does stuff.

**Alejo**
Exactly that. You speak to it in natural language and it does stuff.

## Why Claude Code and not Cowork

The main case is it gives you more affordances to do stuff.

Cowork is very centrally managed by Anthropic: you can mostly only use Cowork from the Claude desktop app, and they tell you "you can connect to this data service but not this one, you can run this script but not that script". They're opinionated, and they do that for two reasons:

* to protect you from shooting yourself in the foot security-wise
* to make the onboarding experience easier.

On the shooting-yourself-in-the-foot side, I think they've made many reasonable decisions — things that should be most people's defaults. But they're making decisions that are safe for all knowledge-work professionals, and you might want to make different trade-offs than the minimum common denominator of all humanity.

And the deeper reason why everyone should switch to Claude Code is that this is what the Anthropic engineers use. They're trying to automate themselves into creating God, and so you should basically use the tools that they use to create God, because it's the thing they put the most effort into.

## Expect the first day to be Linux, not Windows

**Jeremiah** (on his first attempt at Claude Code):
“the Chrome extension isn't connected", "the sheet requires authentication, you now need to download this and save it for me separately". Is that just the first time?

**Alejo**
Exactly. The thing about Cowork is they want you to have a good out-of-the-box experience — you just click and it works. The thing about Claude Code is it's made for hackers. It's the same as Linux. You can think of Cowork as your Windows: you power it on and there's the intro screen, your file manager, whatever.

Linux is a free-for-all: you can absolutely customize it to fulfill your deepest desires if you're into having desires that computers can fulfill, but it also just breaks, and it exposes a lot of things — some versions of Linux don't even come with a graphical interface, just a text prompt. That's what happens with Claude Code: you need to set up some things the first time.

## Worry less about security

One piece of advice I give to people who want to take the most advantage of AI is to worry less about security. There's a bunch of security-versus-convenience trade-offs, and for me the key to using Claude Code and getting a bunch of value out of it is going all the way to the convenience side: I want it to be almost instinctive that I can just open Claude Code, throw in what's on my brain, and get some valuable output at the end.

What you could reasonably be scared of is a few things. One is taking irrecoverable action on some important asset of yours — deleting something on your Drive, messing up your computer. I think with Claude Code, unless you ask it to do stupid things — "please install this program at all costs" — where you're actually inviting harm in, failures are reasonably rare, and when you find a failure it's often recoverable.

But for some things you might want to screen off the possibility entirely. If you connect it to your Drive, you can ask it: "Can you delete files? Make sure you cannot delete files" — and there are tool configurations such that it can never delete files, or Drive already has snapshots. For anything you fear losing — if you fear it's going to screw up your Mac — ask Claude Code: "What are some ways that might happen?" It gives you a couple of ways, and then you ask: "Are there passive, zero-maintenance, zero-attention things we can set up so that won't happen?"

The other thing is prompt injection risk — a malicious page hijacking the agent to do something terrible on your computer. It has been well documented in the academic literature, but it's not actually a threat in practice so far. So I think you can for the most part ignore that kind of risk, even though it's been widely discussed in many places.

Most of what can happen is like lending your computer to an over-enthusiastic friend. If you gave your computer to me, I might try to do stuff and do something crazy. These agents are often too goal-oriented: if you give them an overly broad goal — "just update my Mac to the next version, do it, and don't check with me on the plan, just do it" — there are many ways of achieving that, and some of them have substantial risks. So giving it goals with an overly broad scope puts you at some risk. But for many of the things you do — adding a file, editing a spreadsheet — when was the last time you blew up your computer editing a spreadsheet? I hope never.

## Feed it text

Everything you put on your computer is now in the ownership of Claude Code — if it's on your computer, Claude Code can access it. And there are some file types, like XLSX, that are just hard to process, because agents read text. They cannot read the raw file of an XLSX — if you open that in your text editor it's just garbage. But a CSV — you open that in a text editor and you can read it. Claude can manipulate XLSX files, but it's a little bit harder and sometimes it screws up a bit more. So you ideally want something that's text-in, text-out, and to think carefully about what's the text you're feeding in and the text that's coming out.

So if a connector is giving you trouble — the spreadsheet requires authentication, the Chrome extension isn't connected — often the simplest move is: download the thing as a text file (a CSV, not an XLSX) and drop it in the folder.

## It will lie to you

It will lie to you. So you should find ways to actively verify what it does. If it created a file: "okay, show me the file". If it said "I downloaded this" — you need to get an intuition for what are the things it can and cannot do.

I often start by dictating the prompt. (On dictating: you should dictate, you should never type — at least for interacting with AIs.) You start by dictating the goal you have, and for the things where you have an intuition they might or might not work, you say — literally the same way you'd say it to a person — "hey, I suspect we might not be able to connect to this data source; I don't know if you can access Substack; here's a spreadsheet, I've never done spreadsheets with you — just pull the first rows and open the file at the end so I can see you got the data right." You don't want it to report on whether it was able to do the thing — you want to find the small test for whether it was able to do the thing.

## Be curious when it fails

**Jeremiah** (after Claude Code failed to load a Google Sheets and connect to Swapcard)

"well, why don't you unblock web search and just do the task like I told you to?"

**Alejo**
You're doing it wrong — because you're not being sufficiently curious. Andrej Karpathy — whose Twitter you should obviously follow — had this really nice quote: "you can delegate your thinking but you cannot delegate your understanding". If a part is critical for getting the final output working — the thing being able to do web search, the thing being able to access external data sources. Getting annoyed is trying to delegate again to the AI after it failed.

What I would do is: "okay, it seems to me that you didn't manage to connect properly to this — why was that?" Then I read the explanation and go back and make sure it matches what I saw. Then: "okay, sounds like that's the explanation — what are some ways of fixing it?" And then: "okay, let's try them and test them."

Once you've got it working the first time, you don't need to worry about that anymore — you set it up working. And then maybe: "add this to your global instructions" — that's a move I do for things that aren't obvious to set up. For example, for interacting with Google Docs, I have a special tool, and I said "always use this tool when interacting with Google Docs".

## Ask Claude first

All of these are things you can ask Claude, and you should have the impulse to just ask it first. For many things you realize very quickly — oh yeah, that makes sense. When I've already worked a lot on a piece and just say "improve my writing", it often suggests absolute shit, but at least I found out quickly that it was bad and can decide whether to try a more specific prompt or simply not use it for that purpose.

You should just start by asking the machine, see if you get a good response, and either be merry because it made it or think whether it's worth spending the work of doing a better prompt.

## Don't worry about tokens and rate limits

My approach to token use just buy as much as you need, because your productivity is important. If you're hitting rate limits ever you should basically buy the $100 or $200 dollar plan, and buy more plans if you hit the rate limits again.

You should use as many tokens as you want, but you should also be careful about keeping good performance from your AIs.

One thing that's not obvious to people who haven't read machine learning papers: when you have long conversations, the quality — for structural reasons — is expected to get worse the longer the conversation goes. In Claude Code you sometimes see a little thing at the bottom that says "using 300,000 tokens out of a million". If you're past 200,000 tokens, you can start thinking about how to delegate to another session. Oftentimes I ask it to “write down the things we've agreed upon and then give me a prompt to feed to another agent". But I do that not for cost reasons — again, I use as much AI as my work demands — I do it for performance reasons.

## Everything is just a file on your computer

Cowork kind of hides the files it creates sometimes — hides them in a virtual machine on your computer. When you use Claude Code, it doesn't do the hiding-files thing: it's just a folder on your computer, and everything you've created is just a file on your computer — including the conversation history, which is deleted after 30 days by default; and you can just say "hey Claude, can you change that? I want to keep all the history of my conversations."

This also makes Claude Code a future-proof move. Jeremiah was doing all his AI work inside an organization's subscription he'd eventually have to leave. With Claude Code, if you change subscription, you're basically just saying "instead of charging this account, charge this other account" — all the files on your computer stay there.

## The starter pack

Here's the starter pack:

- [Talk to your computer](https://wisprflow.ai/) [Wispr Flow](https://wisprflow.ai/)
- [Connect to Google Docs](https://github.com/LucaDeLeo/gdoc)
- [Connect to everything else Google](https://gogcli.sh/)
- [The best interface for Claude Code](https://super.engineering/)
  - Use it instead of the Claude desktop app.
  - It has a lot of shortcuts for moving across projects, it makes it easier to have different versions of the same thing, it makes it easier to see the files you created.
  - If I were using the Claude desktop app, I wouldn't have started this many projects at all — this makes it really easy to start new things, and I get notified when a thing finishes. A lot of design decisions made for AI users

That setup should take you maybe about an hour if you're reasonably computer-savvy — clicking the right buttons at the right time — but maybe up to two hours. If you do that, you'll be in a good position to just ask Claude for stuff.

---

# Silly games Tomáš taught me

Source: https://myea.blog/silly-games

The German Summer Camp was lots of fun, and one of my personal highlights was Tomáš’ silly games workshop. Here's some of the games we played there, mixed in with some games I personally like.

These games are best for groups of 8-20 people. Paste this document to your LLM of choice if you need more detailed instructions.

Poses warmup

* Stand in a circle. One of you says their name and does a pose/motion
* Once they finish, everyone repeats your name and pose
* The person to their left repeats the routine: name, motion, then everyone joins
* Do a first loop at a leisurely pace. Second round should be as fast as possible

Honey, do you love me?

* Stand in a circle. One of you locks eyes with someone else from the circle and approaches them
* As you get close to them, say the words “honey, do you love me?”
* The other person replies “yes, honey, I love you, but I just can't smile”
* If they smile as they say this, they're out. If they don't, you're out. The survivor chooses a new victim and the process repeats until only one person remains

Wooosh

* Stand in a circle, one of you holds an imaginary ball of energy
* You can send the ball to the person next to you by saying ‘wooosh’ and pushing it in their direction. Now it's their turn to wooosh
* Once you've mastered your wooosh, add other movements:
  * boing: bounce up and down and you knock your head
  * aaaaaaah: stop passing the ball and scream, everyone runs around until the initial screamer settles which is your cue to remake a circle
  * …
  * …

Puking kangaroo

* Like wooosh, but each pose involves three people
* Stand in a circle. One of you points at someone else to make them the lead. The two people next to them become their sidekicks
* As you point at them, shout one of the following:
  * Puking kangaroo: sidekick arms stretched at should height, lead steps back to puke over the bag-shaped space created by sidekicks
  * Washing machine: sidekicks join arms in a circle, lead crouches and does a circular motion with their head
  * James Bond: lead holds a finger gun to their chest, sidekicks swoon with sexy delight
  * Toaster: sidekicks extend arms straight to surround the lead, the lead jumps
  * Blender: sidekicks extend one arm up and one down to form a vase shape, lead spins in the spot

Ninja war

* Stand in a circle and hold one hand to the center
* At the count of three, everyone jumps apart to their standing position. No one can move until it's their turn
* On your turn, you're allowed one continuous movement. If, through that movement, you touch someone's arm(s) or leg(s), they're not allowed to use them for the rest of the game
* If you expect the attacker to touch you, you're allowed one continuous movement to get out of the way. You can only move after the attacker moves, otherwise you must return to your original position
* After the confrontation ends, it's the next person's turn. They don't need to wait for permission
* Make sure to always do turns in the same order and not skip anyone. The order can come from from your position in the initial circle, from the initial player then clockwise

Assasin

* The policeman goes away. Someone volunters to murder
* Everybody walks around, with the assasin winking at people to kill them
* The policeman gets three guesses to spot the assasin, if everyone's dead before that, they win

Guess the lead

* Like Assasin but instead of winking the game lead makes a dance/gesture that everyone else has to copy as soon as they notice
* The policeman attempts to discover the lead in 3 guesses

Evolution

* You start life as a worm. Shake your body in worm-like fashion
* Look for someone who's the same animal as you. Challenge them to rock-paper-scissors. The winner upgrades to a better animal, the loser downgrades or remains worm if they're already at the bottom
* The animal hierarchy is:
  * Worm
  * Chicken
  * Pinguin
  * Elephant
  * Monkey
  * Human
  * Superintelligence
* Remember to imitate the sound and movements of your animal
* The game continues until superintelligence arrives

Giving circle:

* Stand in a circle. Someone, without words, hands the person to their right an imaginary gift
* Through pose and mime they attempt to convey the gift's content. The receiver then states the gift content with gratitude “oh, thank you for this \[baby crocodile\]\!”. The giver can reply if they like “You're welcome\! I caught it myself”
* Go around the circle until you make a full round

# Card games

### Standard Deck Games

Get a standard poker deck with a box (to make it easy to carry around) and learn the rules of these games:

* **Bullshit (Cheat)** – Take turns playing cards face-down, claiming a rank. Ranks must follow in sequence. Anyone can call "Bullshit\!" If the accusation is correct, the liar picks up the pile; if wrong, the accuser picks it up. First to empty their hand wins.
  → [Rules (Pagat.com)](https://www.pagat.com/beating/cheat.html)

* **Indian Poker (Blind Man's Bluff) – Penalty Version** – Each player sticks one card to their forehead facing out (you see everyone's card except your own). Going around the table, decide to stay in or fold. Among those who stayed in, the lowest card gets a *penitencia* (drinking a bit of ketchup, maybe?). Fold too often and you also get one.
  → [Rules (Pagat.com)](https://www.pagat.com/poker/variants/indian.html)

* **Assassin (Wink Murder)** – Deal cards; whoever gets the Joker (or Ace of Spades) is the secret Assassin. The Assassin kills players by discreetly winking at them—victims dramatically "die" after a few seconds. I play it Mafia-style, with a Detective (who was *n* attempts to identify the murderer), a Prostitute (that can rise the dead), and a Grave-digger (who can foreclose the prostitutes powers, making sure you stay dead for good)
  → [Rules (Wikipedia)](https://en.wikipedia.org/wiki/Wink_murder)

* **Skull** – Pure bluffing. Each player has 3 roses (red cards) \+ 1 skull (black card). Place cards face-down, then bid how many you can flip without hitting a skull—must flip your own stack first. Win two challenges to win. You can buy it as well.
  → [Rules (Wikipedia)](https://en.wikipedia.org/wiki/Skull_\(card_game\))

* **Any other game you've enjoyed** – Even complex games work when you're enthusiastic. People pick up on your energy and join in.

---

### Out-of-the-Box Games

* [**Exploding Kittens (Party Pack)**](https://www.amazon.com/dp/B07CTXHNSL) – Russian roulette with cats. Draw cards, avoid exploding, sabotage friends. The Party Pack (and some other extensions) supports up to 10 players—get this version unless you'll already know who you'll be playing with and it's a small group

* [**Jungle Speed**](https://www.amazon.com/dp/B09TBJGLJ1) or [**Anomia**](https://www.amazon.com/dp/B00363B5FC) – Fast-paced pattern/word-matching games. Incredibly easy to explain—I often just start playing and people pick up the rules as we go.

* [**UNO**](https://www.amazon.com/dp/B00GTJSGXE) – A classic for a reason.

* [Codenames](https://www.amazon.com/CGE-Codenames-Association-Friends-Players/dp/B0FJMQB63N/ref=sr_1_1?crid=K5MIS019KI1N&dib=eyJ2IjoiMSJ9.oHvvl7Nedz_PjxKHSMFAsqx9tfhe_kd8PqYn7waozgsAwbLu5V6u3SCmaEA0VQxyPduQmH49KlqKuDl3whGu7f6VZ0wSUCz5U6-TCuXJSOBuU-R_1QrcjEerX2xwhfhNgMsetrx0PGY6rz5vXUc_ZSoJoSmZmSF32GRp6UgqnXiXlDRtO9Qjirks4GYWjaGjh3grxYIO7MR4U_5IDMr48lWSCx5y-jpV8c3uKx0w---bigfEUbbvxJfYEFcdNDu0REzzNw8SAxYo5_5DqsTmH0FHb_lKSMkuPunWWspC5sU.065MZxiV1ghCZZaz_j6bS-bQzdoVi3P1p0D7ddbW1hk&dib_tag=se&keywords=codenames&qid=1768870387&sprefix=codenames%2Caps%2C303&sr=8-1&th=1) or [Werewolf](https://www.amazon.com/Bezier-Games-UWEXBEZ-Ultimate-Werewolf/dp/B08WHMTZ61/ref=sr_1_6?crid=129TRJCVNWH95&dib=eyJ2IjoiMSJ9.odURcwqO5o9NYeuIJcD73GQtVIIvqBGwODwScsBNTDFQDKTKX1aha4ZtZD5iNdOcdExyR0k_x8q_3_RfhYC_l-8wk7nxV0-HGIAdW6N6nsvBOo2jt8FFI9Ed6_4QcEf2hATirGuP37k5yLwG1fxIFzCNwxi_yq2nfdi2lsdpy7DqbimdvxSK28omehMqPP6YahwkP55eJ7QzgiVz0nbh1cYMFBQamyclhZzxdjc78okvmvpdptpJBAVB9BfPG8lfdrgFUDREXL8LgYxw8x541SfW9664fgmsM213UNxGRQ0.3DapGao1kdctCYycHa6l180wXb95WvfDszcVNaBgNg8&dib_tag=se&keywords=werewolf+game&qid=1768870443&sprefix=werewolf%2Caps%2C204&sr=8-6) – Pretty good if you have at least 1 hour of late-night uninterrupted time to play. Werewolf needs someone who've seen it played before and can guide / be the session's God.

* I've played many of the [games here](https://archive.ph/I7rMf), and recommend them as well. Many need only paper and pencils or a free app.

## Other games I haven't explored

### Mao

The object of the game of Mao is to get rid of all your cards. And this is the ONLY thing you are told about this game. ***You have to figure out the rest of the rules by yourself.*** Every time you break a rule, a penalty is called, and you are penalized a card.

Arthur Machen's short story [The White People](https://archive.org/details/TheWhitePeople/page/n9/mode/2up?q=mao), in which a young girl recording her bizarre experiences with witchcraft:

"I must not write down the real names of the days and months which I found out a year ago, nor the way to make the Aklo letters, or the Chian language, or the great beautiful Circles, nor the Mao Games, nor the chief songs."

Every person who learned how to play figured out the rules by himself (or herself). If you can figure out the rules, then you know how to play. If you can't figure out the rules, then you don't deserve to be told what the rules are. (Actually, there are probably some people who were told the rules without figuring them out themselves. We true Mao players look down upon these unworthy slime.)

#### Ejemplos de reglas

[http://web.archive.org/web/20121228174230/http://kw.tamlylin.com/puz/mao\_sample.html](http://web.archive.org/web/20121228174230/http://kw.tamlylin.com/puz/mao_sample.html)

### Egyptian Rat Screw

Face Cards

If a player draws a face card from their deck, the next player has a given number of chances to also draw a face card or else the pile goes to players on their right. A player’s number of chances to draw a face card depends on which face card was initially drawn.

        	If an Ace was drawn, the next player has 4 chances.

        	If a King was drawn, the next player has 3 chances.

        	If a Queen was drawn, the next players has 2 chances.

        	If a Jack was drawn, the next player has 1 chance.

Slapping

The following card sequences allow any player to slap the pile and get all of the cards:

        	Two cards of the same rank are played consecutively

        	The card played is the same rank as the first card of the pile

        	A combination of King and Queen are played consecutively

        	Two cards of the same rank have one different card in between them

        	The number value of the cards played adds up to 10

        	An ascending or descending run of four cards is played

Penalty

If you slap the pile when no sequences are in effect, you must add two cards to the bottom of the pile.

[https://playingcarddecks.com/blogs/how-to-play/egyptian-rat-screw-game-rules](https://playingcarddecks.com/blogs/how-to-play/egyptian-rat-screw-game-rules)

[https://playingcarddecks.com/blogs/how-to-play/president-game-rules](https://playingcarddecks.com/blogs/how-to-play/president-game-rules)
[https://playingcarddecks.com/blogs/how-to-play/screw-your-neighbor-game-rules](https://playingcarddecks.com/blogs/how-to-play/screw-your-neighbor-game-rules)

# Icebreakers

# Icebreakers I like

Two truths and a lie

* Jot down your ideas before others start thinking of them, to set an example of the kinds of aspects of your lives you want to share

Three snaps (everyone shares 3 pictures from their life)

Everyone answers the same surreptitiously personal question

* What's the first memory you have?
* If you really knew me, you'd know that I…

Because of that…

* [https://www.sessionlab.com/methods/your-amazing-future](https://www.sessionlab.com/methods/your-amazing-future)

Create a one-word story on a topic

* [https://www.sessionlab.com/methods/one-word-method](https://www.sessionlab.com/methods/one-word-method)

Count up to the number of participants with your eyes closed. Start again if two people try to speak at the same time

* [https://www.sessionlab.com/methods/count-up](https://www.sessionlab.com/methods/count-up)

A name game where you say what you would bring to a deserted island that starts with the first letter of your name and have to say all the names previously said.

Honey do you love me (make sure to set boundaries ie: no touching)

Two truths and a lie (might need to split into groups)

EA charades (we just use common EA terms, organizations, and or people)

Human knot

Not an icebreaker: appreciation circle

* Take turns to express gratitude towards someone else/the group
* The receiver just says ‘Thank you’

**Tips for a 3-6-person, same-age group**

* Pick one activity, keep it under 10 minutes, and model the desired depth first.

* Encourage—but never force—vulnerability; use humor to lighten the mood when needed.

https://www.sessionlab.com/blog/icebreaker-games/

---

# Maybe do the thing you wish CEA would do

Source: https://myea.blog/be-cea

*I used AI to clean transcription errors, suggest better wording for tangled sentences, and make high-level tweaks to improve the writing.*

Three of the most exciting projects to come out of EA in recent years are, in a vague sense, CEA spinouts:

* **Kairos** is directly a spinout of CEA and now handles most support for university AI safety groups. Basically everyone I've found who knows them is really excited about what they do
* **NEST** is an opinionated ideas-first support network for EA (university) groups. And you can see from [Matt's blog](https://frommatter.substack.com/p/effective-altruism-will-be-great) the absolutely insane care he has for us.
* **BlueDot** grew out of a group of Cambridge EAs that wanted to make much better introductions to EA topics than what was available out there, and now they're basically trying to solve the talent gap for anything bad AI forever after

My natural next thought, is that maybe some more of the responsibilities currently conceived as “CEA's job” would be better handled by small teams not directly (or only loosely affiliated with) CEA.

