In Spring of 2024, Jacob Lagerros and I took an impromptu trip to Taiwan to glean what we could about the Chip supply chain. Around the same time, I read Chip War and some other sources about the semiconductor industry.

I planned to write a blog post outlining what I learned, but I got pseudo-depressed after coming back from Taiwan, and never finished or published it. This post is a lightly edited version of the draft that has been sitting in my documents folder. (I had originally intended to include a lot more than this, but I might as well publish what I have.)

Interestingly, reading it now, all of this feels so basic that I’m surprised that I considered a lot of it worth including in a post like this, but I think it was all new to me at the time.

• There are important differences between logic chips and memory chips, such that at various times, companies have specialized in one or the other.
• TSMC was founded by Morris Chang, with the backing of the Taiwanese government. But the original impetus came from Taiwan, not from Chang. The government decided that it wanted to become a leading semiconductor manufacturer, and approached Chang (who had been an engineer and executive at Texa

In a private slack someone extended credit to Sam Altman for putting EAs on the on the OpenAI board originally, especially that this turned out to be pretty risky / costly for him.

I responded:

It seems to me that there were AI safety people on the board at all is fully explainable by strategic moves from an earlier phase of the game.

Namely, OpenAI traded a boardseat for OpenPhil grant money, and more importantly, OpenPhil endorsement, which translated into talent sourcing and effectively defused what might have been vocal denouncement from one of the major intellectually influential hubs of the world.

No one knows how counterfactual history might have developed, but it doesn’t seem unreasonable to think that there is an external world in which the EA culture successfully created a narrative that groups trying to build AGI were bad and defecting.

He’s the master at this game and not me, but I would bet at even odds that Sam was actively tracking EA as a potential social threat that could dampen OpenAI’s narrative flywheel.

I don’t know that OpenPhil’s grant alone was sufficient to switch from the “EAs vocally decry OpenAI as making the world worse” equilibrium to a “largely (but not uni

Back in January, I participated in a workshop in which the attendees mapped out how they expect AGI development and deployment to go. The idea was to start by writing out what seemed most likely to happen this year, and then condition on that, to forecast what seems most likely to happen in the next year, and so on, until you reach either human disempowerment or an end of the acute risk period.

This post was my attempt at the time.

I spent maybe 5 hours on this, and there's lots of room for additional improvement. This is not a confident statement of how I think things are most likely to play out. There are already some ways in which I think this projection is wrong. (I think it's too fast, for instance). But nevertheless I'm posting it now, with only a few edits and elaborations, since I'm probably not going to do a full rewrite soon.

2024

• A model is released that is better than GPT-4. It succeeds on some new benchmarks. Subjectively, the jump in capabilities feels smaller than that between RLHF’d GPT-3 and RLHF’d GPT-4. It doesn’t feel as shocking the way chat-GPT and GPT-4 did, for either x-risk focused folks, or for the broader public. Mostly it feels like “a somewhat better langua

An optional feature that I think LessWrong should have: shortform posts that get more than some amount of karma get automatically converted into personal blog posts, including all the comments.

It should have a note at the top "originally published in shortform", with a link to the shortform comment. (All the copied comments should have a similar note).

I think that, in almost full generality, we should taboo the term "values". It's usually ambiguous between a bunch of distinct meanings.

• The ideals that, when someone contemplates, invoke strong feelings (of awe, motivation, excitement, exultation, joy, etc.)
• The incentives of an agent in a formalized game with quantified payoffs.
• A utility function - one's hypothetical ordering over words, world-trajectories, etc, that results from comparing each pair and evaluating which one is better.
• A person's revealed preferences.
• The experiences and activities that a person likes for their own sake.
• A person's vision of an ideal world. (Which, I claim, often reduces to "an imagined world that's aesthetically appealing.")
• The goals that are at the root of a chain or tree of instrumental goals.
  • [This often comes with an implicit or explicit implication that most of human behavior has that chain/tree structure, as opposed being, for instance, mostly hardcoded adaptions, or a chain/tree of goals that grounds out in a mess of hardcoded adaptions instead of anything goal-like.]
• The goals/narratives that give meaning to someone's life.
  • [It can be the case almost all one's meaning can come through a particula

This post is a snapshot of what currently “feels realistic” to me regarding how AI will go. That is, these are not my considered positions, or even provisional conclusions informed by arguments. Rather, if I put aside all the claims and arguments and just ask “which scenario feels like it is ‘in the genera of reality’?”, this is what I come up with. I expect to have different first-order impressions in a month.

