I’m an AGI safety / AI alignment researcher in Boston with a particular focus on brain algorithms. Research Fellow at Astera. See https://sjbyrnes.com/agi.html for a summary of my research and sorted list of writing. Physicist by training. Email: steven.byrnes@gmail.com. Leave me anonymous feedback here. I’m also at: RSS feed, X/Twitter, Bluesky, LinkedIn, and more at my website.
Steven Byrnes
Karma: 21,639
[CLAIMED!]
ODD JOB OFFER: I think I want to cross-post Intro to Brain-Like-AGI Safety as a giant 200-page PDF on arxiv (it’s about 80,000 words), mostly to make it easier to cite (as is happening sporadically, e.g. here, here). I am willing to pay fair market price for whatever reformatting work is necessary to make that happen, which I don’t really know (make me an offer). I guess I’m imagining that the easiest plan would be to copy everything into Word (or LibreOffice Writer), clean up whatever formatting weirdness comes from that, and convert to PDF. LaTeX conversion is also acceptable but I imagine that would be much more work for no benefit.I think inline clickable links are fine on arxiv (e.g. page 2 here), but I do have a few references to actual papers, and I assume those should probably be turned into a proper reference section at the end of each post / “chapter”. Within-series links (e.g. a link from Post 6 to a certain section of Post 4) should probably be converted to internal links within the PDF, rather than going out to the lesswrong / alignmentforum version. There are lots of textbooks / lecture notes on arxiv which can serve as models; I don’t really know the details myself. The original images are all in Powerpoint, if that’s relevant. The end product should ideally be easy for me to edit if I find things I want to update. (Arxiv makes updates very easy, one of my old arxiv papers is up to version 5.)
…Or maybe this whole thing is stupid. If I want it to be easier to cite, I could just add in a “citation information” note, like they did here? I dunno.
Davidad responds with a brief argument for 1000 FLOP-equivalent per synapse-second (3 OOM more than my guess) on X as follows:
Ok, so assuming we agree on 1e14 synapses and 3e8 seconds, then where we disagree is on average FLOP(-equivalent) per synapse-second: you think it’s about 1, I think it’s about 1000. This is similar to the disagreement you flagged with Joe Carlsmith.
Note: at some point Joe interviewed me about this so there might be some double-counting of “independent” estimates here, but iirc he also interviewed many other neuroscientists.
My estimate would be a lot lower if we were just talking about “inference” rather than learning and memory. STDP seems to have complex temporal dynamics at the 10ms scale.
There also seem to be complex intracellular dynamics at play, possibly including regulatory networks, obviously regarding synaptic weight but also other tunable properties of individual compartments.
The standard arguments for the causal irrelevance of these to cognition (they’re too slow to affect the “forward pass”) don’t apply to learning. I’m estimating there’s like a 10-dimensional dynamical system in each compartment evolving at ~100Hz in importantly nonlinear ways.
I think OP is using “sequential” in an expansive sense that also includes e.g. “First I learned addition, then I learned multiplication (which relies on already understanding addition), then I learned the distributive law (which relies on already understanding both addition and multiplication), then I learned the concept of modular arithmetic (which relies on …) etc. etc.” (part of what OP calls “C”). I personally wouldn’t use the word ‘sequential’ for that—I prefer a more vertical metaphor like ‘things building upon other things’—but that’s a matter of taste I guess. Anyway, whatever we want to call it, humans can reliably do a great many steps, although that process unfolds over a long period of time.
…And not just smart humans. Just getting around in the world, using tools, etc., requires giant towers of concepts relying on other previously-learned concepts.
Obviously LLMs can deal with addition and multiplication and modular arithmetic etc. But I would argue that this tower of concepts building on other concepts was built by humans, and then handed to the LLM on a silver platter. I join OP in being skeptical that LLMs (including o3 etc.) could have built that tower themselves from scratch, the way humans did historically. And I for one don’t expect them to be able to do that thing until an AI paradigm shift happens.
Self-dialogue: Do behaviorist rewards make scheming AGIs?
In case anyone missed it, I stand by my reply from before— Applying traditional economic thinking to AGI: a trilemma
If you offer a salary below 100 watts equivalent, humans won’t accept, because accepting it would mean dying of starvation. (Unless the humans have another source of wealth, in which case this whole discussion is moot.) This is not literally a minimum wage, in the conventional sense of a legally-mandated wage floor; but it has the same effect as a minimum wage, and thus we can expect it to have the same consequences as a minimum wage.
This is obviously (from my perspective) the point that Grant Slatton was trying to make. I don’t know whether Ben Golub misunderstood that point, or was just being annoyingly pedantic. Probably the former—otherwise he could have just spelled out the details himself, instead of complaining, I figure.
It was Grant Slatton but Yudkowsky retweeted it
I like reading the Sentinel email newsletter once a week for time-sensitive general world news, and https://en.wikipedia.org/wiki/2024 (or https://en.wikipedia.org/wiki/2025 etc.) once every 3-4 months for non-time-sensitive general world news. That adds up to very little time—maybe ≈1 minute per day on average—and I think there are more than enough diffuse benefits to justify that tiny amount of time.