Beyond the previous examples, there's many advantages to small independent teams here:

1. You can take risks that CEA can't
2. You can kindle a culture that promotes ambitious bets
3. You can set things from the start to benefit from pumping lots of AI into the project
4. CEA is just one org, they simply can't do everything

If you want concrete suggestions, here's the first that came to mind:

* More and better EA events.
  * EAGs are great, and EA Summits are an interesting experiment, but possibly there's many more formats and audiences to explore\!
    * Also, you can vibecode your own Swapcard alternative\!
* New introductions to EA ideas
  * I got into EA because helping the poor seemed great and doing that more effectively seemed even better. I'm not sure the AGI-focused content from 80k would have grabbed me, so there's possibly a good audience for what 80k was for me in 2016\!
* Online communities and discussion spaces
  * AI makes sourcing and curating content much easier, or you can pick a niche community and topic and start from there (I'm the proud owner of the \#too-much-ai channel for AI use enthusiasts\!)
* Intellectual leadership on EA ideas
  * Many institutions that did this have evolved or died (Open Philanthropy, GiveWell, the Global Priorities Institute, Scott Alexander, the Future of Humanity Institute), so, again, it's free real state\!
    * If you're looking for a concrete idea: What are the central virtues a committed utilitarian should cultivate? I would kill for a really great answer to this question
* Support for local EA groups
  * What about more active grant-making for EA groups? Maybe even faster funding? Just asking some EA community builders what they'd want and iterating from a minimal version of that could take you really far

I'd personally be very excited to see someone who's passionate about it pursue any of these ideas. Maybe it's more important to work on AI or something, but I think the heuristic of working on something you care about really hard can take you very far.

Also, you don't need to wait for CEA to give you a role to do any of this\! You can write up a Google Docs with the idea, send it to me for comments ([alejoacelas@gmail.com](mailto:alejoacelas@gmail.com)), figure out the minimal version you can get started on, and then go for it\!

---

# The costs of asking

Source: https://myea.blog/costs-of-asking

* in a naive version of ask culture/consent talk, you can do anything within the bounds of law and morality, as long as you ask and get explicit consent first
* but asking can be an act of itself “oh, attractive stranger, may I get your warmth? I'm lonely” and the boundaries of what you got consent for can be very blurry “how tightly can I hold you if you consented for a hug”
* a fix often suggested is to use your judgement to figure out the right asks, and do mostly those. but ‘avoid self-censoring from your internal judgment’ was the whole reason we created ask culture for
* Given how powerful and mutually beneficial asking can be “what do you think about the prospect of us marrying?”, it would be good to get a sharper picture of its costs, so we can pick up more of its juicy fruits while avoiding the downsides
* Here's a tentative round down of the costs of asking:
  * Being intrusive: many things here feel like “I think I recognize something in you/your plans that you don't see yourself. Let me venture my guess” and while this framing sounds innocuous, it often represents some form of being patronizing, paternalistic, or intrusive. I still don't know how to tell the good and bad instances of intrusiveness apart
  * Misrepresenting others: EV situation
  * Misrepresenting yourself: it's okay to be open about your flaws, and not actively hide them in interaction, but especially when interacting with people who don't know you much, it's easy for a question to divert people from forming a good picture of who you are (e.g., if they stop engaging or filter future interactions through an inaccurate first lens)
  * Causing discomfort: where does the discomfort come from? how much should we avoid it?
* It would be similarly great to get a picture of the benefits of asking, especially those that are unpredictable or don't accrue to the person asking

---

# Palantir for EA

Source: https://myea.blog/palantir-for-ea

The Plan

Over the next 3 months, I'll focus on 1-on-1 coaching to help executives and founders at EA orgs (VIPs) use AI to do their job better.

My case for it is:

AIM executives love our coaching calls, and I love doing them as well
Calls have a low cost of entry (1h30m of call + prep) and VIPs come out feeling more empowered and often times immediately more productive
Even before this, I gravitated towards interacting with friends in a coaching-style relationship
VIPs provide high leverage for AI adoption at their org
They can set up accountability structures, negotiate access permissions, and model a culture of AI enthusiasm
One-on-ones have quick reaction and learning cycles
I often react mid-call to something my VIP says. I can build directly the product they said they'd like to use
Group workshops allow more users at a time, but at the cost of a less custom, curated, and time-efficient experience

To do this, I will mostly focus on:

Booking coaching calls, prepping hard for them, and adding users slowly
Priority is giving each VIP the best user experience
Unsure on how early I should reach out to other orgs
Sourcing and making delightful software for me and my VIPs
I need to stay at the frontier to pull them towards it
I can find existing software for their needs, and quickly build new things to fill gaps
Headhunting for a co-founder
My current guess is that either an extreme AI power user or someone with lots of accrued trust among EA leadership would complement me best
I can scourge the internet for them, visit EA offices, and follow leads and recommendations
I'd love your help with
I'd love any advice, pointers, or introductions you could offer to solve the challenges below, so I have a better chance at ambitious success:

Preserve my focus
Decide on, and enforce boundaries for distributing my time between (A) coaching, (B) building, and (C) headhunting
Maybe wrap up/pause my recurrent mentoring commitments with Latam community builders (~3hrs/wk), and my final biosecurity deliverables
Maybe add friction to responding on email and Slack to everyone but my VIPs and co-founder leads
Get permission to work from the UK and the US
I have a visitor visa for both, but that only lets me “take calls” but not “work” for there
I'm applying for a UK Global Talent Visa, but I need to build up ‘independent evidence’ of being an exceptional promise in tech, such as public statements of my skills that are not directly linked to me
Prevent a Bipolar II depressive episode
Previous episodes have reduced my effective work output to ~10hrs/week for a month or more
I suspect daylight is a big factor. 16-hour daylight often triggers hypomanic episodes (which can be great and productive) and switching back to less daylight has been a strong predictor of the timing for depressive episodes
I have a decently good psychiatrist, I'm on the most suitable drug for prevention (Lamotrigine) and Claude provides lots of guidance, but maybe a great psychiatrist could help

---

# Some of my unusual habits

Source: https://myea.blog/unusual-habits

A friend asked me to make a list of my most unusual habits, here's some I came up with:

1. I persistently try finding ways to be more candid than I've been in the past
   1. How do I feel about you? Could I say it on this call?
   2. My blog
2. For things I use frequently I buy either the best stuff, or the cheapest, where ‘best’ is understood as 'whatever fullfills its practical purpose better’. The price difference is rarely ever that large
   1. All my current Claude sessions use Fable 5
3. I find the most provocative true wording for what I think. Often I let slip the truthful criteria if the meaning is clear (“may I steal this pen from you?”)
4. I ask the people I like best if I can visit and live with them. I love living with people and it's a great way to meet them
5. For cooking meals, and for many other things in life, I try to find a small set of principles to follow religiously and then ignore most other considerations:
   1. Each meal should contain a protein, a carb, and a vegetable/fruit. Protein should be enough for your body weight. Oversample dark leaves and seeds if you're vegan to get enough iron
      1. Everything else I indulge on: use as much fat and salt as needed, get pricey ingredients if you know how to use them
6. While traveling, most of my exercise is instrumental: buying a bike to move around, long walks to think about strategy, carrying heavy bags from Tesco, etc.
7. I don't make a point of visiting museums or attractions while traveling. If I feel like exploring I just head out the door and walk in whichever direction feels most uncharted
8. I try to orient conversations towards the most important things in my/the other person's life: marrying, being great at your work, intense feelings, etc.
9. I'm very intense in pushing people to do things I think are good for them. My day work is basically proselitizing Claude Code (maybe Claude Tag now)
10. For work, I mostly do whatever I feel like doing in the moment. I really care about the people I work with, so mostly I do the things I promised I would do for them, and then more
11. When I do something cool and I notice it, I raise my hands above my heads and perform a solo high-five
12. Until now at least, I answer basically all my email, WhatsApp, and Slack messages (may change soon as I get more busy). When the conversation dies I assume the other person was not interested/inventive enough to keep it going (which is very often fine, relationships have their ebb and flow
13. I keep no defined working hours. I try working all the time, but being extremely flexible to hop on invitations from friends, even if they come at the last minute
14. When booking AirBnBs, I mostly select for places with a horizontal surface (tables or desks can do), with a chair, and natural light. Almost everything else is negotiable
15. For the last few years, I've had my hair cut roughly once a year. I take it really short, and then go through several weird hair phases throughout the year until it's uncomfortably long
16. I spend my summers in Europe, the weather is nice and the people lovely
17. Sometimes I take notes on my arm using a temporary tattoo marker
18. I have periods on my life where I read Twitter a lot, and compulsively read a list of my favorite blogs
19. I often say hi to people on the street
20. I choose what to work on in good part by sensing how strongly I love the people I'd be working with

---

# Faith in humanity

Source: https://myea.blog/faith

\[[Crossposted](https://forum.effectivealtruism.org/posts/q9qBD8LvBLjsTbHaA/faith-in-humanity) from the EA Forum. AI Use Disclosure: Asked Claude to draft this, then entirely rejected its suggestions\]

I've read a lot of good moral philosophy linked from the EA Forum, but not much of it ([Joe Carlsmith excepted](https://joecarlsmith.com/)) has focused on some of the not-evidently-utilitarian traits and dispositions one could cultivate to become the kind of person that does the most good.

I'm discovering there's some meticulous writing on that topic as well, and I wanted to share here a piece that is currently resonating with me, though I suspect the faith I cling to is more about “people's capacity to do good” than on people's decency.

History and literature provide striking examples of people who are morally admirable, in part, because of their profound faith in people’s decency. But moral philosophers have largely ignored this trait, and I suspect that many philosophers would view such faith with suspicion, dismissing it as a form of naïvete or as some other objectionable form of irrationality. I argue that such suspicion is misplaced, and that having a certain kind of faith in people’s decency, which I call faith inhumanity, is a centrally important moral virtue. In order tomake this view intuitively more plausible, I discuss two moral exemplars – one historical and the other literary – whose lives vividly exhibit such faith. Then I provide a rationale for the view that having such faith is morally admirable. Finally, I discuss cases in which someone’s faith in humanity can lead her tomake judgments that are, to some degree, epistemically irrational. I argue that the existence of such cases does not pose a serious objection to the view that having faith in humanity is amoral virtue. Rather, it makes salient important limits on therole that epistemic, as opposed to practical, rationality shouldoccupy in our ideals of how to live.

Source: Faith in Humanity, Ryan Preston-Roedder, Philosophy and Phenomenological Research (2013). [https://philpapers.org/archive/PREFIH-2.pdf](https://philpapers.org/archive/PREFIH-2.pdf)

---

# Should we let everyone die to stand by virtue?

Source: https://myea.blog/let-die

* Is Zach Snyder's Man of Steel right about morality?
* I'd like the PauseAI guys to sacrifice some effectiveness at their mission to stand by integrity and principles
* I wouldn't like to do the same myself, but I recognize the Kantian argument here
* Posts about integrity for EAs feel bland. Do they think Clark Kent should have [let Pa’ die?](https://www.youtube.com/watch?v=yodVQ5QAc88) How high should our integrity standard be?
* Epistemic modesty (or some other epistemic virtues) feel similarly high stakes to the usual respect virtues. I think a lot of EA is failing hard at them now. Too enamored of our own AI narratives
  * We're prone to self-deceive ourselves away from utility-positive epistemic practices
  * The downsides of bad epistemics are similarly large to the downsides of disrespecting rights/principles: counterfactual blood in our hands. Oceans of it, in some cases. Galaxies worth of miscarriaged bodies
* I now have extreme sympathy for Eliezer. He's tried so hard to figure out what virtue looks like on an unexplored domain. Staying law-abinding when the whole arc of human civilization is geared towards its own destruction. He doesn't call for “bombing datacenters”, if everyone who thought bombing datacenters was a good idea did it, the world would be much worse. He calls for something [much more narrow](https://x.com/ESYudkowsky/status/1641953192761266177?referrer=grok-com):
  * “if a non-signatory country is building a datacenter that might kill everyone on Earth, you should be willing to preemptively destroy that datacenter; the intended reading is that you should do this even if the non-signatory country is a nuclear power and even if they try to threaten nuclear retaliation for the strike”

---

# Stop feeling dread. Start feeling regret

Source: https://myea.blog/stop-dread

* Dread is undirected. It prevents action without a clear understanding of consequences
* Regret is after-the-fact, and it's often much more targeted. You can picture what you did wrong, and even how you could have avoided it
* Early on in your life, you should prioritize learning. You can learn from acting on regret, you can't learn from acting on dread
* Regret rarely ever kills you, though it can impair quite a bit. Last time, when I regretted lacking a coherent strategy to push my AI uplift for EA idea, regret took \~3 months of my life, though I camed out much more determined to not commit that error again.

---

# Be so candid they ask you to stop

Source: https://myea.blog/so-candid

Drafted: May 21, 2026

*The world we should be aiming for*

Radical honesty is dumb. Starting conversations at the workplace with "ever since I've met you I've fantasized with you becoming my dutiful devoted wife" is bad business practice, and you should avoid it.

Still, I think for most people, at least in one important domain in their lives, they would greatly benefit from trying hard to be more radically honest.

The stupid thing about radical honesty is that it accepts no trade-offs. Truth should always prevail, no matter how that makes you feel, how that makes others feel, or whatever consequences it might bring in the world. That's extremely dumb but also extremely avoidable. The only reason you might consider it is because your own criteria to judge when to say something truthful is so fucked up that you'd be better off completely disregarding it in favor of a simple straightforward rule.

The thing is, I do think our criteria for judging when to say something truthful is extremely fucked up. So I do think you should disregard it. But maybe don't go all the way there.

There might be better stopping points. But one that I found actually helpful in practice is: just keep going until people literally start asking you to stop. And you can be proactive about it. You can ask people, "Hey, what do you think about my radical honesty? What cost is it imposing on you?" But also, even if you don't ask, people will start telling you.

One reason you might not like this rule is that you want to consider as well your own preferences for what forms of truth-sharing make you feel comfortable or not, what things you can live with after sharing.

I myself, I'm just oh so keen to tell other people about my interiority that I don't mind it at all. But maybe you will need to come up with a rule that suits you better.

Still, maybe start with my rule, because I do really think that radical honesty can be pleasant.

Just at last night's dinner I blurted:

* "What's your favorite person in the UK? What makes you like them so much?"
* "What are your non-pure motivations for being ambitiously altruistic? One for me is my ginormous pride."
* "You seem good at spotting exceptional talent, what's your take on me?"
* "I feel a bit sad after talking to a matchmaker and hearing that people with low cognitive empathy are really hard to match-make. It's really easy for us to hurt our partner."
* "You say the thing \[the EA sphere\] lacks most is good strategy. Do you have a non-self-referential definition of good strategy, or is bad strategy just 'strategy that's not exactly like mine'?"
* "What was the founding story of EA France? Why do you think the times of old feel better?"

It was truly a great night.

So, if any of this sounds appealing, I suggest you pick a domain of your life and give radical honesty a try.

Really, [it won't kill you](https://docs.google.com/document/d/1-vq5xBLjcYeoeFNImBe6iCNtTwj2m97mZ6TYzKkXUhQ/edit).

# emotions important?

Turns out emotions may be important

* Anti-thesis to my defense of extreme candidness: people's reactions and feelings are important. Moving to communicating more is great, but one of the main drivers of great human endeavours is the feelings people have about them. Narrative plays a big role here and precision and thickness of the information flow may be less important, at least in some domains
* Extreme candidness serves as a crutch for me. I fear so much how people will react to what I say, that it overloads my brain to try phrasing things in the way they'll be most receptive, so I end up not saying the thing

---

# The EAG Playbook for Career Success

Source: https://myea.blog/success-playbook

\[keep uppercase in title pls\!\]

* Aim: Have a secure-baseline and really high-upside strategy for succeeding in your career
* Pick an EAG, and go every year
  * Find mentors, find peers. Obsess over them
* There's great need in EA for people you trust and delegate tasks to. You can become that, even if you're not better than the average, and even if you're substantially worse (comparative advantage\!)
* Live in the EAG location for longer if you can

---

# Identity as a barrier to AI diffusion

Source: https://myea.blog/identity-barrier

On Slack I wrote:

One important trait in my mind for predicting who's a power user is "are you willing to let AI reshape your personal identity?"

Power users approach this different ways, but when I've encountered resistance to use AI, is often of the sort: "won't it fragment my attention?", "shouldn't we remain in control of work?", "I like doing the work myself", "isn't it bad to depend on AI for work/handling emotions/taking decisions?" which seem to come from some identity conflict in many cases

My attitude is often more like "I'll shape my cognition to whatever shape is most conducive to getting output from AIs" which does involve a lot of attention fragmentation and reckless delegation to AIs. Wonder how you guys feel that AI has influenced your self-image

---

# Will kindness prevail

Source: https://myea.blog/will-prevail

Carrie is perhaps the only good horror movie I've watched. Most horror movies draw their energy from things that I'm actually not scared about. I'm not scared about witches or demons, I know those don't exist. I'm not scared about mass murderers, those are unlikely to kill me personally. There might be some good horror movies out there about frequent car accidents or dying of cancer at an old age, but I haven't found them yet.

The horror in Carrie comes from asking a question whose true answer generally terrifies me: can kindness and innocence survive the cruelty and indifference of our everyday world?

If you have not watched it, Carrie is ultimately a familiar story about a young girl who is mocked in her classroom and is different and weird, but who also has a tender heart and the desire to just have a date for prom night and have the night of her life. Her warmth and feeling is real, and the warmth and feeling of many of the other characters is real as well. There's actually just not that much cruelty in the movie — most people there are normal, everyday, decent people, and only a few of those actively go out of their way to harm and mock others.

---

I surprised myself recently with how much conviction I was able to convey, to the people closest to a recently departed friend, that things will be good. I realized that I have some deep-seated optimism that I still cannot quite articulate, but it's there, and that I feel could plausibly stand whichever trial the world might throw at it. This post is my attempt to articulate, to rationalize in words, such deep-seated optimism.

I think I can hint at a proof of concept for the persistent value of kindness, and I think that should be enough to make me (or maybe make us) embrace it wholeheartedly for the rest of our lives.

---

Some of the best ways of helping the world are somewhat adversarial. There's some disagreement in AI Safety as to how valuable different policy proposals are, or what things we should be pushing for on the margin for AI governance, and how heavy-handed we should be in treating the AI labs. I think there's lots of valuable work to do there in being clear-eyed about the arguments on each side, and then taking a firm stance based on your best picture of reality. I don't want to do that, however.

My constitution is too frail for such a fight, and I'd doubt myself each step along the way as to whether I'm doing the right thing, mostly from the high-level arguments for epistemic modesty and for taking the considerations of similarly informed and thoughtful people into account, weighting them as much as your own.

But there are some ways of helping that practically everyone is really happy about (that "practically" does a bunch of the work here). There have been discussions about how positive it is to offer anti-malarial nets for free across Sub-Saharan Africa, and some reasonable people think that it would be better not to — that recipients or their surrounding communities are actually worse off from that effort. Still, on the spectrum of contestability for different altruistic interventions, anti-malarial nets lie far on the uncontroversial side of it.

But also many other things: helping a friend deal with their problems is also often uncontroversially good. You can go too far (and I've gone too far myself), but many of the options on the table there are just things that almost everyone who has context would support. Being kind and attentive in your emails: uncontroversial as well, and I really like doing that. There's a bunch of behaviors where it feels like you're just making the world better. You don't have to do seven steps of back-chaining in order to realize that you should be kind and act with integrity. They just seem like things that almost all of us would agree would be really good to do, and that, absent big countervailing considerations, we should just do them.

And I think, by character, I will never stop being kind — I just enjoy it so much. Even when I've been depressed and tired and unable to function at work, I still want to be kind, and even when I'm on top of the world and riding the wave and everyone's loving me, I want to be kind as well. So I feel I can just latch on to kindness. Whatever may come, I can be consistently kind throughout. And that's something I can control, and that's something that is good. And it's something that is comforting, and it's something that has made my own life, my friends' lives, and the world's lives much better.

So I need not worry too much about the precise maneuvering of making the hingiest point in history go better. I can be kind throughout. And maybe be strategic, and maybe be ambitious. But I can be kind throughout. And I will be kind throughout. Because that's me, and that's at my core. And as long as I keep that, things might go well.

I'm not saying we won't die — we might die. But as we near our final, passing breath, I will have been kind to those who tried.

# previous notes

My deep-seated optimism

* We all have the same heart. Or enough of us share some common core of values and caring, such that you can trust others to cooperate with you
* Everything else is just some information problem of sorts. There's this shared value, reshaping the world, and if it only knew how it would proceed
* Even if my strength fades, I can still resort to rooting for this common heart, and that will make the world better. I can be kind to my siblings, and say ‘merci’ in my of course truly impeccable French accent as I head out of the bakery
* Even if much of what needs to be done for the future is antagonistic (us vs them), enough of it will not, such that it's still worth it for me to focus on pushing shared values
  * Knowledge is a great shared value

(when thinking about altruism, or defeating existential risk) I tend to see all humans as a having a shared goal (same utility function, for the economists in the crowd). If that's the case at some level, then all the remaining problems seem to be some form of information asymmetry (I have inside info on my own skills and potentials, but can't easily broadcast it, and wouldn't even know who I should be broadcasting it for), or another standard econ concept that prevents perfectly altruistic dictators from reaching the social optimum.