Crucially, none of the following is making claims about the intelligence explosion, and the details of the intelligence explosion (where AI development goes strongly recursive) are crucial to the long run equilibrium of the earth-originating civilization.

My headline: we’ll mostly succeed at prosaic alignment of human-genius level AI agents

• Takeoff will continue to be gradual. We’ll get better models and more capable agents year by year, but not jumps that are bigger than that between Claude 3.7 and Claude 4.
• Our behavioral alignment patches will work well enough.
  • RL will induce all kinds of reward hacking and related misbehavior, but we’ll develop patches for those problems (most centrally, for any given reward hack, we’ll generate some examples and counter examples to include i

New post: Some things I think about Double Crux and related topics

I've spent a lot of my discretionary time working on the broad problem of developing tools for bridging deep disagreements and transferring tacit knowledge. I'm also probably the person who has spent the most time explicitly thinking about and working with CFAR's Double Crux framework. It seems good for at least some of my high level thoughts to be written up some place, even if I'm not going to go into detail about, defend, or substantiate, most of them.

The following are my own beliefs and do not necessarily represent CFAR, or anyone else.

I, of course, reserve the right to change my mind.

[Throughout I use "Double Crux" to refer to the Double Crux technique, the Double Crux class, or a Double Crux conversation, and I use "double crux" to refer to a proposition that is a shared crux for two people in a conversation.]

Here are some things I currently believe:

(General)

1. Double Crux is one (highly important) tool/ framework among many. I want to distinguish between the the overall art of untangling and resolving deep disagreements and the Double Crux tool in particular. The Double Crux framework is maybe the most

Old post: RAND needed the "say oops" skill

[Epistemic status: a middling argument]

A few months ago, I wrote about how RAND, and the “Defense Intellectuals” of the cold war represent another precious datapoint of “very smart people, trying to prevent the destruction of the world, in a civilization that they acknowledge to be inadequate to dealing sanely with x-risk.”

Since then I spent some time doing additional research into what cognitive errors and mistakes  those consultants, military officials, and politicians made that endangered the world. The idea being that if we could diagnose which specific irrationalities they were subject to, that this would suggest errors that might also be relevant to contemporary x-risk mitigators, and might point out some specific areas where development of rationality training is needed.

However, this proved somewhat less fruitful than I was hoping, and I’ve put it aside for the time being. I might come back to it in the coming months.

It does seem worth sharing at least one relevant anecdote, from Daniel Ellsberg’s excellent book, the Doomsday Machine, and analysis, given that I’ve already written it up.

The missile gap

In the late nineteen-fi... (read more)

New post: What is mental energy?

[Note: I’ve started a research side project on this question, and it is already obvious to me that this ontology importantly wrong.]

There’s a common phenomenology of “mental energy”. For instance, if I spend a couple of hours thinking hard (maybe doing math), I find it harder to do more mental work afterwards. My thinking may be slower and less productive. And I feel tired, or drained, (mentally, instead of physically).

Mental energy is one of the primary resources that one has to allocate, in doing productive work. In almost all cases, humans have less mental energy than they have time, and therefore effective productivity is a matter of energy management, more than time management. If we want to maximize personal effectiveness, mental energy seems like an extremely important domain to understand. So what is it?

The naive story is that mental energy is an actual energy resource that one expends and then needs to recoup. That is, when one is doing cognitive work, they are burning calories, depleting their bodies energy stores. As they use energy, they have less fuel to burn.

My current understanding is that this story is not physiologically realistic. T... (read more)

New post: Some notes on Von Neumann, as a human being

I recently read Prisoner’s Dilemma, which half an introduction to very elementary game theory, and half a biography of John Von Neumann, and watched this old PBS documentary about the man.

