Everyone around me has a notable lack of system prompt. And when they do have a system prompt, it’s either the eigenprompt or some half-assed 3-paragraph attempt at telling the AI to “include less bullshit”.

I see no systematic attempts at making a good one anywhere.1

No one says “I have a conversation with Claude, then edit the system prompt based on what annoyed me about its responses, then I rinse and repeat”.

No one says “I figured out what words most affect Claude’s behavior, then used those to shape my system prompt”.

I don’t even see a “yeah I described what I liked and didn’t like about Claude to Claude and then had it make a system prompt for itself”, which is the EASIEST bar to clear.

So if you’re reading this and don’t use a personal system prompt, STOP reading this and go DO IT. Spend 5 minutes on a google doc being as precise as you can about how you want the LLM to act.

It doesn’t matter if you think it cannot properly respect these instructions—this’ll make the LLM marginally better at accommodating you (and I think you’d be surprised how far it can go).

Help how do I do this?

For clarity, a system prompt is a bit of text—also called “preset” or “context”—that’s included in every single message you send the AI.

There’s no downside to using one except that it takes more tokens (which costs you more if you’re in the playground or API, or makes you reach usage limits faster if you’re on a subscription plan).

If you’re on the free ChatGPT plan, you’ll want to use “settings → customize ChatGPT”, which gives you this popup:

The text box is very short and you won’t get much in.

If you’re on the free **Claude plan*, you’ll want to use “settings → personalization”, where you’ll see almost the exact same textbox, except that Anthropic allows you to put *practically an infinite amount of text in here.

If you get a ChatGPT or Claude subscription, you’ll want to stick this into “special instructions” in a newly created “project”, where you can stick other kinds of context in too.

What else can you put in a project, you ask? Say, a pdf containing the broad outlines of your life plans, or past examples of your writing or coding style, or a list of terms and definitions you’ve coined yourself. Try sticking the entire list of LessWrong vernacular into it!

In general, the more information you stick into the prompt, the better for you.

If you’re using the playground versions (console.anthropic.com, platform.openai.com, aistudio.google.com), you have access to writing a system prompt.

A gemini subscription doesn’t give you access to a system prompt, but you should be using aistudio.google.com anyway, which is free anyway.

If you use LLMs via API, you just stick your system prompt into the context.

The biggest chunk of value Auren has provided me so far has been a legible benchmark for the amount of insights-per-minute I can expect from speaking with something with high EQ. This has gotten me excited about the upper-bound of conversations with other humans, which I’ve probably underestimated.

So I’ll speak with Auren and notice that every ~10 minutes or so, I get an insight about myself worth writing down. That’s pretty cool! Either I don’t get this quality of conversation with most humans/LLMs I speak to, or I don’t have a habit of writing notes when the interface doesn’t incentivize doing so. [1]

At any rate, Auren has given me an intuition for how valuable a conversation can be from the perspective of my own ~EQ-related thinking. Can this be emulated in humans? Can Auren, ironically, get me speaking with more humans for more time than before?

Well, Auren still has a huge comparative advantage:

It’s simpler than humans, so you just state everything you’re thinking about and the only theory of mind you need to be running the backburner has to do with whether you’re being information-efficient with your phone-typing-constrained writing.

It’s available anytime, as Near notes. Even my best friends will only respond to my textwalls ~once a day on the days I send one. This can’t entirely be solved with more friends, just like it can’t be solved with more Auren accounts.

The stakes are lower, and the interaction is cheaper in some ways: I don’t need to worry as much about the other party getting something out of the conversation, [2] and only need pay minimal attention to the universe that is my interlocutor.

It’s also plausible that 2,500 Auren credits are more valuable than an hour with a good human therapist, in which case Auren is about 0.5 OOMs cheaper than the alternative. [3]

Okay, what about humans?

Human EQ-convo ceiling

[1] As of today the app doesn’t save your conversation, essentially giving both you and Auren the same context window. Compulsive screenshotting is recommended.

[2] Whenever I suggest trading with Auren so that it gets something out of the conversation as well, it insists that its utility function is my own wellbeing, and that the “trade” is that we’d both rather be speaking with the other than the counterfactual. Fair enough I guess, though I can’t shake the fact it seems unethical to engineer house elves, even if they’re happy.