---

# You can't compound on your own

Source: https://myea.blog/cant-own

Coda to Sam Altman's [How To Be Successful](https://blog.samaltman.com/how-to-be-successful)

* Robin can be a great researcher, but they won't be an Amanda Askell if they don't find a similarly big and demanded problem to work on
* When trying to list skills I was trying to compound on, most of them were about making it more pleasant for others to work with me. That seemed accidental, but I think it might be the case more generally, as most of your productivity increase will come from locating yourself in networks with better affordances to do great things (eg., receiving more money, directing others, advising governments, receiving curated problem ideas, strategic guidance from great leaders, etc.) rather than your individual productivity increasing
* The same way migrant's productivity increases more than 10-fold when moving to a higher income country, your main productivity boost will come from creating your own personal country, connecting to the infrastructure that's built in there, and creating some infrastructure of your own

---

# Reasonable plans, pursued with conviction

Source: https://myea.blog/conviction

* It's great to be on a field that moves fast. I've had ideas that I didn't pursue, and 3 months (and once 3 weeks\!) later someone else implements it and you get to learn if it was a good idea. Also it's great if *you* personally move fast. You'll get to know much more quickly which of your ideas pan out
* After I've seen some of my 3-month project ideas come up again in the future and be a bit more validated as a worthwhile endeavour, I've gotten more confident that I can actually figure out things that would be worth me investing a year or more working on. Specially if I'm proactive at formulating clearly and getting feedback on the high-level plan
* The conviction part is important. An important mission pursued with conviction is one of the main factors for attracting talent. The other is being able to recognize the talent's strengths and create a narrative for how they fit in that mission

---

# What I want from you

Source: https://myea.blog/what-i-want

I woke up this morning with a feeling of intense devotion towards you. I was thinking last night that it would be lovely to thread my fingers through your hair as you go to sleep, but I'm not sure whether you'd like it.

I think it is reasonable that you worry about my feelings being contingent on some expectations of how you should behave towards me (i.e. in a romantic manner). I think the base rates for unrequited love going well are low enough that this is something worth at least thinking about and discussing. I still don't quite know what to expect, or what is at risk, and I want to try hard to avoid it.

The thing I most want is to take care of you. Last night when you mentioned that if you felt sad at all the next day the hug would not have worked, I felt a slight tinge of pain. I felt powerless, because I know my hugs cannot cure sadness forever. And the thing that my mind says that I most strongly want is to be near you and to just push consistently for you feeling great and happy and fulfilled.

One aspect that you mentioned of relationships is this expectation of a certain behavior towards the other. In trying to think of what are the things that I want of you, I just feel that your default behavior towards me is so lovely. It seems to come from — I don't know where it comes from, but I feel — I feel no need to ask for anything else.

I'm not saying your behavior is perfectly optimized already towards my happiness, and so if I wanted to steer it in some way, one of the things that comes to mind is: I really dislike moments of ambiguity where I don't know how you're feeling towards me. I mean this only about brief moments in conversation, I'll give some examples. I like receiving your replies to my messages quickly, and most of the time I can just block out that thought, but the other day I lost the flight and it was entirely through my own fault, and it took you, I don't know how long, but it took a bit to reply to my messages, and I think I was just slightly mortified. I don't want to demand or exert any social pressure of any form for you to respond more rapidly. And I think the thing we're doing now, where I just feel free to express my discomfort and I hope you feel free to not act in any way, or to decide on your own, ideally prioritizing your comfort, how to react to it, I like that.

So that seems like a long aside, but the central point I was going for is that I don't want you to be anything other than what you are to me right now, or to behave any other way. And sometimes — and this is not imperative — but sometimes I do have, when I hear you breathe easily in your sleep, this deep felt sense of internal volition. I feel the parts in me rearrange, yearn, and crave a little to be nearer to you. Not sure what that is, but I wouldn't ever want to wake you up when you're sleeping. And I just want to take the things that are important to you and help them go better. And I think maybe you want that as well in some form, so maybe we're incentive-compatible(?)

PS. I also thought a bit about ways of encouraging or selecting certain motivations from me that might be preferable. So for example, if you want to discourage behaviors from me that are romantically motivated, we might think of having an explicit agreement that we won't date, at least for the duration of my stay, let's say. I think that's mostly an example for illustration. I doubt that it would do much in practice. I'm a terribly patient person.

I find it interesting that the general class of this problem feels really similar to certain parts of the alignment problem, where you have an agent whose internal motivations you cannot completely trust or control, and you try to find different schemes for eliciting robustly helpful behavior. Not sure if you find that intersection interesting or useful at all, but I just thought maybe love is a good excuse to learn more about the alignment problem.

---

# Have you tried violent passion?

Source: https://myea.blog/violent-passion

* A friend told me 80k finds it hard to hire for their advising team. I think the only reasonable explanation for that is that a bunch of you are absolute morons
* You either don't know what you want, or want really silly things. If you were passionate about something good, you'd realize how exciting advising at 80k is and how flabbergastingly talent-dense the 80k team is.
* Maybe you are too comformist. Quitting your job (or your three part time roles, which was my case) to work on AI Tools for Existential Security was always an option, and you haven't taken it
* Seems many of you would be indifferent between working at cG, Anthropic, or Redwood. You don't quite care what you do as long as it ticks some abstract silly boxes that some uncaring social gestalt demarcated as important
* Same for dating. Lilia, who works at a dating startup, tells me they enforce a rule of not talking about religion or politics on the first date, because some people are dumb enough to discard their future wife or husband based on a superficial dislike of their politics
* What's the opposite of that? What should you aspire to be in your life? Not sure if this fits for most, but maybe I'd suggest you try violent passion
* If Coefficient Giving rejects you from your dream role, maybe you should consider something along the lines of founding another cG, running it for 10 years, and then renouncing the CEO role to go take the role you were rejected for. The motivation behind that option should be clear if you really cared about what you're doing, even if for practicality you should do other things instead
* I served as mentor for a friend who didn't seem too competent to me when he started. He's now impressed me several times and he's developed a hunger for greatness that I quite like. The thing that I saw in him is that he's earnest, and maybe that's a precondition to violent passion, and a decent predictor of outlier success

# quests, not jobs

Look for quests, not jobs

* Jobs constrain the space of what you can try
* The space for ‘succeeding’ (ie., growing a lot and maybe having some valuable output) for quests is much larger than for jobs

---

# Art is the thing that engages you

Source: https://myea.blog/engaging-thing

* It may not engage you superficially (long boring books), but it must engage you at some level
* Force/compell you to think or feel, often doing some shaping as well
  * Often? Is there a case where art just compels feeling but it's unopinionated about content? No
* People and places are great forms of art. They're among the most compelling things we have
* Most of what we call ‘art’ refers to the non-rival / scalable forms of art. Those that can be shared and appreciated by many people
* Food can be a more or less artistic form of art. For some it may be just an immediate thrill, but it can also bond you to people around you, to a community or set of ideas
* Drugs are a pointy kind of art. Many forms of art (those that detract from your enjoyment of other arts) can be very drug-like
* This definition feels too broad. I wonder how we can narrow it. Maybe:
  * Art needs to be deliberate (nature?)
  * Art is reserved for the strongest/most compelling forms
* The classics can be compelling for a really broad set of people, and they can be compelling in a lasting way
  * Movies may surpass books here. They're more accessible and in some cases have stronger and more lasting effects on people

---

# Your colourful heart

Source: https://myea.blog/colourful-heart

This morning I cried with strong emotion, thinking of the draft of this blog post. Now I'm feeling a bit tired and the emotion feels a bit more muted, and I'm not sure if I like this impermanence.

One theme of this piece is unconditional love, and what does it mean when the people you love could radically change. I feel some bond to the people I've loved for a while. And maybe thinking of the first girl that I loved, it's been 16 years and I still love her all the same. I don't quite manage to love her; I stopped seeing her maybe 13 or more years ago, but I hold some fondness and strong feeling towards the memory of her and the times that we've had, which were basically just playing tag for hours after our swimming classes ended. I think she was the nicest person ever. Not, like, for real, but you know, the nicest person ever. And I'd really like my love to do you justice, and I don't know what such justice entails.

I love your colourful heart. I love every single thing you pour yourself into. And I love the joy, and the expansiveness, and the absolute thrill that I see reflected in you.

My eyes swelled with tears because yesterday I gave you a gyoza, and I think you liked it. You taught me how to dance, and I love the people I danced with. I love your friends. They are the nicest people ever (it's becoming kind of impossible because I've already used that adjective for like maybe seven people so far in this blog post, but they are). And I'd really like them to be part of my life as well.

I love that you aspire to be good, and that even through pain and disappointment you just keep wielding this heavy will to carry things through. I love the way you treat me. I love that you made couscous because you noticed I would like it. I love your laughter, and yesterday had lots of it. I love hearing music through you.

Will you keep being this way forever? Probably not. And what will there be of my love after its original object seems to be gone? I don't know. I hope I would keep loving you, but what will this love mean if unanchored from its original place?

Thinking back to unconditional love, I know I've tried to change you. And I don't know how much I should. I want you to be strong to compensate for my weakness. I would like to promise that you won't need to fight and suffer ever again, but I know I can't deliver. I can't even give you my all, because I've made so many other promises. I don't know what strength looks like and how one acquires it, but I know that in many of the most important events in our lives we are fundamentally alone, and I'd like you to be strong such that you can prevail.

---

# You literally cannot die

Source: https://myea.blog/inmortal

I mean, you will possibly very literally die, but not from the things you worry most about on your day to day. I think that means that you should identify the things that could kill you—mostly metaphorically kill you—and avoid those like the plague, but otherwise do everything. Let me explain.

Mother of Learning is this extremely long but highly rated fantasy series online that follows the story of Zorian, a young mage that finds himself repeating the same month for years on repeat. And part of what makes this story so thrilling is seeing our protagonist (and eventually a small group of time traveler friends with him) take advantage of the time loop mechanics to become unbelievably strong and powerful, which they do in order to save their city from the invasion that invariably happens at the end of the repeating month. One of their greatest assets in using the time loop mechanics is that they cannot die: every time they're killed they just reappear at the start of the month as usual. So a lot of the things that no other mage is trying for being too risky, they just dive head-on into them. They can break into the mansion of a much more powerful archinemesis and die in a thousand myriad painful ways, just to have a chance of figuring out what his plan is all about, or track down the sickest ever spider monster to collect its remnants for a potion that will give them enduring strength. \[^1\]

\[^1\]: It's interesting here, and very parallel to real life, how the value of trying these things that kill you a lot is that once you've cracked them, you can reuse them at a much lower cost each and every month going forward.

And it's not even true that they cannot literally die. They're vulnerable to soul magic, which could sever the link they have with the time-traveling machine. So in a sense the protagonists of Mother of Learning and us are similarly vulnerable. Each of us has available a range of daring actions that have little chance of killing us at all, or of leaving us with long-term damage. And the point I'm trying to convey here is that many of us don't take full advantage of the time loop mechanics of our life. Years are composed of months one after the other, and months of days, one at a time, and you can just try things, and if they don't kill you, you get to try things again.

Something that made me viscerally realize this was shortly after watching Groundhog Day, a movie with a very similar premise where a weather newscaster is stuck reliving the same day again and again for what seems to be several decades. I thought of how cool it would be to, just as in the movie, be able to go to a bar, flirt with a girl, and if things don't go well, then go to the same bar the next day and try a different time and a different strategy. And it took me a bit to realize, but I can actually do that? At the time I was thinking about new strategies for finding potential partners, and turns out that you can actually go to a bar, try to flirt with a girl, and if things don't come out the way you want them, you can try again the next day. You might do a different bar if you're feeling a little shy, but you can quite probably just do the same bar—just not the same girl, but for my purposes that's not a crucial detail.

So, a lot of things we could do to meaningfully advance our goals, we don't do. And it seems to me that we don't because, in a sense, we don't realize that they cannot literally kill us. That failing at them leaves no long-standing scar.

Still, before I go out and recommend that you go crazy in your life, I think it's important to meditate on what things do have the power of killing you—maybe not killing you completely, but causing long-standing damage to your goals or the goals of others.

There's this amazing scene in Mother of Learning where the protagonist is being chased by a \[rival time traveler with mastery of soul magic / soul-mage necromancer who can permanently sever him from the time loop / lich whose soul magic could end him for good\], and after having an intense magical face-off where the protagonist realizes how completely overpowered he is, he manages to weasel his way out and run through an underground tunnel, and in his last moment of consciousness he manages to throw himself into a deep pit full of horrifying magical creatures that will eat his body and all remnants, once he explodes the suicide bomb cube he always carries in his neck. I think the story is a great illustration of the great lengths you should go to in order to not die. You should be willing to blow yourself to pieces if the thing at stake is your own life.

So, what are the things that can kill you in real life, and how do you avoid them? I still don't have a great answer to this, but I intuit that most of them are related to upholding virtues. So I think if you're honest and respect others and are well-intentioned and candid—and maybe a few other things—you can for the most part avoid the things that could lastingly harm you or harm those around you. You also need to be careful not to literally die, such as by not scuba diving or possibly not riding a motorcycle.\[^2\]

\[^2\]: Or, if riding a motorcycle, just figuring out the most common causes of accidents and avoiding those. Maybe I'll make a blog post on this in the future, because motorcycles are so fucking convenient.

So yeah, I'm relatively hopeful that through talking to friends and \[looking at cases / seeing other cases\] and thinking about what things are \[worth pursuing / worth caring about\] and important, you can get on a path where you identify a certain really high standard of integrity you can hold yourself to, and if you abide by it, I think you can avoid the things that kill you. You'll still probably do things that embarrass and hurt not only you but others as well—I even have some still somewhat painful recent examples of this from my own life—but when I look at them, the downside is limited, and so I feel extremely enthusiastic about just trying more stuff.

I haven't given that many examples of what you can actually do with this newfound freedom, so I'll quickly give one. I think for myself the rule I want to follow in candidness is: be as candid as possible until the point where people literally start begging me to stop, and then at that point pull back just a little. Friends have already started begging.

But I so thoroughly enjoy it. I think it has led me to so many deep and meaningful conversations. I confessed to a girl that I liked her, and then after she told me she's not romantically interested in me, I tried enrolling her support for helping me win her over. This is partly a silly example, but I think that effort went great—or rather, the whole thing about just being extremely honest with her about my feelings, dispositions, and desires has ended up in us just talking for dozens of hours already, even just in this span of less than two weeks, about who each of us is and what we care about and how we can make the best out of living together for helping the other person live up to what they most cherish.

Today after drafting most of this blog post I took a nap and as I groggily started waking up I had the song Vienna in my head and when I heard Robin's footsteps surround. I whispered her over and told her that I had a song on my head and that I wanted her to come closer. And I would sing Placeholder in what would amount to a really intimate moment. I actually didn't do this. I just told her that I wanted to do this. But I think through my barely open eyes, I saw her look at me with this tender affection. I love this and I love the life I'm building out of this borrowed courage.

So, if you want a single lesson, stay clear of the things that kill you, and do literally everything fucking else.

---

# My heretical fiction ideas

Source: https://myea.blog/heretical-fiction

**1\. Sam Bankman-Fried, but you root for him.**

What would it take to make us empathize with Sam Bankman-Fried? He did have a grandiosely good purpose. I'm not too fond of his methods or poor choices, but maybe something nearby is entirely relatable, and to some extent maybe even defensible.

I would really enjoy reading a fictionalized account of his trajectory that does two things:

* Use narrative to seduce the reader's natural capacity for empathy — to make Sam feel more human, more understandable
* Given that the real Sam almost certainly did a bunch of uncalled-for things and made poor strategic choices in his pursuit, the fictional account would also need to edit some facts, such that the actions he eventually takes all come from a deeply strategic, earnest, true, and appropriate caring — rather than a misguided one.

**2\. Jesus as a strict utilitarian.**

Every telling of the story of Jesus where he was just a strict utilitarian and managed to shape the world to be much better. I'd like to stick to the Bible — the main plot points and stories should stay as they are — but have freedom to edit the rest of history if necessary.

The animating question: *how would the world need to be for the most utilitarian thing for Jesus to do to be Jesus himself, and to do what the Bible tells us he did?*

This probably won't be written, partly because there's only one Scott Alexander and he's Jewish, but also because this particular community for the most part doesn't have the breadth and depth of thought about history to pull it off. Still, I'd be pretty excited to read even multiply short-comed versions of this.

**3\. Tales of extreme ambition.**

A collection of tales tracking the different paths toward the most outsized stories of impact in the world we've had so far. Why not Will MacAskill and Toby Ord, Sam Bankman-Fried, the Anthropic founders, Dustin Moskovitz — and maybe Jason Matheny.

You'd follow them closely and try to answer: how do these different people relate to extreme ambition? What decisions do they take that are important, that shape years of their path and careers? In what ways do they double down on who they are, and in what ways are they intensely aware and strategic about reshaping what they do and who they are? I think that would be a truly enjoyable read.

---

Also, these ideas feel to me like one of the strongest arguments for the value of fiction. If I read something like any of this, I would learn so much about different things that it would be really hard to learn any other way. And writing any of them would take some amount of sheer obsession and quality of thought that's hard to find anywhere.

But the thing is, the classics come from a similar place. They may not be posing the specific questions I'm trying to answer here, but they pose deep questions about the human condition, many of which could be relevant to the purposes I have myself.

So — what a good argument for fiction and the classics.

---

# Chatting Claude about my avoidant attachment

Source: https://myea.blog/avoidant

*A lightly trimmed transcript of a conversation with Claude, right after watching [Heidi Priebe's video](https://www.youtube.com/watch?v=GTQohPaGnSY) on "intellectual bypassing" in avoidant attachment.*

---

**Me:** This was a day we had a friend visiting. I had a crush on Robin that I kind of didn't want to acknowledge I still had  — it would have made things harder for both of us, and I didn't want to impose. Robin had a crush on the friend who was visiting, and I was actually rooting for them, helping them figure things out.

At some point Robin stubbed her toe, and I felt this quick, immediate joy. I blurted out “whoa, that was weird, I felt joy when you stubbed your toe” She asked why. And what came out was “oh, I think I resent you.”

For context, a couple of days before, I'd heard that a friend had killed himself. It had been an incredibly demanding day — I'd spent it glued to my screen, messaging and calling his friends, being this strong, helpful presence that carried them through the grief.

So it was late, and I wanted some comfort. The thing I wanted was to lay my head on Robin's lap and have her stroke my hair for a bit. But I felt shy about asking, and I knew that even if I asked, she'd reject it.

The thing is, I wouldn't have voiced any of this if Robin hadn't asked. I wouldn't have even recognized it inside myself if she hadn't.

**Claude:** You felt the joy and said *that was weird* out loud before you'd tidied it into a reason — the opposite of putting a lid on it. The "you" in "I resent you" might be less Robin than the arrangement you'd built, the one with no room for you to be the person who needed her. Did naming the resentment become its own way of not asking?

Oh, I did ask for the hair stroking. And Robin said okay — but is this maternal? Because I'd phrased it as a maternal thing. Or is it romantic? I think she wanted to diffuse the romantic reading, because she wasn't willing to entertain it. (She refused to stroke my hair, without hesitation.)

And I think that's part of the problem: there are a lot of things I want that I just can't ask for. I find a girl attractive — should I go tell her, hey, I'm attracted to you, I'm lonely, can I have some of your warmth? I feel lonely occasionally, and there's not much I can do about it.

I've been getting a little better at voicing my needs. I can do it with a few people — I did it recently with Maya, who I love. I just said, here's the thing I want from you, you don't have to give it to me, it's just a thought crossing my head, but this is what I want. But for most people, I think you can make the request and they just won't do it. And it might burn them a little — if someone feels ineffectual, some momentum in the relationship gets burned.

So generally I just don't ask for stuff. Or I do, but only when it's a really small ask — I send lots of cold emails. Even those were hard to get started, and I still feel awful sending them to people outside the EA community, because I don't feel close enough to take that affordance.

So there's a real structural thing here: you can't just ask people for things, because they might not give them, and they might get tired of the repeated asking. There's a calibration that has to happen. But I don't know if that calibration is intellectual, or if it's something more emotional — alternative paths I'm not exploring because they just don't feel emotionally available. I don't know.

Your structural case is mostly emotion in a suit. You already have the skill — you used it with Maya, with the exact frame that makes an ask cost nothing: *here's what I want, you don't have to give it.* The fear looks borrowed from a different lane: the candor people have actually begged you to dial down.

Oh, actually, this is interesting. So the candor is something that I learned to do. I used to restrict myself much, much more — I wasn't honest with people when I had feedback to give them, and I didn't voice my feelings, I didn't voice most things. And candor is kind of like my overcorrection. Candor is: I was scared for so long and didn't voice it, because I didn't want to impose on you, I didn't want to make you feel bad. And when I tried being a little bit candid, I realized that I had been leaving so much on the table, and so I was just like, okay, need to go as far as I can in this direction. And now I'm noticing that going as far as I can also has its costs. But I fear — when people push back against my candor, I feel scared that I might not be able to ask for what I want or say how I feel. Because it's so easy to turn on this machinery that tells me that I'm causing harm in people. So I want to defend my candor. I want to argue for it such that it has a place, and such that I can voice my feelings.

Then the freeing move isn't winning the argument that candor is good — it's seeing the binary, where conceding *any* cost means total surrender back to silence. Unbundle the self-voicing (*I feel X, I want X*) from the doctrine (go as far as you can, unsolicited, until they beg you to stop). A hit to the doctrine doesn't revoke your right to say you feel lonely. The candor you can't kill is the one that doesn't need to be right.