I’m glad I did. Von Neumann has legendary status in my circles, as the smartest person ever to live. [1] Many times I’ve written the words “Von Neumann Level Intelligence” in a AI strategy document, or speculated about how many coordinated Von Neumanns would it take to take over the world. (For reference, I now think that 10 is far too low, mostly because he didn’t seem to have the entrepreneurial or managerial dispositions.)

Learning a little bit more about him was humanizing. Yes, he was the smartest person ever to live, but he was also an actual human being, with actual human traits.

Watching this first clip, I noticed that I was surprised by a number of thing.

1. That VN had an accent. I had known that he was Hungarian, but somehow it had never quite propagated that he would speak with a Hungarian accent.
2. That he was middling height (somewhat shorter than the presenter he’s talking too).
3. The thing he is saying is the sort of thing that I would

TL;DR: I’m offering to help people productively have difficult conversations and resolve disagreements, for free. Feel free to email me if and when that seems helpful. elitrye [at] gmail.com

Facilitation

Over the past 4-ish years, I’ve had a side project of learning, developing, and iterating on methods for resolving tricky disagreements, and failures to communicate. A lot of this has been in the Double Crux frame, but I’ve also been exploring a number of other frameworks (including, NVC, Convergent Facilitation, Circling-inspired stuff, intuition extraction, and some home-grown methods).

As part of that, I’ve had a standing offer to facilitate / mediate tricky conversations for folks in the CFAR and MIRI spheres (testimonials below). Facilitating “real disagreements”, allows me to get feedback on my current conversational frameworks and techniques. When I encounter blockers that I don’t know how to deal with, I can go back to the drawing board to model those problems and interventions that would solve them, and iterate from there, developing new methods.

I generally like doing this kind of conversational facilitation and am open to do... (read more)

That no one rebuilt old OkCupid updates me a lot about how much the startup world actually makes the world better

The prevailing ideology of San Francisco, Silicon Valley, and the broader tech world, is that startups are an engine (maybe even the engine) that drives progress towards a future that's better than the past, by creating new products that add value to people's lives.

I now think this is true in a limited way. Software is eating the world, and lots of bureaucracy is being replaced by automation which is generally cheaper, faster, and a better UX. But I now think that this narrative is largely propaganda.

It's been 8 years since Match bought and ruined OkCupid and no one, in the whole tech ecosystem, stepped up to make a dating app even as good as old OkC is a huge black mark against the whole SV ideology of technology changing the world for the better.

Finding a partner is such a huge, real, pain point for millions of people. The existing solutions are so bad and extractive. A good solution has already been demonstrated. And yet not a single competent founder wanted to solve that problem for planet earth, instead of doing something else, that (arguably) would have been more p... (read more)

Dumb question: Why doesn't using constitutional AI, where the constitution is mostly or entirely corrigibility produce a corrigible AI (at arbitrary capability levels)?

My dumb proposal:

1. Train a model in something like o1's RL training loop, with a scratch pad for chain of thought, and reinforcement of correct answers to hard technical questions across domains.

2. Also, take those outputs, prompt the model to generate versions of those outputs that "are more corrigible / loyal / aligned to the will of your human creators". Do backprop to reinforce those more corrigible outputs.

Possibly "corrigibility" applies only very weakly to static solutions, and so for this setup to make sense, we'd instead need to train on plans, or time-series of an AI agent's actions: The AI agent takes a bunch of actions over the course of a day or a week, then we have an AI annotate the time series of action-steps with alternative action-steps that better reflect "corrigibility", according to its understanding. Then we do backprop to so that the Agent behaves more in ways that are closer to the annotated action transcript.

Would this work to produce a corrigible agent? If not, why not? 

There's a further question of "how much less capable will the more corrigible AI be?" This might be a significant penalty to performance, and so the added safety gets eroded away in the competitive crush. But first and foremost, I want to know if something like this could work.

(Reasonably personal)

I spend a lot of time trying to build skills, because I want to be awesome. But there is something off about that.

I think I should just go after things that I want, and solve the problems that come up on the way. The idea of building skills sort of implies that if I don't have some foundation or some skill, I'll be blocked, and won't be able to solve some thing in the way of my goals.