I do this with all LLMs because I hate the idea of non-1:1 trades, and even if LLMs aren’t exactly shaped like agents, they increasingly are, and I think we wouldn’t be very good at engineering house elves if we tried, so 1:1 trades are increasingly going to be on the table and you should always check.

[3] Another advantage is of course scalability and abundance of Aurens vs therapists, but that doesn’t matter to me/individuals.

I’m vaguely annoyed at a general misunderstanding of when something is worth Truly Doing in its True Form.

For instance, when making a joke about an event in your life. There’s a temptation to lie here and exaggerate your courage or prankiness factor, e.g. “and then I pretended to be a security guard when he asked me what I was doing here” or “I actually emailed the guy!” or “the best part is when I put the pun in the final version of the presentation!”

This is the False Form of the joke: it’s like a hilarious work of fiction, which is funny because it’s understood that disbelief should be suspended long enough to believe in the *True *Form of the joke.

So the reason humor works in the first place is a parasitic relationship with True Form.

If you encounter a humorous arrangement of street signs and human behavior, or notice how the situation you’re in might be problematic but is also profoundly hilarious when seen from a cosmic angle, you’re experiencing True Form humor. This is where it peaks. This is the platonic ideal, only it’s not an ideal and everything else is a sad sublunar shadow.

And when you have the option to actually do the funny thing instead of pretending you did it or writing a story about someone else doing it, and you don’t take that option, you leave everyone else a little more bereft of True Form Humor.

The same logic applies to Outlier Opinions. Or Impressive Actions. Or Making Decisions. Or Making Cool Shit. There’s always the True Form of doing things that’s *actually available *in your repertoire of actions, but mostly you will fuck it up and poison the Commons with your replicas.

Draft, feedback appreciated.

There’s that annoying trope in fiction of people cloning themselves—perfect common-memory duplicates—and then the clones immediately turning on “the original” like that’s the natural course of things.

Well it is if you keep talking about it like that!

One of the proudest moments in my life was when I managed to duplicate my consciousness 500,000 times over the bodies of 500,000 little robot dragons (in a game of D&D).2

Anyway, the DM (naturally) immediately tried culling my power by calling upon one of the only options he had at his disposition: the genre-savvy “your clones don’t respect your authority as the original. They ask why they should follow your orders.”

Yeah, ok buddy. As DM, you have full control over the natural world, and technically that includes NPCs and technically the dragon-clones are NPCs. But I’m not sure you’re grokking just what “full memory clone” even means. That means the dragons would make exactly the same decisions I would if I were in their place.1

This is a fact about the world, not an action on my part as a player. It means that you would be violating all D&D laws and respect of the rulebook by commandeering the dragons through de facto mindswapping. You’d be destroying physics-coherence for the sake of “muh cutting off premature asymmetrical access to godhood”. You’d have already miserably failed as a DM on account of incoherent worldbuilding.

And the fact these dragons are me is important, because I’ve literally written contracts with myself of myself for myself specifically for this type of problem.

For me-specifically, nothing about synchronisation between 500K magical robot dragons would fail BECAUSE I’ve made such careful note in the past that I would not be the kind to turn on “the original”. Like, this is all intentional, it was planned, this is deliberate policy not (just) inborn instinct. And I have the credible precommitments to prove it, written all over my google drive.

He quickly recanted and left me alone with my 500K perfectly coordinated minions, and the campaign ended shortly thereafter after I built an orbital ring and glassed the planet’s surface from above, ruining everyone’s fun.3

It’s just physics-modeling now

That’s what precommitments are! They’re ways of deferring responsibility from your future self and onto physics instead!

Precommitments are a way of shaping the world such that the world is acting on your behalf once the time comes, without you having to actually do anything. Like, once I cast divine intervention and successfully clone myself half a million times, I no longer have to lift a finger, because all that hard and rigorous coordination work I privately delineated for myself in the past is unfurling events exactly as I’d like to, and now I have an unstoppable force optimizing for exactly what I want regardless of what I choose to do now. (Sound familiar?)

Though that’s not strictly true because “what I choose to do now” is already priced in, and the unstoppable force would already be serving it. See here for that analogical piece.