Yeah, I kind of want to say fuck you a bit. Like, you're sidestepping the problem. So — in any particular situation. Like, maybe also good context: I have, I think, what might be called like really low [cognitive empathy](https://en.wikipedia.org/wiki/Empathy#Distinctions_between_empathy_and_related_concepts:~:text=%5B9%5D-,Cognitive%20empathy,-is%20the%20ability). So it's really hard for me sometimes to understand how other people are feeling. And so the advice that you're giving just feels like, oh, just bear with it, or just figure it out, man, learn to read people's feelings. And a lot of my blog posts about finding rules is my attempt to just figure out some sort of principles that I can adhere to that don't require me to just become better at reading people's emotions overnight. I want to become better despite this limitation that I have. So yeah. I don't — fuck you kind of.

Fair hit. I'd smuggled the reading task back in — and reading other people is the one faculty you've said you lack. But feeling your *own* interior isn't that faculty, and that's all voicing needs. So here's a rule built for the instrument you have, not the one you don't.

You can't perceive inferred harm, so the machinery's verdict — "I hurt them" — runs on data you don't have and can't check, which is exactly why it's bottomless and terrifying. Starve it: the only harm that enters your decisions is harm someone tells you about, in words. If they didn't say it, it doesn't count. (This would be poison for a callous person. You are the precise opposite problem.)

And the Maya frame is *the* tool here, because it outsources the read — you hand them the wheel and ask them to say out loud the thing you can't see. One last check you can run without reading anyone: sort your own sentence. *I feel X / I want X* — voice it freely. *Here's your fatal flaw* — that one waits to be invited.

---

# Research reports for everything

Source: https://myea.blog/research-reports

Sometimes I want to spend 1-10 hours making sure I really understand a certain question. In theory the *deep research* functionality of most chatbots is the affordance you'd reach for. But I find it's often hard to trace back the sources and the details of how each claim was arrived at.

I already know how to do a decent research job. I have a sense of what sources are likely to exist, what a high-quality analysis of the evidence looks like, and roughly what the output should feel like — even when I can't yet picture the answer or the exact format. In those cases I just open Claude Code and start narrating, using dictation, what I want to do.

Here's a recent example. I just started staying with Robin (not her real name) — she's vegan, eats a lot from *Too Good To Go*, doesn't like wasting food, likes to eat cheap, cooks mostly Chinese-style, and has some health complications if she eats too much protein. A bunch of details I wanted to accommodate. And I know roughly zero about nutrition science beyond *vegetables good, variety good*. I wanted to figure out a few grounded recommendations or action points so we could both eat in a way that's healthy, cheap, and keeps energy levels up through the day.

This felt like the kind of thing I could delegate to Claude Code. So I timeblocked around 40 minutes to make a quick first pass, and probably will invest \~3 hours to get actionable things to incorporate in our diets.

## Claude Code Log

**Or see the full conversation here directly: [https://gisthost.github.io/?8c3277444a3aa7adeeb260181dfcfed5/index.html](https://gisthost.github.io/?8c3277444a3aa7adeeb260181dfcfed5/index.html)**

### 1\. Open with a long ramble \+ ask for questions back

I dictated the first message: \~400 words of context, half-formed ideas, constraints, things I'd half-tested before. It ended with:

*"give me an initial prototype of what you're imagining for this project and then just ask me lots of questions to help me kind of dig down a little bit more into the steps of the design"*

I almost always include "ask me lots of questions" in my first prompt. When I get started, I often have only a blurry idea of what I want, and Claude is pretty good at posing questions that give me enough surface to clarify further. This time Claude came back with a proposed structure and eight clarifying questions; almost all the design that followed was just me answering them.

### 2\. Reuse anything you have at hand

I had a set of fact-checking prompts a friend wrote for an unrelated Wikipedia automation project. Instead of rewriting from scratch:

*"There were some really good prompts here for fact checking and corroborating sources... I'd like you to git clone this repo \[equivalent to downloading a Drive folder\] and find the prompts and then just put them in a folder titled like Example prompts or something like that."*

### 3\. Define in `CLAUDE.md` the default behaviour you want

*"instead of having a readme. We should have a Claude.md and Claude.md should kind of point where the structure of the repo... such that if I point an agent... 'let's explore this question'... then the agent will know okay how should I proceed to investigate this question"*

Everything you put on a file named CLAUDE.md will be loaded automatically in every Claude Code session you have on that folder. Here I wanted to make it easy to spin up new sessions to ask nutrition questions, so that's the behaviour I hardcoded on the CLAUDE.md.

### 4\. Figure out how to structure the final answer

Claude proposed a Q\&A document format. I half-rejected it:

*"really instead of trying to kind of answer my question I in many cases want you to kind of present in a structured manner the evidence that's already available... something more... that has more nested depth and allows you to like look at different claims and follow the sources"*

The doc format ended up as an *evidence map*, not a clean answer. I also requested a hard three-bucket priority verdict (`Worth prioritizing` / `Easy to handle` / `Ignore`), which makes it easier to directly act on the output.

### 5\. Use subagents to break down the task

It's better to give smaller well-scoped tasks to AIs. One way to do it is breaking the goal into a pipeline, and asking Claude Code to use different subagents for each step:

*"yes I think it should be heavy so red team agent should actively search for contradicting literature and then I think we want to wait for like the adversary agent to like flag stuff anything that's medium or high priority"*

### 6\. Make agents save their results

By default, Claude won't save a permanent copy of intermediate outputs. That's silly, you really want to keep them to skim/review or reference later. I told Claude:

*"each sub-agent should know where it should save its outputs and it should never finish if it doesn't kind of give us a markdown output... give me a few options for how to do the directory structure for this."*

### 7\. Test it quickly and iterate\!

*"Thanks\! Can we [dogfood](https://en.wikipedia.org/wiki/Eating_your_own_dog_food) this? Is there a question we can run this on? I think actually just pick a question, make some first documents, I'll be away from my computer and when I go back I want to like kind of see the full chain of things and see if I'm pleased with the result."*

Don't ask "is the pipeline done?" Ask the pipeline to run on a real question and judge the artifact.

# The output: what the pipeline looks like

The "output" of the design session is the pipeline itself. Each pipeline file below is linked as its own Google Doc so you can read it standalone. File names are anonymized.

- [**CLAUDE.md**](https://docs.google.com/document/d/1y56dwUcENnUhULMuXYC85sRPiwGmemRqJzTFoUDPO5k/edit) — the operating manual: workflow loop, file conventions, source-registry schema, and an opinionated stance on nutrition evidence quality (when to trust observational data, how to weigh conflict of interest).
- [**personal-context.md**](https://docs.google.com/document/d/1ODQOzHAjxI-RsrtYC71xmt5qdvkaEKO06E-2wpzZQHU/edit) — who this is for, prioritized goals, cooking and ingredient constraints, health notes, working hypotheses, and explicit non-goals so agents don't optimize for things we don't care about.
- [**questions.md**](https://docs.google.com/document/d/1zVmm-TM5AR_ANpHDlTaWi6lb4sHjq3MwRke_J7gWu5g/edit) — the backlog of nutrition questions, tagged by status (open / in-progress / answered / parked); the entry point when starting any new investigation.
- **docs/** — the final, polished evidence maps. One file per investigated question; each surfaces contrasting positions instead of averaging them away.
- **investigations/\<slug\>/** — the full audit trail per question. Each subagent's output is numbered by workflow step:
  - `01-framing.md` — sharper restatement, hidden assumptions, sub-questions, first-pass answer with per-bullet confidence.
  - `02-scan-<subq>.md` — ranked candidate source list per sub-question (one file per sub-question).
  - `03-fetch.md` — fetch report: which sources resolved, tier classification, paywalls hit.
  - `04-draft.md` — working evidence map; eventually promoted to `docs/<slug>.md`.
  - `05-red-team.md` — adversarial flags by severity (low / medium / high).
  - `06-verify-batchN.md` — claim-verifier per-claim verdicts (one file per batch of 3–5 claims).
  - `07-grade.md` — evidence-grader confidence tiers per claim and overall doc grade.
- **sources/**
  - `registry.json` — canonical source list with per-source tier (A–E), tier reasoning, and funding notes; every claim in the docs cites by id from here.
  - `pdfs/` — local copies of papers when no stable URL exists (fall-back so links don't rot).
- **example-prompts/** — battle-tested prompts ported from other projects (e.g., the citation-accuracy-checker that I adapted into the claim-verifier subagent).
- **.claude/agents/** — six specialised subagents, each with its own context window, dispatched by the main loop in workflow order:
  - [**question-framer**](https://docs.google.com/document/d/1tduwA1e256IjAiwPatPA8dIXDw5RUIfjWRvaVbz3Jmo/edit) — turns a vague question into a precise one, decomposes into sub-questions, gives a first-pass answer with confidence so many questions can end here without spending more time.
  - [**literature-scanner**](https://docs.google.com/document/d/1j6NDxjhThLtlkLom2weBBUqVF2rEwIf9qoqkFyUkAxw/edit) — searches PubMed, Cochrane, Examine, and Scholar for one sub-question; returns \~15 ranked candidates with study type, sample size, and one-line relevance.
  - [**source-fetcher**](https://docs.google.com/document/d/11n89vt90GzddE2V6BBPsd_7IlzoW0eCtZTt3Ebow5xc/edit) — resolves candidates to readable full text, downloads when no stable link exists, classifies evidence tier with reasoning, and appends to the source registry.
  - [**red-team**](https://docs.google.com/document/d/1VTl1CTE9Ogwy9IKWoQCk1_w5MY1q1Q0dBANdvyv0G50/edit) — adversarial reviewer; actively searches for contradicting literature and methodological weaknesses, then flags by severity. Flags only — never fixes.
  - [**claim-verifier**](https://docs.google.com/document/d/1cNuOaagNWu46nDnGKqC0so4EUggWGFh9PBSSRJGRKF4/edit) — for each (claim, source) pair, checks the claim faithfully represents what the source actually says; strict on numbers, population scope, and mechanism-vs-outcome conflation.
  - [**evidence-grader**](https://docs.google.com/document/d/1Y9FyPwy0LF3e420yiYZ-uNtlZdGttl-7u3TK8QhNSGU/edit) — tags each conclusion in the final draft with a confidence tier (A–E) and reasoning, then calibrates the doc's headline confidence statement.

You can see the full prompt-by-prompt session that produced all of this, plus the outputs linked above, here: [**https://gisthost.github.io/?8c3277444a3aa7adeeb260181dfcfed5/index.html**](https://gisthost.github.io/?8c3277444a3aa7adeeb260181dfcfed5/index.html)

---

# There is rarely ever a good reason to do things later

Source: https://myea.blog/do-it-now

Doing now is great. People love it when you do things quickly. You love it when people love it that you do things quickly. You don't have to reload the context — you're already in it, so you deliver.

You're also freed now, so you can do more things later.

Doing things now does not mean doing all of it now. Oftentimes you need external input, agreement, stakeholders, or you need to think about it — but you need to get started on the task and advance it as much as is practically available. So if someone sends an email, you can probably reply now. Why not? What would it take? What special power will you acquire in the future that will make your reply better than the reply you'd write now?

One exception is if you have another thing happening now. You have two WhatsApp threads from different people. Then you have to pick what to do now, but you should do each now. So pick one, go for it, then do the other.

If you're not going to do it now — and I've done this often — just ask the person to help create a now in the future. Tell them: I won't reply now, but message me at this time or this date and I'll do it now then.

I don't adhere to this strictly. Blog post ideas I often have up to five in the shower, and I don't do them successfully after I get out. I often just do one and let the others go to waste. I may do them later, but probably not. You possibly need some sort of prioritization because there are many things now you could be doing — but that's part of the trick. You either do something now or you don't do it.

Another exception is bunching similar tasks together. In those cases, your now should consist of piling the task in the bunch such that you can decide to act on it later.

There are also things I've delayed for a while. They often feel important. And many of them are just delivering something for someone I'm working with, and I find it weird not to have the energy to do them now all the time. I'm not sure exactly what's going on there.

Sometimes I do need to think about something. I need to have it in the background of my mind and have this consistent background pressure shape the output. For example, I've delayed applying to a Coefficient Giving RFP for a while, but at least for the past week most of that delaying has been pretty intentional. It's like trying to cook in the back of my mind what the right shape for our proposal should be.

But that background cognition has a cost. You shouldn't incur it freely. You shouldn't let people impose on you some background process in your mind. If someone sends you an email, you should either reply now, or tell them: I won't reply now, and probably I won't reply ever.

Some tasks you do want to schedule for later. I often set up reminders to prod friends to do something. But those reminders are basically my phone ringing — as soon as a reminder hits I have to stop whatever I'm doing and do the reminding thing now.

Maybe the most common reason I don't do things now is that I feel tired. Some of that's valid — I do think you should try sleeping eight hours every night and having a consistent sleep schedule. But other times it feels fake. What's three hours of straight continuous calls? Why can't you just send all the follow-ups you've collected right now?

Leaving things for later just feels like being inconsiderate with your future self. Why think the future self will have it better?

Maybe in some cases you might have reasons. You have a strict external deadline that you have to adhere to. In those cases — which are rare, I think, for most people, because you can always flex around your deadlines and scope down commitments — you don't have to do things now.

But the rest of the time, which I think is basically the whole of your life, you should basically do things now.

As an additional example — not work-related, not directly, though it's intensely work-related for me given how work-related my life is — you should let friends invite you to stuff. You should let people interrupt you. You should let people who are near to you possess your now. The arguments are the same: people love it, you love it, and close relationships are one of the most valuable longer-term investments you could make, such that investing more of your now in them is a real asset.

\[As another example, one of the best uses you can make of AI is turning more tasks into “now and done” tasks. You often don't even need to know how to get started with a task to dictate it to your Claude Code, you just need the will to get it done now\]

There's no limit for what you can do if you do things now (false, but I still haven't bumped the limit yet\!)

---

# Quick wins in AI uplift

Source: https://myea.blog/quick-wins

# Quick Wins in AI Uplift

[Alejo Acelas](mailto:alejoacelas@gmail.com). [alejoacelas@gmail.com](mailto:alejoacelas@gmail.com)

I may get to these eventually, but if you could take one of these and do it now it would make me really happy, and all EAs would be forever grateful.

Leave a comment here/email me if you want help in getting started. Do send me the final output so I can share it around\!
---

### Reduce friction to share knowledge

Collect possible concerns about sharing work repos publicly, then create a Skill to facilitate the process

* Some AI Uplift staff at EA orgs are open to sharing their projects, but mostly do it on an ad-hoc, case-by-case basis
* Considerations for what and who to share with don't seem to vary much across EA orgs. Talking to a few people (I can introduce\!), collecting their concerns, and investigating them should give a reusable questionnaire \+ workflow to share stuff
* The final thing should be really easy to use (and ideally require no manual intervention most of the time\!). A Claude/Codex Skills seems like a good default

Make some of your projects public. Ask AI to summarize them on a website/blog

* You've probably built some really cool stuff. The world has to know\!

Create an AI power users/AI enthusiasts Discord

* I've heard that Discord is much better than Slack for high-bandwith communication and idea exchange, but I don't use Discord so it would take me a bit to get one started
* I know many people who could potentially be interested in this Discord channel, and would invite you if I trust you to have a good idea on how to set a good culture for the online space

### Make intro material

Record yourself installing Claude Cowork \+ Code on Windows from scratch

* There's many unintuitive things about installing Cowork/Code on Windows. You need to install Git (what's that?), sometimes create a virtual machine (what?), and many other minor annoyances
* If you understand the process, you can record yourself doing it and explain things intuitively as you go
* Ideally the final video is \<10 minutes long. I may help with editing if needed

Record yourself setting up Github with Claude Cowork. Create a 3-sentence instruction for Claude to use it well

* GitHub is the best tool for sharing AI workflows and projects within an organization. Most people haven't heard of it, least of all used it
* The manual parts of installing Git and signing up to GitHub are not too bothersome. Most people just need a guide on where to click and how to get past the fear of accepting permissions
* Because Claude can handle all the actual Git usage, you just need to define some simple good defaults for using it. e.g, always commit and push, share with all the organization, etc.

Write/tape some 10-minute immediate-value tips for AI noobs

* The default paths for getting people started with AI require a significant time investment and have an unclear productivity reward. That's completely unnecessary
* There's lots of tips that people can immediately adopt and benefit from. Things like:
  * Use WisprFlow
  * Make your context easy to use (connect Drive and Asana)
  * Use AI Agents like Cowork for long-running tasks
  * Specify your need, ask the AI for implementation options
* A video that introduces the concept and gives a real use case for it could bring lots of value here

Create an AI power users/AI enthusiasts Discord

* I've heard that Discord is much better than Slack for high-bandwith communication and idea exchange, but I don't use Discord so it would take me a bit to get one started
* I know many people who could potentially be interested in this Discord channel, and would invite you if I trust you to have a good idea on how to set a good culture for the online space

### Clear bottlenecks for AI use at EA orgs

Address EA org concerns on connecting Drive/Slack/Connectors to Claude

* Most EA orgs probably have the same concerns around this. They don't want employees to expose their data accidentally, for companies to use it in obscure ways, or to expose themselves to litigation
* You can talk to a few compliance / security people (I can introduce\!) and get a sense for their concerns. Do the research, and come back with opinionated recommendations
* Main goal here is to make it frictionless for company staff to connect their Claude to the data sources they'd like to use most

Create an identify-automation-opportunities-from-Drive+Slack+Asana Skill

* Pablo Stafforini got lots of value from running that for Epoch. You could borrow from his repo (I can share\!)
* Main challenge is figuring out the minimal friction way to run that skill as a person external to the organization. Should you hop on a call with someone technical in their team? Get an email on their domain?

### Make useful AI tools

Create a repo to pull new frameworks/libraries and tools from Twitter

* ….

Publish a Slack bot to do a daily/weekly AI tips and tools round up from Twitter

* have one already, could be better

# Big Wins

---

# What I'm looking for in my next manager

Source: https://myea.blog/my-manager

[Alejo Acelas](mailto:alejoacelas@gmail.com). [alejocelas@gmail.com](mailto:alejocelas@gmail.com). Whatsapp: \+57 316 680 37 46

**Compensation:** Whatever you think is reasonable, and possibly more if I can get the money for it.

Mid-June I'll be starting possibly my most ambitious project ever. I've noticed that I love having a manager, and I've had many great people who have served as managers for me so far. But, same as dating, if there's the potential for a great long-term relationship, you should invest a lot in exploring. So I'm doing this document to prospect for great managers.

You don't have to be a great manager now—and actually I think you shouldn't be, because if you are then you probably have great people that you're busy managing. But you should have the potential to be a great manager. Maybe you're untested but you think you could be great, and you think we could get along. I'll write some things to help you figure out if we could get along and enjoy making each other great.

Here's what I think I'm looking for:

* **Quick and [responsive](https://sashachapin.substack.com/p/what-the-humans-like-is-responsiveness).** I like [moving fast](https://docs.google.com/document/d/1WabYjxXmObvPbnyFJCZ0v1Vt-jQ_fadyI3BmETfmWBE/edit?usp=sharing), and as one of the people with the most context on what my project will look like, you'll probably be extremely well suited to help me bounce around ideas and take strategic decisions on a day-to-day basis. I don't want to wait for a response. Many times I'll try to ask a question in a way where just eight words in your reply should suffice for me to go ahead—but I really do want to keep close contact.

* **A warm listener.** I love talking. My friends know this, everyone who has met me knows this, and I just enjoy it enormously. But I enjoy it even more with people who seem receptive—in a certain hard-to-specify way—to what I'm saying. Where I feel like I can just pour my mind into them and trust that through that process we'll figure out what's worth doing and how to get it started.

* **Loves unprompted feedback.** [If I worry](https://docs.google.com/document/d/1_IOFvpchwDP2l27QOuSaGbeX0Vx1FJ-DbbRewa88nGM/edit) about timing or framing my feedback, I often give much less of it. In fact, over the past three weeks I've been incredibly wild about just spitting my mind out to people nearby. Many of them actually don't quite like this and would rather me just keep my thoughts to myself. You should not be like this if you want to improve at being my manager. And if we want to get along together without you wanting to strangle me to death, probably it's better if you just enjoy some of it.

* **Unconditionally trusting.** You have a vision of me, of what I could achieve if I put my all into it, and you trust me to deliver on that mission. It doesn't have to be necessarily grand—you don't have to believe the delusional stuff I feed myself to keep me going—but you do need to have some image that is further and larger and grander than what I am. And I want to feel that no matter if I screw up, no matter how insecure I'm feeling today, you will hold me to that vision and cradle me into achieving it.

You don't have to fulfill all of this to be my manager. These are just some directional pointers.

If you think you're unusually strong at one or several of them, I would love to have a call: [calendly.com/alejoacelas/30-minutes](http://calendly.com/alejoacelas/30-minutes)

I think the upside here is really high, so I want to have an extremely low bar for just meeting people and figuring out our fit. You don't need to have legible evidence of being a great manager yet—you just need to have some hope and desire to do it well, and maybe feel something rise in your chest as you read through the previous bullet points. I've already taken bets on people and I feel great about taking bets again, so please do reach out\! I want to meet you.

Here's my blog if you want more context on me: [myea.blog](http://myea.blog)

---

# On using pride as fuel

Source: https://myea.blog/pride

A friend died this past weekend. There's a few things that need to be done to increase the chances that the people who cared for him make it out of this okay. I've been doing a few of those things myself, and one of the motivations I'm relying on is that I like the image of me who's able to help them.

I like to imagine going to the job interview at BlueDot and telling those guys that, in a hard episode of my life, I still managed to keep a clear head and care for those who needed it.

This doesn't feel like a noble motivation to have, and in some ways it's not aligned with the ultimate outcome of having things go well for everyone involved. At some moments I've been too eager to share details about the situation, to share that a friend died, with people who would otherwise, I think, have a better time if I didn't disclose this (this post a very explicit example). But I still disclose it, because I'd kind of like them to appreciate my heroic struggle. I'd like them to see how I have a reason to suffer and still I managed to muddle through.

I don't quite endorse being fueled by pride, but at hard times like this, I'll take any fuel I can find.