But that doesn't actually sound right. Like it seems like the main important thing for people who do incredible things is their ability to do problem solving on the things that come up, and not the skills that they had previously built up in a "skill bank".

Raw problem solving is the real thing and skills are cruft. (Or maybe not cruft per se, but more like a side effect. The compiled residue of previous problem solving. Or like a code base from previous project that you might repurpose.)

Part of the problem with this is that I don't know what I want for my own sake, though. I want to be awesome, which in my conception, means being able to do things.

I note that wanting "to be able to do things" is a leaky sort of motivation: because the... (read more)

I think I no longer buy this comment of mine from almost 3 years ago. Or rather I think it's pointing at a real thing, but I think it's slipping in some connotations that I don't buy.

What I expect to see is agents that have a portfolio of different drives and goals, some of which are more like consequentialist objectives (eg "I want to make the number in this bank account go up") and some of which are more like deontological injunctions ("always check with my user/ owner before I make a big purchase or take a 'creative' action, one that is outside of my training distribution").

My prediction is that the consequentialist parts of the agent will basically route around any deontological constraints that are trained in. 

For instance, the your personal assistant AI does ask your permission before it does anything creative, but also, it's superintelligently persuasive and so it always asks your permission in exactly the way that will result in it accomplishing what it wants. If there are a thousand action sequences in which it asks for permission, it picks the one that has the highest expected value with regard to whatever it wants. This basically nullifies the safety benef

When I wrote up some notes about Moldbug’s political philosophy last year, it seemed (when you strip away a bunch of flavor-text and non-load-bearing details) to reduce to a proposal to impose market discipline on governments by having them compete for citizens. I ended with the question “wait, how is Yarvin’s proposal any different than Anarchocapitalism?” They sound like they’re basically the same.” (I have since removed that line from the post, but it’s still there in the revision history.)

A few weeks ago, I read most of David Friedman’s The Machinery of Freedom and incidentally, I now know the answer to that question

The character of the overall political system—whether anarcho-capitalism degrade back into a collection of microstates— depends on the geographically-localized economies of scale to rights protection.

If rights-protection doesn’t benefit from large geographically localized economies of scale, we could end up with an anarcho-capitalist equilibrium of many different rights protection companies serving a similar local, and competing to better serve their customers, and generally relying on arbitration to settle disputes peacefully.

But if it’s a service that is sufficien... (read more)

I’m no longer sure that I buy dutch book arguments, in full generality, and this makes me skeptical of the "utility function" abstraction

Thesis: I now think that utility functions might be a pretty bad abstraction for thinking about the behavior of agents in general including highly capable agents.

[Epistemic status: half-baked, elucidating an intuition. Possibly what I’m saying here is just wrong, and someone will helpfully explain why.]

Over the past years, in thinking about agency and AI, I’ve taken the concept of a “utility function” for granted as the natural way to express an entity's goals or preferences. 

Of course, we know that humans don’t have well defined utility functions (they’re inconsistent, and subject to all kinds of framing effects), but that’s only because humans are irrational. To the extent that a thing acts like an agent, it’s behavior corresponds to some utility function. That utility function might not be explicitly represented, but if an agent is rational, there’s some utility function that reflects it’s preferences. 

Given this, I might be inclined to scoff at people who scoff at “blindly maximizing” AGIs. “They just don’t get it”, I might think. “T... (read more)

Eliezer claims that dath ilani never give in to threats. But I'm not sure I buy it.

The only reason people will make threats against you, the argument goes, is if those people expect that you might give in. If you have an iron-clad policy against acting in response to threats made against you, then there's no point in making or enforcing the threats in the first place. There's no reason for the threatener to bother, so they don't. Which means in some sufficiently long run, refusing to submit to threats means you're not subject to threats.

This seems a bit fishy to me. I have a lingering suspicion that this argument doesn't apply, or at least doesn't apply universally, in the real world.

I'm thinking here mainly of a prototypical case of an isolated farmer family (like the early farming families of the greek peninsula, not absorbed into a polis), being accosted by some roving bandits, such as the soldiers of the local government. The bandits say "give us half your harvest, or we'll just kill you."