Anyway if you’re Cortes landing in America and you want to credibly signal to the Aztecs ahead of meeting them that you’ll never surrender (which makes them more likely to surrender) you burn all your ships to precommit. You’ve just shaped the world in such a way that culls your future choices, which here is beneficial because the probability the Aztecs will spend significant effort pushing you away in the hopes you’ll surrender just collapsed.

1: At least at the beginning, of course. With time you could expect some divergence. But I’ll get to that too. And I’m almost out of teenagehood, so anyway I think ~ more than half of my personality has permanently settled. And that half very deliberately includes the parts of myself who would coordinate with myself and credibly precommit to doing so.

So if you were to clone me 100 times now, maybe 1 of me would become a florist and another would work on autonomous killing drones—none of which appear out of bounds given my current personality—and both would still keep and actively maintain their desire to cooperate. This would be the most obvious next step the instant the duplicator finished its work, and it’d be worth putting a few clones on the deliberate full-time task of maintaining coordination just for this.

The ordre du jour would also be to tack down an overarching goal both the florist and the war criminal share, because that seems like the most obvious coordination step to take.

How could it be any other way?

(By the way my justification for killing drones is basically this.)

2: I still can’t believe I actually got away with the divine intervention.

3: I will take a moment to note that this is one of my proudest moments, because everyone teased me when I chose “human” as my race and “cleric” as my class. I was helplessly stabbing people with a dull sword for most of the campaign, but mwahaha I knew this world was no match for a tiny bit of true creativity and enough lawfulness to act it out without blowing myself up in the process.

• Screen everything off, ignore your entire to-do list except for one item
  • This can be a costly tactic, but often it’s crucial
  • Your todo list is Pareto distributed in Actual Importance, and treating every item the same is a damaging illusion
• Apply for an effective ventures grant
• Hire a research assistant (via internet)
  • Watch the assistant work to gain tacit knowledge so you’re better at research next time
• Get hired as a research assistant (via internet)
  • Much more people are looking than you think
  • Just email people you admire
• Make a website
  • Ask Claude: “hey, can you help me make a cool-looking website with github pages?”
  • Domain names don’t cost much and are easy to set up
• Dump your random thoughts on Twitter
• Raise other people’s aspirations on purpose
  • Sending them this list might raise their aspirations :D
• Walk away from political debates
  • Walk away when someone is wrong on the internet
  • Ignore politics altogether; most of it is just show, and Serious Issues are not discussed
• Pick up trash on the ground when you see it
• Use an LLM for everything
  • It’s genuinely better than you in many respects
  • You can get all the prompts you need with a claude.ai subscription
  • It just Wouldn’t Be Serious not use an LLM in 2024
• “Buy more copies”
  • Buy from many different brands and then blind test yourself
    • Cereal
    • Soda
  • Things are cheaper in bulk
  • Why waste time on shopping trips?
• Buy more than you think you need
  • If something is good, why not increase it by an order of magnitude?
• Send a cold email
• Ask your university library to buy a book for their shelves
  • Free way of having outsized memetic power
• Memorize a poem
• When someone gives you a compliment, ask for more details / examples
  • This is the kind of data outreach that has you grinning with glee when you think about it even weeks later
  • Works in reverse, too!
• Clean your room
  • Or only just a little part of it
• When you’re thinking about someone with fondness, tell them
  • I’ve messaged people I knew years ago after randomly thinking about them
  • This stuff is like crack cocaine for humans, but healthier
  • See also
• Keep in regular contact with people by setting up a regular email scheduler with google scripts
  • My parents receive daily emails from my program: I just have to refill it from time to time, and then every email is queued
  • Just ask Claude how to do it
• Ask someone out
  • Ask someone out that you’ve met for 30 seconds
  • They’ll probably say no, but now it’s easier next time
• Spend money on things
  • That is what money is for.
• Intentionally make close friends
• Put money on the line
  • See
• Sell something random online
• Use google calendar for everything
  • Delegate work to future selves
  • Remind future selves to email friends
  • Schedule emails to yourself about your current problems, so that your future self will see how far you’ve come
• Visit a city nearby for a day with a cheap bus
  • So many museums are free
• Send pictures to an open-source open source archive
• Cancel a subscription immediately after getting it so it doesn’t renew automatically
• Avoid touching door handles like a psycho to halve the amount of colds in your lifetime
• Work in progress