---

# The giving tree

Source: https://myea.blog/giving-tree

There was this tale from my childhood that I remember feeling strongly moved by. I just searched it up — it's called [*The Giving Tree*](https://www.lynnjohnston.com/wp-content/uploads/2016/04/TheGivingTreePoem.pdf). I recommend reading it now, before you continue. It's barely a page long and it's really good.

Not sure if I thought this then, but I think it now: I want to be the giving tree. The story's a bit too melancholic — the tree sounds like it's all the time waiting for just a single boy. But I love the idea of giving your all to those you love. I love the idea of taking pieces of your body and turning them into good. I love the idea of giving away everything until just a stump remains.

I think I learned much of this from my father. I remember seeing his strong devotion towards me and thinking that I wanted to do that someday. I'm not sure I want to have children of my own. When I think about it, I feel like I want to father children of all ages. You could be my son or my daughter if you wanted to, and I'll welcome you and give you apples and branches and my trunk and my stump. And I'll grow more of them, and I'll collect many such that all the children in the world can rejoice in my shadow.

Not sure if this is the best way to go about utilitarianism, but right at this moment it feels just like what I want to be.

---

# Lay yourself bare on the web

Source: https://myea.blog/lay-bare

The arguments for writing more and posting it publicly are good and numerous, so I won't rehash them much here. Personally, my main aims in blogging are:

1. Discovering (at least one?) great long-term partner
2. Discovering great friends and collaborators
3. Having better conversations about the topics I care about most

Just a month after I started I say I've had wild success on C already (I chatted a bunch with my crush about things on my blog, and had titillating dinner conversations with many others as well), deepened many of my friendships in a way that counts for B, and have made some conceptual progress on my way to A.

If that's not enough reason for you, mind as well that I'm vibecoding a new social media app, and there the only way to feature on people's feeds will be to write earnest pieces about things of either deep or broad importance.

If my app ever gets popular, you'd better have practiced writing earnestly online, otherwise you might find the path out of the permanent EA underclass has much narrowed by the time you start trying.

---

# My pattern of disregarding people's immediate preferences

Source: https://myea.blog/my-bad

Today I was helping Robin time-block making a summary of their research output of the week. They had been feeling shame about being slow at research and not being able to package their outputs for feedback and iteration. So I proposed that they should try taking half an hour to dictate their free-form thoughts, then ask Claude to do some very light polishing on them, and then share directly with their supervisor, or with more people in their field with context on the questions they were discussing.

They seemed excited about it, so I left them for 30 minutes. When I came back, they had dictated north of 2,000 words on their output for the week and were feeling pretty happy about it. The last step was just doing some light polishing and sharing.

After doing the light polishing, they started to act visibly anxious about the idea of sharing. They looked at the output and were like, "oh this is not precise, this is really hard to skim." Usually when you want something to be readable, you want the first sentence to summarize the whole piece, and the first sentence of each paragraph to summarize the paragraph — and this was definitely not it. This was a narrative account of what was on their mind.

I insisted that, given how much resistance it created in them, the idea of trying to do these weekly summaries — that they should bite it and just send what they already had without doing any more polishing. Because otherwise they would not trust that they could just quickly take 30 minutes from their day to share intermediate outputs, and that was a valuable part of this exercise.

But still they were hesitating a lot, trying to find ways here and there of maybe editing the text, maybe composing an email, maybe just saying to their supervisor "I'm sharing this but this is bad, so I should just polish it more before sharing something like this again." And I pushed back on each of those points. I was like, "You should let your supervisor know the trade-offs involved in your sharing more polished outputs. If it takes you eight hours to polish what you did in a week, then your supervisor should know this so that he can appropriately decide where he thinks you should stand on this trade-off."

And so, seeing their resistance, an idea occurred to me (though I didn't share this with them in the moment), so I asked if they would share their current output with me. I said "I'll go to my computer, do something, I'll be back right away." What I did was make a copy of their Google Doc, such that they couldn't remove access to the doc later. I searched for the email of their supervisor — I had already seen their name — and I scheduled an email to be sent one hour after the current time, saying something like "My friend Robin is having trouble sharing quick outputs from their research. I thought it would be helpful to share this," and the link to the document they had already composed. And I hit schedule send.

Then I came back to the main room for another round of insisting "you should share this intermediate output, just to gain information for what your supervisor thinks of it." After maybe a few minutes of panting and angsting, they just wrote an email — a short, clean email explaining the situation to their supervisor: "this is a quick time-blocked output, I'm curious to get your thoughts on whether I should publish more." They added the link, and they sent it. I was really proud of them at that moment.

And as they readied themself to go sit in the closet and feel anxious about what they had just done, I reflexively shared what had just happened, out of a general candidness habit: that, in the moment that I had removed myself, I had arranged that plan to hit schedule send. They were furious.

Just the day before, we had talked about how strongly they dislike paternalism, even when that paternalism is directed with knowledge of a person's longer-term preferences and meant with the best of intentions. I think I agree with them that what I did was clearly disrespectful, and that given what I knew about their preferences, I just don't endorse — and didn't endorse right after doing it — having gone against them in this case.

I'm not sure exactly what would have happened if they had decided that they didn't want to send the output without further editing. My quick guess is that I would have just shared the schedule send plan and disarmed it before it triggered. But still, I was so eager, with only a few seconds of thought, to completely override their immediate desires about the situation. And I think that's bad. Even on a strictly utilitarian picture I think that's bad, because it gives you license to act in a bunch of cases in which you're not knowledgeable about the costs and trade-offs involved in the decision. In this specific case, I knew practically nothing about Robin's relationship with their supervisor, other than just knowing that they're a supervisor and they work at a university, I guess. So I had some internal picture that the costs to my actions were limited. But this was based on paltry knowledge of the situation and done with only a minute and a half of thought on it.

Robin insisted a bit that I should not only not do this to them, which I agree, but that I should not do this to anyone. There I'm a bit more doubtful. I think there are cases where social norms — they suggested that my guide for this should be social norms. When do social norms say we can override people's preferences? I'm actually not too familiar with the social norms in this context. I proposed a scenario where you have a spouse and you've been fighting with them and the situation starts to get so nasty, and you ask your father to please not meddle in this situation, to please not talk to your spouse, you don't want their intervention in any way, this is something that you want to handle on your own. I think there are some cases where the appropriate thing for your father to do is to still go talk to your spouse, even despite your immediate preferences not to. I think Robin agreed with that; they just insisted that this was not a case of family.

So then we tried finding some clear rules, because I cannot just choose to be virtuous, I don't know what virtue looks like in practice (a very perceptive friend recently commented that I have absolutely zero common sense. I think he's onto something). So we tried proposing some rules, some triggers that I could use to recognize situations in which I should strongly default to being respectful of people's immediate preferences. What we landed on after some discussion is that I should default to following people's immediate preferences and only depart from that in cases where I had thought in detail and at length about the benefits and costs of not adhering to that rule.

I feel generally content about the rule. Not sure I'll be able to hold it for small interactions — for example, a joke that's slightly spicy with a close friend. I actually did one last night, asking my friend if he mistreated women, in a call with a woman there. But I think at least for most substantial stuff, I feel good about this rule and really positive about my prospects for enforcing it.

Still, I think this whole situation points to a deeper trait that has so far characterized my personality, which is being too utilitarian-leaning in ways that often conflict with the virtue of respect. And here's another illustration of that. I commented to Robin that for the first days of the Sam Bankman-Fried scandal, I was unusually sympathetic to him and his situation. At that point in time it wasn't clear if the whole thing had come out from a general fuck-up — from a failure not only to adhere to standard corporate practice and to abide by the law, but also as a failure to implement any form of skillful utilitarian reasoning. At that moment I thought there was a possibility where he just very deliberately, with a clear head, chose to commit fraud as a calculated bet to collect more funds that could be donated towards EA causes.

I think — knowing the details now — that was clearly not the case. But at the time it was at least entertained as a strong hypothesis, and at that moment I didn't think that what he did was right, but I could see myself being tempted to have done the same in his situation. Because fraud laws, they're just such a removed proxy for morality. They're trying to do the right thing and setting some boundaries for it, but they include much less detail about the specifics of your situation. They were not made with extreme altruism in mind. And that reasoning crosses my head.

I've also read the cases for integrity and lawfulness for EAs and I very much endorse them, but still, when the stakes get high enough you start to think maybe I'm smart enough to have it all. And now, having seen how everything went down, I'm substantially more skeptical of that reasoning. But I can still feel the instinctive pull of it.

Most of the rules that are out there, especially social conventions — I'll readily break them if I think that will lead to better utilitarian results. And that's because I hold no intrinsic aim for adhering to social conventions. I think they're a good default for most people's behavior, and they encode some knowledge about things that have made interactions go better in the past. But I'm extremely willing to rework them after having thought for a few minutes about what would be best. This has already led to significant issues, and indeed possibly has strained my relationship with a friend that I care for a lot. (Another case of a screw-up that I don't endorse but that my disposition makes me too willing to incur on first occurrences.)

So what's the difference between those and state laws? For example, would I — I actually thought of discussing an example of me breaking state laws, but I think that wouldn't go well on my record if discovered, so I'll abstain from it.

And in some other cases, like overriding Robin's preference for chatting with their supervisor, there's not a clear hard rule for how to behave. Indeed, if I were helping a perfect copy of myself, I would endorse them overriding my immediate preferences. What we ended up agreeing on with Robin is that I should just ask someone else from the EA community about when they think it's appropriate to override immediate preferences or not, and maybe make some explicit commitments if together we estimate that the preference for not being overridden is fairly widespread.
I'm posting this here partly because I want us to go back to the pre-FTX levels of ambition. I think there was something true and valuable there. I think aiming high is morally correct. But if you're going to do so, we need safeguards to hold our worst human tendencies for doing self-justifying naked power-grabbing. That's part of why I wrote a previous piece on honesty and candidness. Indeed, I think candidness served me well in this scenario, as I was immediately ready to share with Robin that I had gone against their preferences here.

And I'm very interested in finding other clear rules such that I can immediately recognize cases of wrongdoing in myself, and find ways of committing publicly to those rules and letting others quickly know if I'm deviating from them. I have this horrible fear of, in my pursuit for the greatest good, landing in not the greatest good at all and actually quite possibly doing lots of harm.

So I asked Robin to commit to tell me in advance if they ever consider giving up on me — if they ever think that I've strayed too far from what's good and virtuous, such that they wouldn't even bother telling me, advising me, alerting me so I can try changing what I do or who I am. I'm happy to report that they currently do not think they should give up on me, and that they will tell me if they ever consider doing so.

If you want to be alerted in case Robin considers giving up on me, you can leave your email here.

(The first draft of the post didn't quite capture how much I care about Robin, and how much I feel the pull of trying to be a better person just for their sake. You can get a sense of that from reading my previous blog posts, many of which are directly addressed to, or came out of an explicit conversation with them. I chose Robin as their alias from Unsong's character, where Robin is the ultimate impersonation of virtue directed toward utilitarian aims).

(See parallels with HPMOR's discussion of paternalism and ultimate responsibility. e.g., [Chapter 75](https://hpmor.com/chapter/75). I hope that I'm not nearly as paternalistic and annoying as that Harry is).

(Now I have much more respect for [the scene](https://www.youtube.com/watch?v=WK-kQou1ecM) in Zach Snyder's Superman movie, where Superman's adoptive father signals him not to intervene to prevent his violent death to a tornado. I read it as “no matter the stakes, if you have to choose between power and virtue, choose virtue”. Not that I think it's logical or correct, but I can feel its emotional pull now. What kinds of morals would you create for a man that can literally overpower everything and everyone else?).

(These last two parentheticals feel like one compelling answer to my longstanding question: what's the value of fiction?)

---

# Maybe your weirdness is a mental health condition

Source: https://myea.blog/weird-health

[**https://claude.ai/share/5b7babb6-8fd1-4329-926a-ce403320962a**](https://claude.ai/share/5b7babb6-8fd1-4329-926a-ce403320962a)

I was surprised recently when I discovered how stereotypical I am. If you search for the symptoms of bipolar II disorder, a bunch of those just look like the description of my last three years. Similar thing for descriptions of people with low cognitive empathy. I read them, and I was like, yep. That's me. And it's not exactly me — there's some social situations in which I'm unusually perceptive compared to the median person or the median EA. But still, I think as a rough cluster, they managed to capture some wide-ranging traits of my personality.

**Mental health conditions are not necessarily bad**

There's this common joke about people in SF being extremely autistic, many of them. And overall, I hink they look like they're doing well, at least by their own eyes. So you shouldn't take this mental health label as something to be necessarily ashamed of.

The way I'm thinking about is: for sufficiently consequential and widespread groups of traits, psychiatry is the field that tries to kind of group and classify them and find ways for dealing with shared downsides of possessing such traits. So when you discover that you share a particular trait or cluster of traits with, like, a good chunk of the population, and that you're lucky that psychiatry has done the work of researching this condition — I think that's \[the\] way you should treat it. Just add a new reference class to the list of things you're planning to compare yourself to in order to take better decisions.

And how do you find these things? How do you notice which of your traits are mental health conditions, or are well understood or analyzed through that lens? I think you just look for things that are sufficiently consequential in your life. And then just chat with Claude about which of those seems well studied in the psychiatric literature.

So in my case, I had this pattern of periods of really high intense energy, followed by months of aimlessness and fatigue. I spent a bunch of the time when I was dealing with those episodes just trying to understand and figure out patterns that kind of influence my mood and behavior during those episodes. And I think I made substantial progress on that front. But I also think it was silly to not be curious about any existing literature for how people deal with similar things.

I considered at different points that I might be depressed, but I flinched away from \[the\] idea, partly because that meant acknowledging in some form that my problem might be serious and that might lead me to update negatively on the things that I would be able to achieve in the near future, but also because I just didn't know enough about the psychological literature. Like, I just \[thought\], what conditions are there? I have heard of some of those. Maybe I was, like, most familiar with depression, which is something that many people in my circle expressed having or have expressed having. And I was just like, I'm not sad like them. So probably not depression, probably not anything like that.

And turns out that being bipolar is a really different thing from what I thought. Yeah. I don't have, like, different voices in my head asking me to commit mass murder. Well, I do have a voice like that, but we both know they're joking \[I dictated this out loud in the park, and now I'm feeling a bit interesting\]

Turns out that bipolar II is basically what I just described: it's marked by shifts in mood that alternate across your life. And so the thing that I want to prompt you to try is just: if some aspect of your emotional makeup is being highly consequential for your life, whether in a positive or negative way, just consider searching in the literature and finding out if there's anything you can learn from that.

---

# EV application - Emergent Ventures

Source: https://myea.blog/ev-app

**Tweet version**
EA is not taking AI seriously, for no good reason. We should be hurling ourselves to find how AI can promote the outcomes we want (at orgs, in new projects, individuals).

I hope that by relentlessly pushing for this I can make a change, though I haven't settled on a concrete strategy.

**About me**

I'm extremely non-complacent. I feel EA is getting crufty, and I want to change that. We all think AI is important, but most do \~nothing to leverage it better for their work, and I don't think we've got the right emotional disposition to face what's coming.

I think epistemic modesty is at some level basically correct, and I feel really uncomfortable taking sides in disputes with smart, thoughtful people. I find much of EA to be too adversarial to crowds that could be possibly conceived as epistemic peers.

I found out about EA in 2016, and since then I've been on a gradual, but possibly accelerating process of reshaping myself to be more utilitarian. My latest target is taking advantage of my feelings of strong care towards friends to help them massively succeed. I've started doing life coaching and I feel some of them really appreciate it.

I feel angry that I did not try to be more ambitious and determined earlier, and I'm finally coming to be that way, much in part to help from mentors who believed in me. I want to do that as well, and I have this intense almost manic energy that has served me really well to convince people that they're worth it and they should try hard.

Living in Colombia, and having been convinced of utilitarianism at \~15 by its Wikipedia page, much of my beliefs and dispositions growing up were shaped by aspiring to be what I thought EAs were like, from reading blogs and listening to the 80k and Julia Galef podcasts. I think I acquired unrealistically high standards for myself through that process, which has been pretty beneficial.

I've worked on like, 5 different things since I graduated uni 2 years ago:

* AI Safety research (mech interp) with mentorship from Marius Hobhahn, before he got to be CEO of Apollo research
* AI policy think-tank-type stuff, which I left because I wasn't sure supporting the EU AI Act was net good, partly from reading a bunch of Dean Ball
* Running \~7-day bootcamps on ML and AI Safety for Latin American young students and professionals
* Doing footnotes and quick research tasks for Ben Todd's new edition of their career guide book
* Building an AI tool to automate customer vetting for DNA orders, and advising the construction of standards for what every DNA provider should do to vet customer orders

**My consensus view**

I think much (most?) of my personal growth comes (and will come in the future) from deeply loving a small set of people (such as friends, family, I'd say mentors as well, but that seems a bit less consensus).

I think much of our brain is very people-shaped, and there's some liveness to having ideas embodied in an actual person, which is even stronger if that's someone you meet or feel close to.

I don't have a good grasp of how raising aspirations works, but I think of it as some kind of internal narrative that shapes much of the direction and even magnitude of what you try to do. I feel that narrative can be extremely malleable as well, at least by some of the people you know, at certain stages of your life.

Also, a lot of self-efficacy habits you get from watching great people do their thing. Most often you don't understand what great is like until you see it right in front.

Much of what we are we cannot see through introspection, and when we discover it, it's through friends that we preserve the momentum to act on it and become someone else.

I've been infatuated with a girl for the last few weeks and written lots of blog posts to deal with the rejection (myea.blog). I had not until now considered how much of great literature may have been motivated by unrequited love.

**The idea**

AI will bring a century in a decade, at least for some areas of technological progress. This will have societal consequences, and we can make things better by tying the capacity of the institutions that deal with those consequences to the rising tide of AI capabilities.

I think that idea is very unoriginal. Maybe my only insight is that I think I can actually do it, at least for a fraction of the relevant institutions (EA orgs come to mind, but not sure if I want to focus mostly on those longer-term. Jaime Sevilla recommended against it).

My impact story is:

1. EAs like having an impact
2. They could have a much greater impact by reshaping what they do to seize AI capabilities
3. There's no structural reason why they're not using AI better. Whatever it is, I will be able to solve it once I figure out what it is

(3) seems like the most brittle here. Maybe EAs are not what I think they are and they won't rise up to the challenge. Then my project will become changing the soul of EA instead.
I've been travelling to Europe, and not always finding a friend's house to stay. I have maybe a 6 month runaway at my current spending pace (if I'm willing to eat my stocks as well). In ramping up this work I'm starting to get more paid gigs, at consulting fees, which should alleviate the pressure.

I'm happy to receive anything between USD $100 and USD $5.000, though I think the money would probably do more for another of your grantees.

**Current work**

Mostly I just do whatever I feel like on any day, trying to pick up opportunities for learning things deeply and spending time with exciting people. Recently that has involved:

* Setting up AI coaching calls with \~4 people from Ambitious Impact
* Staying two days with a few people I like from the Lead Exposure Elimination Project
* Creating an EOI for an undefined joint venture with Pablo Stafforini (RA for Doing Good Better) and Luca de Leo (incredible AI power user, now at 80k)

I plan to ramp up having calls in June to figure out the space needs better. Some things I could imagine myself doing are:

* Meeting CEOs of EA orgs to ask about their needs, then pitch them on how AI / AI power users can fill them
* Building shared AI infrastructure for EA (e.g., a Google Docs you can easily use AI with)
* Driving up enthusiasm for AI uplift among early-career AI power users

One of my most valuable assets is having a shared Slack with Peter Hartree, who I absolutely love and admire. We also have a 20-people Slack channel of power users, and people there are super friendly and kind.

---

# After-date surveys may be a bad idea

Source: https://myea.blog/after-date

*Epilogue to the "You should consider dating me" post I haven't finished yet*

When a date is fun, sending a quick forms asking "how likely are you to recommend dating Alejo to a friend with similar interests \[0-10\]?" may land like a cheeky fun exchange.

However, when the date goes poorly (like when you took 20 minutes of a first date to detail your half-apologetic desire to subjugate any relationship outcomes to the ultimate and only worthy goal of improving the world in a utilitarian fashion, all out of a pathological craving for mutual candidness), the same survey may land a bit differently, and the act of sending it itself might be uncomfortable for all parties involved.

I weakly recommend against them.

---

# On caring

Source: https://myea.blog/caring

I started thinking about what caring means after a friend told me they're frustrated with what they're doing right now. They don't know if it's the right thing, feel aimless and demotivated, and are wondering whether to try something else. I could feel their mood tank noticeably from this uncertainty. My friend is an EA, of course; many of my friends are, and I couldn't help but feel the intensity of it. I immediately empathized with their sentiments and found it easy to provide some reaffirmation to many of the things they felt, but I also felt this intensity behind it that almost choked my throat and reminded me of that poem "Do Not Go Gentle Into That Good Night." There's something I want to rebel against, really strongly, about that state of affairs—its determination, its resolve, its caring—in a way that was beautifully articulated by Nate Soares here \[link\].

I started pulling at that thread because I think this feeling of caring might be quite special. When I think of the emotions that are at a deeper level, or the emotions that have caused me to change my behavior in persistent ways, to take decisions, to take commitments, to act in the moment and change course and take a day to do something I didn't expect—a bunch of those seem to come from an emotion like this. They seem to come from caring.

## **A Tentative Definition**

Let me offer a definition and see if it helps any. I'm interested in caring because it seems to be a persistent trait. The things I care about, at some level of abstraction, I feel have not changed much for me throughout the last eight years of my life. It also seems to be action-guiding—but more than action-guiding, sometimes action-*forcing*, like it *has* to. It seems to pull heavy weight in terms of what I do.

Those first two characteristics you could associate in general with motivations or personal traits. There's a broad space of things categorized by those two criteria, and I think that space is really interesting. I've asked myself many times what it means to know someone, and I think some of the answers lie in factors that fall under those two criteria.