The argument above depends on a claim about the cost of executing on a threat. "There's no reason to bother" implies that the threatener has a preference not to bother, if they know that the t... (read more)

Old post: A mechanistic description of status

[This is an essay that I’ve had bopping around in my head for a long time. I’m not sure if this says anything usefully new-but it might click with some folks. If you haven’t read Social Status: Down the Rabbit Hole on Kevin Simler’s excellent blog, Melting Asphalt read that first. I think this is pretty bad and needs to be rewritten and maybe expanded substantially, but this blog is called “musings and rough drafts.”]

In this post, I’m going to outline how I think about status. In particular, I want to give a mechanistic account of how status necessarily arises, given some set of axioms, in much the same way one can show that evolution by natural selection must necessarily occur given the axioms of 1) inheritance of traits 2) variance in reproductive success based on variance in traits and 3) mutation.

(I am not claiming any particular skill at navigating status relationships, any more than a student of sports-biology is necessarily a skilled basketball player.)

By “status” I mean prestige-status.

Axiom 1: People have goals.

That is, for any given human, there are some things that they want. This can include just about anything. You might wan... (read more)

In his TED talk, Eliezer guesses superintelligence will arrive after "zero to two more breakthroughs the size of transformers." I've heard others voice similar takes. But I haven't heard much discussion of which number it is.^[1]

There is an enormous difference between the world where there 0 insights left before superintelligence, and the world in which we have one or more. Specifically, this is the difference between a soft or a hard takeoff, because of what we might call a "cognitive capability overhang".

The current models are already superhuman in a several notable ways:

• Vastly superhuman breadth of knowledge
• Effectively superhuman working memory
• Superhuman thinking speed^[2]

If there's a secret sauce that is missing for "full AGI", then the first AGI might have all of these advantages, and more, out of the gate.

It seems to me that there are at least two possibilities.

We may be in world A:

We've already discovered all the insights and invented the techniques that earth is going to use to create its first superintelligence in this timeline. It's something like transformers pre-trained on internet corpuses, and then trained using RL from verifiable feedback and on synthetic data generate

So it seems like one way that the world could go is:

• China develops a domestic semiconductor fab industry that's not at the cutting edge, but close, so that it's less dependent on Taiwan's TSMC
   
• China invades Taiwan, destroying TSMC, ending up with a compute advantage over the US, which translates into a military advantage
   
• (which might or might not actually be leveraged in a hot war).

I could imagine China building a competent domestic chip industry. China seems more determined to do that than the US is.

Though notably, China is not on track to do that currently. It's not anywhere close to it's goal producing 70% it's chips, by 2025.

And if the US was serious about building a domestic cutting-edge chip industry again, could it? I basically don't think that American work culture can keep up with Taiwanese/TSMC work culture, in this super-competitive industry.

TSMC is building fabs in the US, but from what I hear, they're not going well.

(While TSMC is a Taiwanese company, having a large fraction of TSMC fabs in in the US would preement the scenario above. TSMC fabs in the US counts as "a domestic US chip industry.")

Building and running leading node fabs is just a really really hard thing to do.

I guess the most likely status scenario is the continuation of the status quo where China and the US continue to both awkwardly depend on TSMC's chips for crucial military and economic AI tech.

Something that I've been thinking about lately is the possibility of an agent's values being partially encoded by the constraints of that agent's natural environment, or arising from the interaction between the agent and environment.

That is, an agent's environment puts constraints on the agent. From one perspective removing those constraints is always good, because it lets the agent get more of what it wants. But sometimes from a different perspective, we might feel that with those constraints removed, the agent goodhearts or wire-heads, or otherwise fails to actualize its "true" values.

The Generator freed from the oppression of the Discriminator

As a metaphor: if I'm one half of a GAN, let's say the generator, then in one sense my "values" are fooling the discriminator, and if you make me relatively more powerful than my discriminator, and I dominate it...I'm loving it, and also no longer making good images.

But you might also say, "No, wait. That is a super-stimulus, and actually what you value is making good images, but half of that value was encoded in your partner."

This second perspective seems a little stupid to me. A little too Aristotelian. I mean if we're going to take that ... (read more)