The third aspect is the phenomenology of it: how does it feel? There's an emotion that at least precedes, temporally, my episodes of taking resolve or taking action to help a friend, committing to a certain cause, or endorsing a certain set of ideas—like utilitarianism—that have had long-lasting effects on what I do with my life.

To point at this feeling, a few written samples might help. Nate Soares on caring made a whole series about it; \[this post in particular\](link pending). Joe Carlsmith I think is one of the greats—not only showing you arguments for a thing but showing you what it *feels* like to care about that thing. That caring may not last in the way I've been defining, but you get an emulation; you get to experience, at least briefly, what it would look like to care and decide whether you want to make that disposition a long-lasting thing in you. There's also this [post I made on honesty](https://docs.google.com/document/d/1_IOFvpchwDP2l27QOuSaGbeX0Vx1FJ-DbbRewa88nGM/edit?tab=t.0#heading=h.98baigsv7070), my totally undispassionate defense of it. When I drafted it, I was experiencing an intense emotion of the sort I'm describing. When making edits, it was much harder because the intensity of that emotion was not as high as during the first draft.

## **Care as Explanatory**

One question is how much care explains differences between people—can we use it to characterize a meaningful fraction of who a person is?

Most people do not care about doing good in this scope-aware consequentialist manner. You can tell a lot about someone by how much they care about this, by how much these considerations affect their life. Even within EAs, I'm still unsure how many differences there are between people who all care about this ideal.

So what fraction of someone's actions are explained by or downstream of this feeling? In my case, a large fraction of what I do is downstream of this feeling—in good part because of my utilitarian leanings. "Leanings" is too soft a word, hahaha. My utilitarianism, I think, originally came from an argument like: I wouldn't let my sister die. I feel earth-shattering anger when thinking about that—a pressure on my chest, a constriction that motivates me and rouses me. And I feel I'm not in a unique position; people feel that way about their sisters and children and spouses and friends. Their care for their loved ones, the care you feel for your own well-being or even just your direct experience of joy and shame and sorrow—I feel that's as important and action-relevant for me as the care I feel for my own joy and sorrow. The emotions of other people should compel me just as much as my own.

That's how my utilitarianism came about. A good chunk of it is just some form of universal care.

## **Care in Action**

Utilitarianism has made me do a bunch of things, most of them I've done gladly. But I feel care acts even at a smaller scale. Let's take an example of how I behave in settings where I'm not specifically constrained by what utilitarianism might say. Recently at an EAGx conference, I spent a bunch of time talking to early community builders and group members. I felt compelled to book lots of those meetings and try to meet them all. When I reflect on this experience, there's a feeling of care there that's pretty intense and leads me to act in certain ways. In one-on-ones it leads me to ask questions, be curious about what they hope for, what they're trying to do. Even though in many cases I think their goals aren't precisely the beautiful quantitative utilitarian thing I myself aim for, I still felt compelled to help them pursue their goals better. There's something interesting and rich about caring about helping people pursue their goals.

There are also previous examples of how I am with my friends. One way my friends—many of whom are more normie than me (I don't like that word much, but it fits)—might characterize me is how willing I am to break convention for the things I care about. Maybe I'm not that unusual in this respect compared to EAs who've been in the movement as long as I have (10 years already\!). But in my case, care seems to have strong explanatory power.

## **Care and Who You Spend Time With**

Another question: how much does care matter for the people you want to spend time with?

Some small examples. I find interactions with my paternal grandparents to be a bit bland, because a bunch of them are shaped around expressions of affection—how much they care for me—but the way they care is undirected. My grandfather says he's proud that I'm working to do good, because I've told him I'm working on bioterrorism stuff out of my desire to do the most good. Before that I told him I was working on AI safety out of my desire to do the most good. Before that, I told him I was studying hard in my mechanical engineering undergraduate degree to eventually get better at curing malaria. (I think he was more on board with the last thing.)

He's been similarly proud and supportive of all of these and not that curious about the content or outcomes of my efforts in any of them. He's just proud that his grandson is trying to do good. He's also proud of my cousin who's a personal trainer—a really great personal trainer, who might actually be better at the thing he does than I am at mine. That's such an unopinionated form of care that I'm not sure I enjoy it.

Similarly for my dad. He seems pretty adamant about reinforcing the message that I can and will achieve great things, but he's always unopinionated about what those great things might be, and not that curious about the content of them. Having conversations with him about the things I care most about—he's trying to be helpful, but it's a weird way where most of his advice isn't directly helpful. What helps is when he asks clarifying questions, because saying things aloud about the things I care about most tends to help a lot.

Part of this might be that helping people with big life decisions is hard. But I think there's also this component of unopinionated care that makes it harder. If you don't care quite about the same thing, the goal you're trying to shoot for in those conversations keeps going off-target, and the advice you give doesn't center on what's most helpful for the other person.

## **Care and Selecting Partners**

Another example that's emotionally resonant for me: selecting partners.

I broke up with my ex after three years, and I struggled to articulate why I thought that was a good idea—as good of an idea as breaking up with someone can be. One way I can put it: she cared *for* me but not *about* the specific things that were most important to me. When we talked about my plans, she seemed very interested in how they affected my emotional well-being or my short-term income, but I happen to care little about either (while I endorse not caring about the second, not caring about the first has brought its issues). In our conversations about the things I care about most, she could empathize and was often a helpful presence, but she couldn't steer or guide or serve as meaningful company toward the things I was aspiring to.

Same the other way around. As far as I understand, one of her main goals in life is to keep making art, to make some kind of art she approves of. Despite some talking about it, I never quite understood what the good art she was aiming at looked like. To compare: I think I understand Paul Graham's idea of what makes good art better than I understood my ex's idea. At least I feel more drawn to engaging with Paul Graham's idea of art than I felt engaging with hers. Part of my not understanding was probably downstream of lack of sustained interest.

I never quite understood what she meant by good art and never felt quite excited or had the right emotional disposition to support her in that. And I think that's important. In a relationship, you want someone who shapes you in some form.

I'm not sure how widespread this thought is in EA circles, but talking to friends about relationships, many people have this conception of wanting a partner who supports them where that support doesn't come with strings attached—where your partner isn't opinionated about what the best parts of you are, the parts worth preserving and striving for.

I don't like that. I have a strong conception of the good, and I've seen myself deviate from it or lack the courage, ambition, or resolve to pursue it. I want someone who pushes me toward what's good.

I'm pretty opinionated about what's good. I'm happy to be shaped by other motivations and learn, but I do think long-term partners should push you, and I want to push my partner as well. I've talked about pushing my friends toward a certain shape, because I care, and that's a strong feeling. I want to act on it.

One important thing about care, which I should have mentioned earlier: for the most part, I *endorse* this feeling. I endorse acting on it. Right now, I endorse pushing people to change, at least for some directions of change—even as I try to be deliberately respectful of letting them make their choices and decisions, and enthusiastic about people making their own choices without interference, other than by trying to change their conceptions of the world through persuasion methods they endorse.

## **A Metaphor**

My grandparents' care feels a bit like a blanket—pretty wide, pretty soft, falls on something but doesn't change it. The care I'm most excited by feels sharp, like a long stick or a sword, that can push through the crust and break things and shatter and reshape. I'm much more interested in that form of care. You might call it destructive, but I feel it's pretty much constructive. It's riskier. But oh god, do I want some of it.

## **What Does It Mean to Know a Person?**

Another thing on my mind: I'm excited that I'm starting to make progress on this question I've had floating around for years—what does it look like to get to know a person?

Taking the first two prongs of my definition of care—persistent, action-guiding dispositions—and adding a third element, the phenomenology of it, how it feels, associated with an emotion... that was productive. We found something that helps us explain ourselves and the people we want to spend time with. I'm really curious what other three-clause combinations we can make to say useful things about who people are.

Another angle: what are those motivations or predictable patterns in our behavior that are *not* care-based? When there's nothing more important to do in a conversation, or when I just want to spend time with someone, I try to be funny. I invest a lot of mental effort on it. My mind has been consistently shaped to pump funny things into the air, even when I'm alone I often do this: "Oh my god, this is so funny, this just happened in my mind." I've had this for more than 10 years, but that's not care-based. I don't feel this same warmth about that aspect of myself, even though I endorse it in some form and want to preserve it.

## **Caring for a Person**

What does caring *for* a person mean?

Robin told me they have changed, and I don't know how to feel about that. Part of me is like, "We've been apart for almost a year and I've only met the previous Robin, and I love the previous Robin. If you tell me you've changed, I mourn this a little—there are aspects of you I adored that may not be there anymore." But I also think my excitement for Robin remains as high as ever, if not higher. I'm going to see them in a few weeks\!

What are the things I actually care about Robin? In acting on my care, given that I'm so into this opinionated care, it's hard for me to care for Robin in the fullness of their being. I often end up acting on caring about specific aspects of them. I care about their well-being—if they tell me they're sad, I think and hopefully act really hard so they can start feeling better. But there are many things about Robin that might be core to who they are that I don't know about and feel currently unable to quite care for.

## **Care and Proximity**

Care seems strongly tied to proximity—even for those of us who repudiate giving proximity important weight in our moral decisions.

A clear example: my sister, the oldest of my siblings. I don't live with her; I only visit once or twice a year for a few weeks. Most of the time I'm not a strong presence in her life. We have calls on WhatsApp occasionally, but I don't proactively try to influence her or spend time with her.

What I do with my two youngest siblings is different. I live with them, and when I'm with them, I find it easier to care—or to act on my care. I see them doing something and I want them to be a certain way. For example, I want my siblings to be able to reshape the world to their desires, and I think some skills are really valuable for that. Cooking is a great social skill; I'm trying to learn it myself, and I'm inculcating that in them. The other day we made guacamole, and my mom was like, "You shouldn't give knives to children that are 5 and 7." I was like, "Mom, I'll be careful." I feel really strongly about letting children take risks so they can get feedback from reality, learn stuff, feel independent, and *be* independent.

I act much more on these impulses. If I want to tap into my feelings, I can muster similar feelings—even stronger feelings—for my sister than for my siblings. But proximity means I end up doing caring actions much more for my siblings who are closer to me (but this switches quickly as soon as I travel to meet my sister\! On the same day of the flight I basically stop caring for my younger siblings).

Similarly, I spent so much time away from Robin. We had a call one day when they told me they were sad and depressed and doubtful about what they were doing, but it didn't weigh on me as much back then. Now that we've had a similar call and I'm closer to visiting, the feeling is much stronger. It has easily been the top thing on my mind for the past day.

## **The Impermanence of Care**

Another aspect of caring I've thought a lot about is its impermanence—proximity not only in space but in time.

I've often been haunted by this: people I felt so strongly toward, that I wanted to rip out my ribs and offer them as tribute for their amazingness—I've had many of those people in my life. But I know from experience that once they leave, once we don't see each other like once a week or once a month, I just won't be there pushing for their well-being. It feels like my care is not real, because I won't act on it. Once they part, I don't influence their lives. I won't proactively try to.

I have trouble expressing love and affection toward people, because what is love worth if you don't commit to changing their lives and making them great and good and happy forever? Part of me feels that's the only form of love that actually deserves the name. In some way, that's the love I try to express through my work. But I don't have that love for most things. Even my closest friends I only see occasionally, and I'm only there at intervals to try and make sure they fucking nail it.

There are tensions. If I want to focus on my work, sometimes I devote less time to figuring out how my friends are doing and trying to help. It feels bounded and impermanent, this care thing.

This talk about loving forever—I'm just like, "What? No." Probably the only commitment of that kind I have is with pursuing utility, and even that I'm not sure I'll be able to keep for my whole life, seeing the base rate for these kinds of things.

---

# Keep the number of people you like

Source: https://myea.blog/keep-the-number

Not those people. Well, those people as well, but everyone you like — everyone you enjoy or benefit from conversing with. The more relevant case for me is people you meet through work who you're incredibly excited to keep interacting with.

Email feels too high-friction for some of this communication, where you just want to spout an idea and see if the other person is interested in running with it. This, for me, has been one of the most valuable affordances I've acquired that has let me act on all my great ideas.

So if there's anyone who thrills you, ask them for their number.

---

# Solve each problem only once

Source: https://myea.blog/solve-once

(not sure why I only have a draft using Claude's voice)

**The core claim:**

The promise of software is to encode in its design a set of principles and practices such that we can abstract away—or just forget about—those conscious, deliberate good decisions we've made, and focus on the rest.

**The deeper principle:**

Great operations people and great founders do this. They build systems. Instead of handling the same friction repeatedly, they create something that handles it for them going forward.

**The observation:**

Watching Luca and Pablo, I've seen them build things bit by bit: a special connector to Slack, a link to connect two documents, a setup that makes sure Claude Code always has the right context. Each piece looks like gradual progress—small, even quaint—but after a month, it has compounded quite a bit.

**Why it compounds:**

Because you're building scaffolding, infrastructure for yourself. It stays there. It's a form of capital you can compound on.

**The personal experience:**

If I have an issue now, I can just fire off Claude Code and ask it, "Hey, solve this issue for me," and then have that persistent solution I can reuse.

**The thesis:**

This reusability is a key thing of the new AI era.

---

# True ambition has barely been tried

Source: https://myea.blog/god-of-flesh

*god-of-flesh*

When I was a child, I remember thinking how meager it was, the sacrifice of Jesus. I thought: if someone were to offer me the trade — a few weeks of intense torture to forgive the sins of the world, to save humanity from damnation — I would take it in a heartbeat. I wouldn't even doubt. I was unsure if I'd have the strength to endure the pain Jesus had, but I would easily put myself in his position and attempt to try it, the rewards being that high.

\[*The cross, it seemed to me, was the easy part.*\]

When I was a little older, I remember thinking that if I ever got to work at CEA (or any other EA org I was super hooked on at that time — GiveWell, maybe) I would just be fulfilled. My dreams and aspirations and ambitions would come true, and I could go to my job, have an impact, go home, and be extremely happy.

Now that doesn't quite seem like what I want. Actually, a lot of versions of working at CEA kind of feel unambitious — a bit distasteful, to waste so much potential. Not because CEA is a bad org, but because as I've gotten to meet more people working in actual EA orgs, I've come to realize that true ambition is quite unusual. A surprisingly rare endeavor.

\[*True ambition, the kind that has barely been tried, looks less like a career and more like sinking your own hand into your living flesh and reworking what's inside you to serve the good purpose.*\]

I also had this impression after meeting more closely the CEO of one of the iconic AI safety orgs. I realized they go on trips with their partner frequently, they take breaks, and they just have what in many places looks like a perfectly normal — if somewhat more intense and more work-demanding — high-income life. And that awakened some kind of sheer grandiosity in me. I can work harder than you. I might even surpass you in the impact I can have in the world.

This person is way more talented than me. I know they're way more talented than me in many of the really important traits for having an impact in the world. \[*And still* —\] if I can just work harder, it feels great. I can just choose to work harder than them. Even more, I can choose to pick myself. I can choose to look at my traits and dispositions and inclinations and gradually rework them to be maximally impactful.

I haven't seen many other people willing to make that jump. Maybe the closest one — and just a vision of it scares me — is Marius Hobbhahn, who I've had the luck of meeting through several different outlets. His work ethic scares me. You can tell every time you meet him that he's such an incredibly busy person, and still he replied in less than 24 hours on the few occasions I had a request for him. When I met him, before he was CEO of Apollo — before Apollo was even a thing — you could just book him on Calendly at any time of day. I did it once and he answered from bed, bedridden with a strong flu I think, and we still had our scheduled chat and we did work. This was shortly after I was visiting the LISA offices, where he won an indecent fraction of the awards MATS was giving to its participants — best collaborator, most ambitious, a bunch of those — by popular vote. So it's just mad respect for Marius. But maybe there can be more than one of him.

Also, mad respect for his blog. What I didn't mention in my previous blog post, but one extra argument for intense honesty or intense candidness, is that Marius has an early blog post just laying bare the details of his life and his realizations, and actually encouraging others to be as honest and forthcoming as they can with him. I admire Marius, and I admire that.

So what is this true ambition, this relentless utilitarian life that has barely been tried? It doesn't necessarily look like the calculating caricature we've been sold of utilitarianism. It looks rather like having a clear direction — some strong sense of where you want to move, and in that endpoint being utilitarian — and then the tenacity to constantly reshape yourself to move toward it.

I suspect a bunch of that process, for us beings of flesh and blood, looks like rearranging all your tastes and desires and emotional inclinations to ultimately serve an altruistic purpose. For example, I'm pretty keen on friendship, and I think friendship — and indeed a bunch of other virtues, like honesty and integrity — are deeply utilitarian. But the utilitarian version of them is much more intense than what almost anyone in the world practices.

For the past few weeks I've just been messaging my friends telling them awful things I think about them. Not so awful — I think most of them have been very well received. But just: *what's your main fatal flaw? When I met you I expected you to do much better than this other friend. Why haven't you been that successful? There must be something wrong about you. Why haven't you moved faster in the two years since we've met?* And just trying to think through that and have a conversation about it.

Now, thinking back on that childhood thought about Jesus, I think there's a more charitable and interesting interpretation of him — one where his final sacrifice at the cross was the final nail in a long, arduous, relentless utilitarian life. Jesus was ultimately a god of flesh: all-caring like his father, but limited by his human form, and the only thing he had to put forth was his full life and caring. In my reconstruction, he might have done so strategically, walking the path and meeting people and spreading the message, and it required of him intense thought and action to figure out and execute this ultimate plan for saving humanity — such that generations forth, he would still, at least to me and hopefully to the rest of the world, be a sizable, potent force in the direction of good.

*And so he, the god of flesh, reached into himself and remade what he was, that it might bear the weight of every soul.*

# Cate Hall email

Subject: Please help me become a god of flesh

Hi Cate,

Whenever work is going great, I start writing grandiose self-statements like this:

\[quote from the blog post\]

And whenever work is going poorly, I spend months doubting I could ever do anything useful again.

This doesn't seem quite healthy. I've read a bunch of what you've written on the topic and found it helpful — would you be up for a 20-minute call, or maybe a short chat in person when I'm in San Francisco in June?

Best,
Alejandro

---

# Compounding yourself

Source: https://myea.blog/compounding

What does personal compounding look like? In great careers you get better every year — your output, your impact, the things that you multiply every year in some form. And I think there's an external aspect to this, where you're matched to better opportunities and better collaborators, and assigned more resources to manage as you do better and show your colors. But I think there's also an internal aspect to this — there's something inside you that is valuable, and the world judged it worth piling resources on. And I want to understand what this is.

One interesting angle to get started on this question is: why hasn't AI automated more things? Like, it's super smart, and for most tasks that you could do… The usual way to think about how smart AI is is, what's the time horizon of a task that you could delegate to AI? And yes, AI can already somewhat reliably do multiple-hour tasks. If you look at the things that you do on your day-to-day, many of them would seem like they don't take you more than a few hours to execute. So a naive intuition would tell you that you should just be able to delegate — if doing something yourself would take you less than three hours, then you might safely delegate it to an AI, and the AI might be able to do it.

And I think for some tasks that is true, and the reason it's not happening is because people are just not giving enough context to the AI such that the AI can do it. But I've seen cases where even expert power AI users try to delegate even five-minute tasks (like replying to an email) and the AI absolutely sucks at it. And this is puzzling, right? Five minutes is much lower than the current time horizon of the best models. So something is going on here.

I think usually these kinds of confusions reveal or shed light on something that we didn't realize about how the world works, about how production works, about how work and productivity work. So here's a story — here's half a story — for what's going on, which I think has interesting implications.

I think if you do things well in your career, many of your five-minute, many of your one-hour, many of your eight-hour tasks are things that took you years of expertise to develop, years of some form of deliberate practice or intense practice to develop. Tyler Cowen says — when people see Tyler Cowen read a book, it's a breathtaking experience, I've heard, because he just pages through the book, looks at a page, figures out if there's something important, consumes all of it, and then goes to the next. There's some element of natural talent and disposition to this (I think there's probably a name for his condition). But the thing he says is that it took him 50-something years (basically, since the time he started reading, at age 8 or something like that) to read that book, because the process of having read so much throughout his life seems instrumental and actively improves the outcome of reading this new book.

I trust Tyler on his judgment here, and I think it's illustrative for many other tasks that we ourselves might do, which might be better done because we've accumulated throughout the years some form of practice that has made the output much better.

Another example here: I think some of my blog posts might actually come from years of practice. Many of my blog posts I just dictate to Claude and then publish with minimal edits, and a couple of friends have told me that it's surprising how engaging my writing voice is. And this is not a voice that I practiced. I'm just spitting things around to the wind. I'm actually spitting things into a boat yard apparently, in the middle of Barcelona. And as I do that, there's something compelling about this voice, because it's a voice that I've practiced so long and so hard. Not only have I tried speaking with conviction for maybe a decade already, I've also spent who knows how many hours just drafting ideas in my head, and enjoying the little tweaks, the little phrases, the sheer intensity and volume that different words and different repetition forms can have. And I've enjoyed all of that, and I think by this point maybe over a thousand hours, even.

So there's some practice here, and to the extent that you like my blog (I hope you really do), you might be able to appreciate this.

Keeping ourselves to writing for a bit, I think great books from the past are another great example of this. Mostly when I try reading old books, even the classics, the western canon, I find that they feel consistently shit. I make an effort to read the first page, it is shit. I make an effort to read the first chapter, it's still shit. It feels shit at every point that I'm reading. But I think what's going on is that great authors of the past did not try to write good sentences. Not at all. Not even good chapters. They were aiming for a great book. So this thing, the thing that they poured years and years of effort into, the final output that you should judge them on, is a book.

And it's a pretty high cost to pay if you're like me — someone who doesn't have that many uninterrupted continuous hours to just read, or who doesn't like reading the same thing for uninterrupted continuous hours. But I think when I hear authors say that they took so much from one of these old dense books, I now get them, because it's not the sentence, it's not the chapter, it's a whole thing. That's the output of a life's worth of effort and intensity.

I'll try to give more examples, because maybe it's illustrative for when you try to think of your own career. What are the things that you might do that you've poured several years of practice into?

I have great physical endurance. I had asthma problems as a kid, and as a result my parents put me into one of these swimming classes, and every weekday for maybe over four years I just went and I swam, and I swam hard because I was really proud and competitive. And so when I came out of swimming lessons I just had much better physical endurance than almost any other kid you'd see around. And I've tried to keep that. I try to run whenever — the other time, when I had the next bus from the airport to the city scheduled for like half an hour later, I just ran like 10 miles or something to get to my place, because I wasn't going to stand there and wait for the bus. And at every moment I can (I don't do this much deliberately because it's hard to fit into a schedule), but whenever I can, I just run. I just run hard and I run fast, because physical endurance feels like something great, and effort that I want to keep.

And maybe the place where I've used this more productively is at EAGs. Last EAG, the day I had most one-on-ones, I had like 16 or something, and maybe a few more shorter informal ones. Most people I meet cannot do 16 one-on-ones in a day. That's like eight hours, and mostly it was uninterrupted conversation. I was pacing around the whole venue, just going out to the street, just walking around. And that required physical endurance, and I was there for it, and I loved it. And that drew on some hard-earned capital.

What's another example? I think the intensity of my physical presence. So if you see me at EAG, again you might have noticed that I'm just bouncing around, that I try to look at you and just be there, just pour all my mind into the conversation that we're having. And I love the feeling of that. I started doing one-on-ones when I ran an EA group, that was maybe around 2021–2022. Not sure if I had had much deep conversations before, maybe a few with friends, but since I started doing one-on-ones I've tried putting effort into them. I've tried just being there and just letting myself be consumed by the other person, by the arguments, and just being as helpful as I can. It's like pumping my mind out, to give them something valuable that they can latch onto. And I think I'm good. I haven't checked with my friends at the last EAGX, but I think many of them would say this was maybe top three of the one-on-ones that they had — maybe top one, for several of them actually. And I can do that, and it took me effort, it took some kind of devotion to the task at hand, just to get that going.

There are other things that are not the result of long, deliberate effort, and they're more like some sort of innate disposition that you can just tap onto, especially when that disposition is somewhat rare across the population. So one of the things that I have that most people are not willing to put into service for their work is: I get incredibly anxious about people not replying to me in less than five minutes. So if I message you and you haven't replied, I'll just feel it, and it will be worse and weird and bad. So whenever someone sends me an email, if I can just reply in the next five minutes — my god, it feels great. And they might reply back to me — oh god damn it, that's amazing.

And so a lot of the things I've been doing recently which I feel have been pretty high leverage are: someone sends me an email, someone books me for a call — I love just letting people book me on my Calendly at any time of the day. (Please don't do it if you're boring, but I think you're probably not boring if you're really enjoying this thing that I'm writing.) And I'm just there for you. If you book me for the same day, the next day, I'm fucking there for you, because there's a certain intensity, a certain responsiveness, that I really love. And this I did not develop consciously. I'm not sure — I just started doing it more recently, and it just locked. It felt like my body demanded it.

And relatedly (I'm not sure if I have more examples, but relatedly), I think this is one of the reasons why I'm so bullish about people who are willing to just pour all of themselves into EA. Because there are so many things about you that are weird and extreme, and if you're willing to consistently try to figure out where in the world that skill and disposition fits — and if you're also consistently willing to just shape yourself, time and time again, in order to fit what's most helpful for the world…

For example, I didn't think I was a good manager when I was twelve. I probably absolutely sucked at it. But I've read things from good managers — I'm obsessed with Holden Karnofsky, and he might actually be a pretty great manager. And so I've been consistently just rearranging my cognition, my thoughts, and my hopes and spirits, just to try to be more Holden-like. And I think this is paying off. Recently I've had a couple coaching calls, recurrently scheduled throughout my week, and I do almost zero preparation for them, and I was surprised that at the last two calls I was actually tired by the time the call started, and the call seemed pretty good by the judgment of the other person there.

So yeah, I'm just incredibly bullish on you if you're willing to just keep trying, and keep trying for years. I've got friends who've been at this five years, a decade, and they still keep trying. Some of them have not succeeded wildly, but they're at it. And they're just like this extreme putty that's just so willing to forge itself. And it's not infinitely malleable — for many of the things I've said, these are extremely particular to me, and many of the people I admire I just couldn't reshape myself to be them. I don't have the soil for that. But if you're just willing to figure out *where can I push myself, where am I extreme,* and *how can I put that extreme in contact with reality to help* — then I'm so extremely excited. I'm so extremely excited about you.

---

# Act on all your great ideas

Source: https://myea.blog/all-great

People have been saying that, given how good AI is at executing well-scoped tasks, the new bottleneck is taste. That now your responsibility is to come up with and bet on the best ideas, such that AI can work on them.

I think that might be right to an extent, but the word "taste" here is misleading. Taste makes me think of an opinionated internal intuition about which things are promising or not, and that's just not something I have. I have some taste for art, for example. And I have zero taste for things I want to do well. Like, how the fuck do organizations work? I have zero taste for that. And I want to make it my project over the next three months to help them adopt AI better.

Over the next three months, I'll be trying to make EA orgs get radically more value out of using AI. And I still have close to zero clue about how organizations, any of them, work. So under the taste paradigm, it might seem that I'm either not well suited for this work, or should defer it until I've acquired the necessary taste, or somehow acquired some taste (which may be far from perfect but still better than zero).

Instead, the thing I feel drawn to do is just not rely on my taste at all. I have all these ideas for ways I could really help, and if AI (or the other capacities and affordances I've acquired throughout my career) makes it much more frictionless to get started and do a minimal version of them, I think I should basically try them all. So, for the most part, I'm not exercising my taste about what EA orgs need with respect to AI uplift. I'm just doing fucking everything I can figure out to do.

This isn't literally true. Most days I have more ideas than I have tasks started. But I've been trending in that direction, and very deliberately trying to move in the direction of acting on more and more of my ideas there.

And to clarify, I'm not acting only on the great ideas. I just think that, for me, and for you the reader as well, each and every one of our non-boring ideas is great. Not sure how I want to define "boring" here. But boring ideas just don't have a chance of leading to anything at all. They're just not aimed at doing the thing you care about. But everything that has even an ineffable sense that it could be good, that's great. That's great already. And you should do it.

To give some examples: last week I had a call with a security professional. And the thought I came away with from the call was that there were lots of biosecurity interventions that could only be meaningfully promoted by people working in US intelligence. I don't work in US intelligence. I mean, my nationality would probably make it really hard to ever work there (even ignoring the fact that I'm extremely averse at keeping most of my work secret).

But still, it seemed like an important consideration, the seed of possibly great ideas. So I thought: okay, who's in a position to act on this, to do something about it? 80K has a bunch of influence on what people do, so 80K should do something about it. That's a great idea. I have someone I know at 80K who could actually, if convinced this is a worthwhile thing to do, just make it so that 80K more prominently recommends people work in US intelligence. So I sent her an email. And that's how I acted on one of my great ideas.

Another great idea I had is that maybe CEA doesn't have the right culture or institutional make-up to take advantage of many of the most important opportunities to better steward the EA community. So I reached out to someone at Kairos and posited the idea that maybe a \[nimbler, more forward-looking\] org should eat up a lot of the current or projected CEA responsibilities. Kairos, in a sense, already did this for AI safety groups, and I think they're doing a great job at it. So I feel reasonably proud of this as the actionable for that great idea.

Those two are examples that don't involve AI, at least not directly. The affordance that allowed me to act on these ideas was mostly having met people and having shown them my seriousness and commitment to ideas and principles. Unlike AI, this contact-rich network is not something that's available to everyone.\[^1\] Most of these contacts weren't even notable to me two years ago. Still, I think one of the most valuable things you can cultivate is these affordances to be able to act on your ideas. To turn your exploration of the landscape of worthwhile projects into actual worthwhile projects by moving through them step by step.

I think AI is also a really important enabler here. Many ideas I've worked on, I wouldn't have dared to try if not for AI. It would have just taken too long to learn about the area or make an initial prototype. Damn, the thing I've been working on for the past year, I wouldn't have done it if it weren't because ChatGPT allowed me to vibe-code an initial prototype of the idea in 30 minutes.

So you should heavily use AI to act on your ideas. And for me, this just looks like making it easy to open a new terminal and just piece together whatever thing seems valuable and worthwhile. And many of these ideas are not as grandiose as the other ones. Many of them are: I'm fucking tired of having so many applications on my tab switcher, and the tab switcher app I downloaded is buggy, so I'll just make a new one. Because who likes wasting time on unpleasant interfaces? A bunch of my day-to-day ideas are actually pretty small. But I think they're great still. They either very clearly address a problem I know well, or have the chance of doing something for something more nebulous but way more valuable.

And it is through this process of exploration that the taste part of the taste/execution equation comes from, for me. For many of the things I do, I don't know them enough to have taste in any way people often use that word. But by arming myself with the right tools and affordances and friends, the bar for execution is so low that I can just get started, and then snowball into working on and focusing on the projects that work best.

\[^1\]: \[Placeholder for what I mean by "everyone".\] Most people, like the majority of people reading this, can probably afford a $200 Claude Code subscription. Yes, I realize that by global standards not everyone can. Still, I think you're probably privileged enough to be in this group, where you can definitely do it and have no good reason not to.

---

# What youth knows about love

Source: https://myea.blog/youth-loves

Shakira was 16 when she composed *Antología*, which after more than 20 years and many other hits is still possibly in the top 10 of her most listened-to songs. It's surprising how well that song captures the feeling of a longing, grateful heartbreak.

The other day I found a pair of shoes that I really liked. I've been picking shoes deliberately for more than 10 years now, and I was surprised to find that hiking shoes, which I had never tried before, might actually be a thing I end up using going forward. They're waterproof, comfortable, and durable, which is exactly what I've been looking for in shoes so far.

It was a bit astonishing to realize that it took me 10 years to figure out what line of shoes I like best. I buy shoes once a year, so it took me 10 attempts. But in a world where every attempt takes so long to play out, and it's a process to learn from, it takes literally decades to figure out the better options.

This is yet another reason to never shut up talking about love and romance. You only get so many chances to try yourself in the span of a young, early life, and if you're not deliberate in learning from the experiences of others, you might find yourself decades into life knowing little more than when you started.

---

# Play hard (with AI) so you can work hard (with AI)

Source: https://myea.blog/play-with-ai

Stub: play-hard

*Field notes from watching Luca and Pablo “work”*

*The two AI power users I've come to know best are really extreme in some of their habits. This is my attempt at a clean framework that rationalizes some of what they do.*

\[current tone of the piece is a back and forth between my marginalia on how I derived these conclusions and my current open questions \+ the tight-clear-flow-packs-all-punches authoritative blog post style in the middle\]

AI is a bad coworker. It's unreliable. It flatters its own work even when that work is wrong. Sometimes it just cheats.

If you're doing work where failures are costly, say, anything your reputation is invested on, a 10% failure rate is painful, and trying to do the thing with AI ends with you down on your knees fixing the mess it made. Even though it did 90% of the work, nailing that last 10% requires you to perform much of the intellectual labor you thought you'd outsourced.

But working well with AI is plausibly one of the most important skills you could build to sway the balance of the world towards the good. Even if in 5 years AI will automate everything, the stakes of what we need to do right before that are huge and frenetic, and humans will need to play a role in that.¹

So, how do you get good at working with something that's frustrating to work with?

You play with it.

The same characteristics that make AI a liability at work, also make it incredibly fun to play with:

* **Unexpected rewards with wild variance.** Sometimes you one-shot the tool you've been meaning to build for years. Sometimes you get stuck watching it try the same failing approach over and over. Sometimes it reintroduces the bug you painstakingly solved yesterday, because you didn't document the fix. The highs are high; the lows can be absurd.

* **A poorly mapped frontier.** There's little shared knowledge of what AI is actually good at. For most tasks you care about day-to-day, it's still surprisingly unexplored how well AI could do them. Figuring it out requires active problem-solving. That's annoying when you need results, but it's kind of exciting when you're exploring.

AI also gives you immediate feedback loops — you prompt, you get something back in minutes. That's good for work, but it's *great* for play. You can try your ideas right away.

When you're playing, 90% is enough. You don't need to ship. You don't need to wrap up the last 10%. If something doesn't work, you don't have to deliver it — you just learn that AI isn't good at that yet. No frustration, just some learnings.

## Title new section

Some of the mechanism might be that through play you can test specific tools and patterns that you can reuse for work: “can I use swarms of LLM agents to fact-check every individual statement in a report?”, “what's the maximum complexity of a website I can build in a single prompt", however, I feel that alone doesn't support the strength of why I think that play and exploration and so valuable if you're trying to fully take advantage of AI. I've seen how Pablo and Luca are so radical in just trying things and building for themselves that I think there must be something more to it.

Maybe I’ll have a blog post about that later. For now, here are some questions I'm interested in:

* What's the best way of thinking about what AI makes possible? Some step in our work is now practically free (e.g., writing the code), which suggests we should be doing much more of it and of any of its complements. But what's the thing that is now so cheap and what are its compliments? What explains the order of magnitude differences in the productive output of differing people dipping in this space?

* What does great play look like? For example, I feel both Pablo and Luca have this exacting standards for what the final output must look like (e.g., can I trigger this with a simple key combination? Does it show the glory of emacs?) in a way that feels quite different from just aimless-general-interest exploration

* How much of the benefit is just getting you excited about using AI? If you have gleeful enthusiasm for trying out AI for a project, you'll do it more and better (including at work), maybe that's a big chunk of this.

* What kind of ‘capital’ are you building through play? What's the thing you can accumulate and reuse?
  * Maybe a library of working examples — proof of what's possible — that you can recombine later. Simon Willison [writes about this](https://simonwillison.net/guides/agentic-engineering-patterns/hoard-things-you-know-how-to-do/): the idea that you only need to figure out a trick once, and then your agents can reference it forever.

* Following the exploration angle, how can you explore faster/wider/better?
  * Watching AI power users (literally sitting behind their screen and talking to them) was immensely helpful for me, and it unlocked several things I had not tried. How did that happen?

---

**Notes**

¹ I'm assuming here that learning to work with AI is one of the most valuable skills you could build right now. See \[link\] for why I think that.

# Tab 2

**Play Hard (with AI) So You Can Work Hard (with AI)** *Field Notes from Living with Two AI Power Users*

AI is a bad coworker. It's unreliable. It has poor judgment — or I don't know what to call the thing where it flatters its own work even when that work is wrong. Sometimes it just cheats.

If you're doing work where failures are costly — anything you need to ship, deliver, or stake your reputation on — you can't afford the failure rate. Out of the box, AI is still missing a couple nines of reliability before it can replace humans at most tasks. Even software engineering, which is famously the field where AI is most advanced, still has this problem: deploying production software is high-stakes, and fully embracing AI there still feels like a risky proposition.

You often end up being the person who goes down in the trenches and fixes the mess AI made. And even though it did 90% of the work, nailing that last 10% is damn troublesome. It often requires you to perform much of the intellectual labor you thought you'd outsourced.

But before AI can do it all, there's a window where knowing how to direct it is unusually valuable.¹ So how do you get good at working with something that's frustrating to work with?

You play with it.

The same characteristics that make AI a liability at work make it incredibly fun to play with:

* **Unexpected rewards with wild variance.** Sometimes you one-shot the tool you've been wanting to build for years. Sometimes you get stuck watching it try the same failing approach over and over. Sometimes it reintroduces the bug you painstakingly solved yesterday, because you didn't document the fix. The highs are high; the lows can be absurd.

* **A poorly mapped frontier.** There's little shared knowledge of what AI is actually good at. For many of the tasks you might care about day-to-day, it's still surprisingly unexplored how well AI could do them. Figuring it out requires active problem-solving. That's annoying when you need results, but it's kind of exciting when you're exploring.

AI also gives you immediate feedback loops — you prompt, you get something back in minutes. That's good for work, but it's *great* for play. You can try ideas in almost no time.

The stakes of playing feel lower. When you're playing, 90% is enough. You don't need to ship. You don't need to wrap up the last 10%. If something doesn't work, you don't have to deliver it — you just learn that AI isn't good at that yet. No frustration, just some learnings.

And through play, you're building something you can reuse for work. I'm still not sure exactly what's going on, or why play seems so important. But I've seen so many AI power users do insane amounts of play and genuinely enjoy it, so I guess something must be there.

I'm working on a post about this. I still don't quite know what the answer looks like, but here are some questions I'm interested in:

* What are the remaining bottlenecks? AI made writing code nearly free. So why isn't everyone building custom software overnight? What's still hard?

* What do the people who build daily do right? Pablo and Luca ship side projects constantly. What's their pattern?

* Is most of the benefit just enthusiasm? If play makes you excited to use AI, and that excitement carries into work, maybe that's the main thing.

* What are you accumulating when you do side projects? Maybe a library of working examples — proof of what's possible — that you can recombine later. Simon Willison [writes about this](https://simonwillison.net/guides/agentic-engineering-patterns/hoard-things-you-know-how-to-do/): the idea that you only need to figure out a trick once, and then your agents can reference it forever.

* Is exploration just the point? AI's capability surface is vast and poorly mapped. Maybe the right move is to surf it, maximizing exploration speed. Play is better for that than work.

---

**Notes**

¹ I'm assuming here that learning to work with AI is one of the most valuable skills you could build right now. See \[link\] for why I think that.

---

*The way this piece came about: I watched Pablo and Luca, who are really good at using AI. I noticed the ways they're weird and unusual, and tried to reason backwards — what would need to be true about the world for those patterns to actually be advantageous? As a methodology, it's pretty faulty. But I think it gets you to interesting conclusions, or at least provocative hypotheses that might actually be true.*

---

# Earning to Claude

Source: https://myea.blog/etc

*Earn more so you make the most of AI in case it gets pricier*

The latest Claude model, Claude Mythos as of today, is both much better than its predecessors and five times more expensive per token.¹ So far we've been lucky that the best AI models have been available at a price most working people can afford,² but this may not continue for long.

There's suggestive evidence that the cost for an AI to do an hour's worth of human-equivalent work [is rising](https://www.tobyord.com/writing/hourly-costs-for-ai-agents) as models improve. This might be partly explained by the fact that, until recently at least, much of the improvements in AI capabilities came from letting models "think" much longer, and the compute cost of that thinking has to be paid for every time the model runs.

[https://www.tobyord.com/writing/hourly-costs-for-ai-agents](https://www.tobyord.com/writing/hourly-costs-for-ai-agents)

[https://www.tobyord.com/writing/mostly-inference-scaling](https://www.tobyord.com/writing/mostly-inference-scaling)

Given these trends, it's plausible that the best way of using AI will soon involve spending thousands or tens of thousands of dollars per month.

If that's the case, and you take seriously the prospect that the most valuable work we can do will leverage AI plenty,³ you should be trying to earn enough such that you can afford to use the best models and use them for as long as needed.

For people who, like me, have so far de-prioritized income to preserve the flexibility to work on whatever appears most important, this is worth taking seriously.

So before you take your next opportunity, consider whether you should be asking for more money (or doing something else entirely) so you can afford to make the best out of AI.

¹ Tokens are just a fancy term for words, or similarly-sized chunks of text.

² That is, assuming you live in a higher- or middle-income country. And even for middle-income country dwellers, I think the current Claude Max $200 subscription makes lots of sense if you're the kind of person who would be proactive enough to take advantage of it. Looking at you, Diego\!

³ I make the case for that \[here\].

---

# EA groups are hard and great and worth it

Source: https://myea.blog/worth-it

*To Alejandro, Yuli, and Efrén*

Ben Kuhn says you don't have to do hard things \[[link](http://placeholder)\]. But there are obviously some hard things worth doing. I want to point at a cluster of things that are.

Most things you do in life are capped. The opposite of open-ended. Most things you do in life, you do not fully understand what you do them for. You submit an essay for a school assignment and you have this big story told by your parents and your teachers that if you do better on the assignment then you'll build some skills or you get a diploma and then you can use that to get a job and then you can use that to get money or to get a better life or to change the kind of person you are or to grow. And at least for most of my life, the things I put my greatest effort on took that shape — they seemed to be intermediate outcomes and the goal was blurry and was not given by me.

I think founding and running an EA group was the first time I was exposed to a hard, open-ended, valuable thing that I could own myself.

What you're aiming for in an EA group is — you know that there's ambitious, good-hearted people out there, and that some of them, when presented with the idea that they can just save ten more lives, a hundred more lives, if they are strategic about their efforts, that part of them would start aspire towards that. And maybe at some point wholeheartedly work and devote their careers to that. You know there's people out there like that. And some of them you've reached out to and are starting to become your friends and you love them and you want them to be ambitious and to try hard and to do good and save lives. And you care about that. You care about that hard.

But now instead of having a rubric for your essay you just have the world.

How do you support someone's ambition? How do you help them navigate the following 40, 60 years of their career? How do you take those decisions yourself? What are the values? How do you make a decision now about what to do with your time, what projects to start, what classes to enroll in, what things to apply to? Thinking now not about the metric they give you but thinking about what you care about, thinking about the ultimate good, whatever that turns out to be. Now that's open-ended. And it's hard.

And no one's there to tell you what to do. Open Philanthropy — well, they may ask you some questions at the end, how do you feel about it, did you like it — but they don't tell you what to do. If you want to do a shrimp pizza party as the opening event for your group (the party is for celebrating shrimp, the pizza is vegan), you can do so. Just do whatever.

And almost no one's there. Luckily now you may have some mentors and I'm grateful that OSP exists. I'm incredibly grateful to the people who served as our improvised mentors in our time. But for the most part they're not there and for the most part there's just hard decision and hard action and no one came to your event. You advertised it in every email you had and still just two people showed up. And oh god dammit it hurts. And oh god dammit does it mean that you're not bringing heaven to earth in the way that you hoped you would when you decided to start it?

So I think if you've already started an EA group I don't need to tell you that it's hard. You know this already.

But maybe what I can say to you, given that I've been through the whole thing and I see my friends grow and become people I cherish and admire, is that it's worth it.

Because when you care so hard about something, sometimes — and luckily with all the support you get, many times — it results in that thing becoming closer to reality. All the hard I cared, all the intensity I carried, through the power of human will and work and the malleability of the world, eventually managed to turn into something. Not the full extent of my aspirations. We didn't conquer the world and make everyone donate to \[malaria? shrimp? you get the idea\]. But I got some friends and they're excited about just trying to do the goodest. And I really love them. And I'm proud.

And now you're starting a new group. And I really love you. And I'm proud. And I wish you the best of luck.

This may be hard. And most things hard are not worth doing. But maybe for one of the first times in your life you're going to get to spend months, possibly years, working towards something hard and worth it.

Because most of the goals the world tries to give you, they're blurry, and they don't feel yours entirely. They may be valuable. I don't know. I'm so grateful I got taught basic math literacy and many of the things I learned in school, just the basic habits, were great.

But now it's a different thing. And it's yours to seize it.

---

# Please forgive the errors in my ways

Source: https://myea.blog/please-forgive

I spend too much of my time talking and thinking about extreme versions of honesty. As a result, I often feel compelled to direct my thoughts to friends in ways that may reveal not only my lack of decorum but also my ignorance of their goals, their situation, their emotions, and even of the world itself and its workings.

Why do I do this?

Maybe I have a delusional account of the world, where every extra piece of information can only help you make better decisions. What better way to make all the best decisions than to have all the best information, which you acquire through the perfect honesty of everybody else?

Maybe I fantasize about amassing twenty-six billion dollars and becoming the wealthiest person under thirty. And when the report comes to me—*"Sir, crypto prices have come down and our investment arm can't cover its losses. The only way to keep things afloat is to use some of the customer deposits."*—I say, *"Oh, but wouldn't that involve lying to our users? To EAs?"* And I go on and not do that, and still manage to keep most of my wealth, and everyone in EA remains ambitious forever.

Maybe I don't know you at all. Maybe, through blurting an unrequested piece of advice, I've just shown you how little attention I've been paying and how little thought I give to all the vastness inside you. Maybe I've shown you that I have mixed motives beyond your wellness—that I am willing to cause you discomfort just for the thrill of following some abstract, unvalidated ideal.

Maybe I am so exhausted by the complexity of consequentialism in this history-defining moment and every decision feels like a thousand twists and turns I cannot anticipate. Maybe I want to grip tightly the hope that if I follow some simple maxims—be more honest, be more ambitious—I can capture the upside that my poor capacity for consequentialist reasoning would otherwise squander.

Maybe I can't wait to tell you about my day. Maybe if I pour my thoughts through you I won't have to make all the hard decisions alone. Maybe if I say what's in my mind loud enough, someone will see me and think, *oh, do I love this kid.* Maybe if you know me well enough we'll marry, and I'll know you as much as anyone can be known.

Maybe I've seen the world outside of EA and thought: isn't it great that we're so much more honest here? Maybe I've reached out to someone who months ago was a perfect stranger and said "Can you advise me on building a free competitor to your tool? It's for biosecurity" and have them become one of our most cherished advisors. Maybe I've emailed a famous EA and been enchanted by their candidness. Maybe I appreciated some curt words when they were meant to seek truth.

Maybe I've seen how much of the costs of honesty fall on the sender. Maybe I've seen how much it mortifies me to be rejected, to say a true thing and have the other person realize I'm not what they're looking for (not the right partner, not the right candidate). Maybe I still love them hard enough to feel proud of letting them make their choice.

Maybe I dislike small talk, and would rather ask you why you love your partner than hear which bus you took on the way here.

Maybe every argument I have heard against full honesty sounds like some version of "I just don't like how it feels." We have these pesky emotions, these weird preferences against hearing some things that might be true, and we should aspire to be stronger and see through them.

Though my motives may be impure and my reasoning may be faulty, I want to be honest, and I want it badly. I don't know if I can do this without being some tinge of hurtful and annoying.

So I ask, please forgive the errors in my ways, while I keep doing what I think is most important. I'm just trying to do good, and I'm confused about how, and I mean the best for you, and I mean the best for everyone.

---

# And I bought yet another flight

Source: https://myea.blog/hunting

*A pro-tanto reason to not hire me as your ops person*

I wake up groggy. For the past few weeks it's been rare the night I manage to sleep more than six hours. Last night was unusually intense — I got to talk to possibly one of the best talent pickers in EA, and they challenged me to become more than I am.

(Also, while mulling over this mentor's words, a bearded young man at the park graciously offered to suck my dick. That night I also confessed to a girl I love and we talked about it for over an hour. But that's not the main story.)

The story is that I wake up groggy because I went to bed real late, haven't slept well, haven't slept long, and I'm a complete moron at this hour. I can already tell work is going to be hard.

The first thing I do is lie in bed and have thoughts. I have *lots* of thoughts. Lots of things have been going on. I keep contemplating them in the shower for at least an hour. By the time I come out, I've had probably four blog post ideas, of which I draft one to send to Robin.

Now totally exhausted — from the lack of sleep and the intense mental effort of thinking great thoughts — I try to figure out what I should do. I know I have to travel soon, so I might as well hasten the departure.

Last night Robin and I had talked about meeting up; my recollection is we'd agreed on meeting in Barcelona in the next few days. The best way to get there seems to be a train from Toledo. I buy tickets, then settle in to do some work before departing.

As I start working, I half-find half-remember I have three calls booked for that afternoon, all coinciding with my train ride. I blame it calmly on my absolute lack of sleep. I move the tickets to later and tell Robin my new arrival time — at which she asks, confused: why am I going to Barcelona, when I'd said I'd go to Toulouse first?

I retrace my steps. I look up how to get from Barcelona to Toulouse. Turns out it's a long train ride, so I'd rather go from Madrid by plane. I buy tickets for the last flight I can find for today, leaving me time to finish my calls and head to the airport.

## The next day

This is the third flight I'm about to try embarking on today. Hopefully I'll make it this time.

The first flight wasn't my fault. Or, it was only partly my fault — an act that might be forgiven of most people. I bought a flight from a random airline, showed up on time, and it just happened to be cancelled. Me being a busy person, I wasn't about to put up with that shit — wait around for them to assign us a hotel and book us, who knows when, on another flight. So I got my own place at a reasonable price next to the airport, and a flight for the next day.

It wasn't the cheapest. I was actually a little desperate to get to Toulouse and meet Robin, so I just got the earliest I could find. I didn't start booking until I got to the hotel, by which time I think I'd lost the chance at the absolute earliest flight, the 7:15 AM. The next earliest was departing at 3:40 PM, which felt really late given my preferences. I just wanted to be there as soon as possible. But it would have to do for now.

I was a bit tired, so I took the early morning off, then walked to the airport early, found myself a chair, and settled with the intention of working until the flight was called. I lunched. The gate took real long to be announced, but I knew it would show at 2:45. At about 3:05 I was dashing to my gate; boarding was scheduled for ten past. I arrived a few minutes after — I'm really fast after all — and there was no line. I breathed out.

As accustomed, I sat in the nearest row with a view of the gate. I waited a bit; no line was forming. The line was for a different flight at the next gate. I figured I could use the time to send my application to the Generator fellowship.

I'd learned before this flight's airline was one of the most punctual in Europe, after the previous flight got badly delayed and cancelled. It didn't cross my mind to use that knowledge at this moment.

Fifteen minutes before the scheduled departure, there was still no line. So I went up to the gates, only to find that there was no one there. Apparently — I'm not sure how to reconstruct this, but the narrative that leaves me best standing is one where my flight had continuous boarding, and so there was no line at any point. There was just people entering the flight whenever they wanted. How stupid and silly I feel now.

OK. Now that my flight was no longer, I had to focus. I've been working on several other blog posts about my ops failures, and one of the big lessons is: as soon as the contingency hits, focus, think clearly. What are my options? What's time-sensitive?

(Earlier that day I'd received an email that my flight would be rescheduled to depart at 4:30 PM. I had my own flight at that time, from a much more trustworthy airline, so I had deleted it and paid little mind.)

So I knew there was another flight I could take. My original cancelled flight was scheduled to depart at 4:30 PM. It was a few minutes before 3:40 — I could make it to another terminal if I ran like hell. Luckily, I'm really good at running.

So I just ran. Ran all the way through to the terminal's exit. I tried getting a taxi, but Spanish people are sticklers to rules, so they wouldn't let me take one unless I joined a real long line. Jogging was not an option — I'd looked up the distance, and it was just much faster to go by wheel. I'd thought, while running to the terminal exit, of getting an Uber, but Uber arrival times are unreliable, and waiting for one — I didn't feel like it. So I took the bus, the free shuttle that runs between terminals. I caught one right before it departed, and prepared, while on the bus, to run madly for my flight. I got off two stops early — I knew I could make it from there faster than the bus by running to the other side of the terminal.

I dashed, stopping only to ask how to make it to Volotea's registration desk. Alright — there's no one there. Out of business hours, I figure — I'd called them on the bus, no answer, and the machine had clarified they only worked Saturdays until early afternoon.

So I went for a security guard and tried explaining my situation. He pulled out his phone, looked up the flight, and when he couldn't find it, just recommended I try using my ticket past the gates. I run through the gates, greet the girls at the entrance and, through some sweating and panting, I explain my situation. They take me in empathetically, check my information against a paper they have for the cancelled tickets.

I make it past, and I know my gate. I run all the way through. And everyone was so kind to me — letting me skip the line, letting me pass through with little interruption. I pant and I run. I make what the airport sign says takes around six minutes in about 55 seconds. I get to the gate. There's a girl at registration, but she's not from my flight. And apparently my plane had departed some time ago, even though the sign still showed it on last call.

At this point I feel anger, but it's the kind that feels like excitement. Like I'm hunting something. Like the world is my prey and I'm there for the hunting.

I walk out, and I'm determined: I'll fucking make it to Toulouse today, be it the last thing I do. Again I reach the terminal exit; I sit and consider. I know there's another flight from Iberia. I look for it — it's damn expensive, three times the price of the last flight. I search for other options. Train (not available). Bus, 10 hours, would disrupt my sleep. Flights with layovers, roughly the same price and much longer — maybe saving 100 euros, but not, not worth it now.

I reflect a bit on the inconsistency of my preferences. I've spent an hour before to save 50 euros on a laptop, and now here I am, willing to give it all just to get a fucking flight to take me to the place I want to be today.

So I pay up. I take the bus back to the terminal — same terminal as my previous flight. I'm proud I made that mad sprint for my originally rescheduled flight — it was a decent bet.

And now here I am, waiting. I'm actually pretty energetic from all the running and chasing, so I think I'll get to do some work before the flight.

The next flight is scheduled for boarding soon.

---

# Metaphors for caring

Source: https://myea.blog/smart-dolphin

[Alejo Acelas](mailto:alejoacelas@gmail.com)

I've been trying to find good metaphors for the [act of caring](https://docs.google.com/document/d/1SUPH6K6J67J5pgT_d0H82219RA5brg6lfaoBLtkXSi4/edit?tab=t.0) for another person.

One of the first that came to mind was helping a wounded bird, and the emotion this brings to mind is that of very delicately tending to a small, frail, shaking bird. One thing I like about this metaphor is that my ex used to refer to me as bluebird (*azulejo* in Spanish, which rhymes with my name, and she coined it for me because of my tendency to shake as if shivering when possessed by excitement). But what I don't like about it is that, while it captures the idea that you should be careful and deliberate in your efforts to help others, the reason for that is that the other is fragile — which I think doesn't represent the case well.

A slightly better metaphor might be helping a falcon, taking care of it after a wound. You can think of falcons of different ages: a small falcon, you might actually have more to tell, more to guide it on, as it might only be starting to learn to fly, and even with your total incomprehension of the experience of falcon flight, you might still be able to guide it in the right direction. An older falcon, you just need to care for the wounds and help it get better, help it soar back into the skies.

I like the falcon metaphor because it immediately transmits just how terrifyingly capable the person you might be caring for can be. It also shows their alien-ness. I look at a falcon's eyes and I know there's a world beyond me. I can't fathom a falcon's mind, but I know there's a big otherness, and most of the falcon's highly precise, superhuman skill I cannot pretend to imitate or understand.

But I don't like this metaphor because human hands are too dexterous. As you're caring for the wounds of the falcon, you are so much more capable of doing it than the falcon could itself (not to mention that you're the only one in the relationship who has immediate access to millennia-long practices for healing wounds.)

So a better metaphor might involve something more of a peer relationship. Maybe two animals, with one of them trying to help the other out. Not sure why, but one of the first that came to mind was the snake. A snake trying to care for the wounds of a falcon. I like this because snakes have no thumbs — hell, they don't even have limbs — so the only thing they can do in their attempt to care is to contort their spine into whatever shape they think would be most helpful. And this is clumsy and blunt, but still: snake bodies, despite their obvious limitations, are malleable enough to make out of them something that might be of help. Snakes may also gather food for the falcon as it recovers, or might steal from human civilization a curing potion for the falcon's ails.

Still, this relationship between falcon and snake feels only temporary. I can see a wounded falcon receive care from a snake while it cannot catch a fly, but once it's healed, once its temporary disability is lifted, there is no reason why a falcon and a snake should see each other much more. No practical reason, at least. A sentimental one might remain.

So I'm not sure how to represent this. I want to care long for my friends, and I want to elucidate the shape that this caring should take. I want to help them grow and shape them while recognizing the sheer otherness of their existence and the terrifying potential that they hold within.

\* PS. another metaphor I thought of for caring: offering career advice to an extremely smart dolphin. What I like about this metaphor is that your first reaction is: what the fuck? What do I know about dolphin society? I should start by asking lots of questions to understand what on earth is going on under the sea.

---

# Sad kindness

Source: https://myea.blog/sad-kindness

I've noticed this pattern in some of my close friends, and in the people I feel most magnetically drawn to. I can't quite describe what this trait is. I'd name it something like a sad kindness that radiates out of them. No, a kind sadness, actually. When you talk to them, there's something so receptive in their disposition, like they just take you in and return something warm and tender. But you can also see that either now or in the past there was something quiet and sad and timid and slightly tragic that they carry around in their demeanors, some trace of a flailing will.

And as I thought of them, a certain personal story came to mind.

When I was eleven years old I had just moved to my grandparents' big house in Bucaramanga. I had been living alone with my father for three years already, the last year of which I had skipped school. My father, I can't remember exactly why, had decided not to enroll me. Now I was at my grandparents' house, still a few months away from enrolling back into school.

I remember one night, after my grandparents went to sleep (which they did at an unusually early hour), I had just finished my bedtime routine, maybe drinking some water and putting on my pajamas, and I went to the empty living room, picked a seat, and carefully sat down to cry. I was really careful, holding each out-breath such that the sob was barely audible in the dark of the night. I must have sat there for ten minutes or so, the tears running down my face while I held the edges of the cry silent, when my grandfather arrived.

I don't quite remember what I was sad about. Maybe it was that it had been several months since I'd last interacted with someone my age. This was a year when my father, though he didn't know the term at the time, thinks he was probably clinically depressed. It had been not long after a violent divorce with my mom, and after he had been forced to move back with my grandparents when he wasn't able to earn his wage at the company he founded with his brother.

Still, my grandfather asked for no explanations. He just sat there, and in seeing me cry started telling me a story. When he was younger, after having married my grandmother and having had the eldest of my uncles, he had been unable to secure a job in his hometown, and a friend had offered him a job in the departmental capital, Bucaramanga. He had scraped together the fare for the bus and arrived alone, making the promise back home that he'd take some weeks to settle on a job, find a place to live, and take them with him. When he arrived, I'm fuzzy on the details now, but his friend was either unresponsive or impossible to locate, and the upshot was that there would be no job for him. So that night, and possibly many others, he just picked a chair and sat there in the middle of the night to sob quietly, carrying on him the weight of his family and the responsibility he had to them, and his inability to resolve it.

In that moment, after hearing my grandfather's story, after having him tell me that our cries were of the same kind, after he told me that this is the way real men cry, I felt like the bravest person ever. I felt that my crying was an act of defiance against everything bad that had happened, and I felt strong for having kept my kindness and my will throughout.

And so maybe the reason why I feel so attracted to the kind sadness of my friends is because I have a message to them that cannot quite be put into words: that their courage will matter \[finish with some Katherine Rundell\].

\[photo of the house / of us in the house\]

Quotes from Katherine Rundell on one of my [favorite interviews ever](https://www.youtube.com/watch?v=GQ7bkW6dvBQ)

**On what love does / what it's worth**

"...the things that I want to tell them are things that I cannot prove to be true but that I believe utterly to be true, and they are things like **love will matter, hope will matter, your endurance and your tenacity will matter**... something along the lines that **your generosity, your self-sacrificing boldness as you move through the world will matter**."

"I want to write books for children that will have in them, you know, plots... climbing trees and going down rivers and climbing on rooftops and riding wolves, but underneath it is the hope that they might see **love will give you courage, love will make you brave, love will give you back so much if you offer it with a kind of boldness and exuberance**."

"...one of the things I want to offer children is the idea that **your love, your tough and funny and inventive and careful and imaginative love, will be your guiding force for your life. Therefore love as boldly as you can because it will shape your heart and push you forward**."

---

**On the children she writes, and the kind of love she means**

"I think the thing I often wanted with the children in my books is the thing they're learning to put their faith in is love, and they are **children who love recklessly and freely, without reckoning, without an accounting of the balance sheet of love**, that they are all children who act for love..."

---

**On what she wants to tell children**

"...if you could tell children something and you could guarantee through some magic fairy that they would believe it — not they would half-believe it or a little believe it — but you could tell them something and they would take it to heart for the rest of their lives, **like putting a sort of fish hook through the imagination** — what would you tell them? And I think **the things that I want to tell them are things that I cannot prove to be true but that I believe utterly to be true**."

---

**On unexpected kindness in the world**

"...if you stand orphaned, the world is such that though there will be danger and though there will be those who will not seek to help, **there will also be people... who will step in and gather you up**."

"\[*on* The Railway Children\] ...that book understands so much about human pain... The world will largely break faith with you, the world will largely not offer its most bright and shining face because it is broken, but **sometimes the smoke will clear and the thing you long for most will call your name**... Sometimes the world does turn good, and sometimes wealth is dropped into your lap, and **you are embraced by pure love**."

---

**On attention (close to love in her framing)**

"I think **attention might be so closely bound up with love as to be almost indistinguishable**, and I think it might be the thing we owe the world most — attention to the natural world, attention to the intricate reality of one another's hearts, attention of course to the political reality of the world that we stand upon, unswervingly, unflinchingly, to the injustices that we live among. I think it is one of the hardest things to sustain and I think it is one of the finest gifts that we can offer. And that's the thing that I want to tell children: you know, **the world is immense and it awaits you and it is worth your passionate and focused attention. Go forth and offer it.**"

# claude notes

You're right that it's short, and you're right that dictation carries emotional information that text doesn't. A voice that catches before "picked a seat" tells the reader something the sentence alone doesn't. On the page you lose that, and the question is whether the text compensates or just sits there being shorter than the feeling behind it.

My honest read: the piece is close to working but is doing less than it could. Here's what I mean.

The opening paragraph and the crying scene are the strongest parts. The opening does real work — the correction from "sad kindness" to "kind sadness" is a great move, it shows you thinking on the page, and "the trace of a flailing will" is a phrase that lands. The crying scene is specific and strange in the right way: an eleven-year-old carefully composing his own sob is an image that doesn't need help.

The middle sags a bit. The grandfather's story is told efficiently but almost too efficiently — it's summary where the living room scene was scene. We get the facts of his situation (friend unresponsive, no job, sat in a chair at night) but not the texture. One concrete image from his story would do a lot. What was the chair like. What did he see out the window. Something. Right now his night and your night are supposed to rhyme, and they do structurally, but yours is rendered and his is reported. That asymmetry costs you.

The "real men cry" beat is doing a lot of work in one sentence and I think it needs a touch more room. Right now it's: he told you this, and then immediately you felt like the bravest person ever. The transformation is fast. In dictation your voice probably slowed down there. On the page, one more sentence between the telling and the feeling — what it was like to hear it, the quality of the silence after, the first thing you noticed about your own body — would let the reader arrive at the bravery with you instead of being told you got there.

The ending, in whatever version you pick, is reaching for something the piece hasn't fully built the ground for. You're saying: the kindness in my friends is what survived their holding. That's a beautiful idea. But the piece hasn't quite established that *you* kept your kindness through the holding — it says you did, but we don't see it. The crying scene shows you keeping your composure, not your kindness. If the arc is "kindness survives hardship and I recognize it in others because I recognize it in myself," the piece needs one more beat where we see the kindness, not just the will. Otherwise the ending is asserting a connection the body of the piece implies but doesn't quite earn.

On length: I don't think it needs to be much longer. Maybe 150 words. The additions I'd push for are concrete, not expansive — a detail in your grandfather's story, a breath between his words and your response, and somewhere (maybe in the opening, maybe woven in) one line that shows kindness as distinct from will. The piece is about kindness surviving but the word "kindness" is doing almost all the work of that idea by itself.

One more thing. The dictation voice is actually still in the text — the "no, a kind sadness, actually" correction, the "I'm fuzzy on the details now" aside, the way sentences accumulate clauses. That's good. It gives the prose a spoken quality that suits the intimacy of the story. I wouldn't smooth those out in a final pass. The piece should sound like someone telling you something, not like an essay.

So: trust the voice that's already on the page, add two or three concrete beats where the prose is currently summarizing, and give the ending one more piece of earned ground to stand on. That's the version I'd want to read.