I feel like I’ve really struggled to identify any controllable patterns in when I’m “good at thinky stuff”. Gross patterns are obvious—I’m reliably great in the morning, then my brain kinda peters out in the early afternoon, then pretty good again at night—but I can’t figure out how to intervene on that, except scheduling around it.
I’m extremely sensitive to caffeine, and have a complicated routine (1 coffee every morning, plus in the afternoon I ramp up from zero each weekend to a full-size afternoon tea each Friday), but I’m pretty uncertain whether I’m actually getting anything out of that besides a mild headache every Saturday.
I wonder whether it would be worth investing the time and energy into being more systematic to suss out patterns. But I think my patterns would be pretty subtle, whereas yours sound very obvious and immediate. Hmm, is there an easy and fast way to quantify “CQ”? (This pops into my head but seems time-consuming and testing the wrong thing.) …I’m not really sure where to start tbh.
…I feel like what I want to measure is a 1-dimensional parameter extremely correlated with “ability to do things despite ugh fields”—presumably what I’ve called “innate drive to minimize voluntary attention control” being low a.k.a. “mental energy” being high. Ugh fields are where the parameter is most obvious to me but it also extends into thinking well about other topics that are not particularly aversive, at least for me, I think.
Sorry if I missed it, but you don’t seem to address the standard concern that mildly-optimizing agents tend to self-modify into (or create) strongly-optimizing agents.
For example (copying from my comment here), let’s say we make an AI that really wants there to be exactly 100 paperclips in the bin. There’s nothing else it wants or desires. It doesn’t care a whit about following human norms, etc.
But, there’s one exception: this AI is also “lazy”—every thought it thinks, and every action it takes, is mildly aversive. So it’s not inclined to, say, build an impenetrable fortress around the bin just for an infinitesimal probability increment. “Seems like a lot of work! It’s fine as is,” says the AI to itself.
But hey, here’s something it can do: rent some server time on AWS, and make a copy its own source code and trained model, but comment out the “laziness” code block. That’s not too hard; even a “lazy” AI would presumably be capable of doing that. And the result will be a non-lazy AI that works tirelessly and uncompromisingly towards incrementing the probability of there being 32 paperclips—first 99.99%, then 99.9999%, etc. That’s nice! (from the original AI’s perspective). Or more specifically, it offers a small benefit for zero cost (from the original AI’s perspective).
It’s not wildly different from a person saying “I want to get out of debt, but I can’t concentrate well enough to hold down a desk job, so I’m going to take Adderall”. It’s an obvious solution to a problem.
…OK, in this post, you don’t really talk about “AI laziness” per se, I think, instead you talk about “AI getting distracted by other things that now seem to be a better use of its time”, i.e. other objectives. But I don’t think that changes anything. The AI doesn’t have to choose between building an impenetrable fortress around the bin of paperclips versus eating lunch. “Why not both?”, it says. So the AI eats lunch while its strongly-optimizing subagent simultaneously builds the impenetrable fortress. Right?
I’m still curious about how you’d answer my question above. Right now, we don’t have ASI. Sometime in the future, we will. So there has to be some improvement to AI technology that will happen between now and then. My opinion is that this improvement will involve AI becoming (what you describe as) “better at extrapolating”.
If that’s true, then however we feel about getting AIs that are “better at extrapolating”—its costs and its benefits—it doesn’t much matter, because we’re bound to get those costs and benefits sooner or later on the road to ASI. So we might as well sit tight and find other useful things to do, until such time as the AI capabilities researchers figure it out.
…Furthermore, I don’t think the number of months or years between “AIs that are ‘better at extrapolating’” and ASI is appreciably larger if the “AIs that are ‘better at extrapolating’” arrive tomorrow, versus if they arrive in 20 years. In order to believe that, I think you would need to expect some second bottleneck standing between “AIs that are ‘better at extrapolating’”, and ASI, such that that second bottleneck is present today, but will not be present (as much) in 20 years, and such that the second bottleneck is not related to “extrapolation”.
I suppose that one could argue that availability of compute will be that second bottleneck. But I happen to disagree. IMO we already have an absurdly large amount of compute overhang with respect to ASI, and adding even more compute overhang in the coming decades won’t much change the overall picture. Certainly plenty of people would disagree with me here. …Although those same people would probably say that “just add more compute” is actually the only way to make AIs that are “better at extrapolation”, in which case my point would still stand.
I don’t see any other plausible candidates for the second bottleneck. Do you? Or do you disagree with some other part of that? Like, do you think it’s possible to get all the way to ASI without ever making AIs “better at extrapolating”? IMO it would hardly be worthy of the name “ASI” if it were “bad at extrapolating” :)
Because you can speed up AI capabilities much easier while being sloppy than to produce actually good alignment ideas.
Right, my point is, I don’t see any difference between “AIs that produce slop” and “weak AIs” (a.k.a. “dumb AIs”). So from my perspective, the above is similar to : “…Because weak AIs can speed up AI capabilities much easier than they can produce actually good alignment ideas.”
…And then if you follow through the “logic” of this OP, then the argument becomes: “AI alignment is a hard problem, so let’s just make extraordinarily powerful / smart AIs right now, so that they can solve the alignment problem”.
See the error?
If you really think you need to be similarly unsloppy to build ASI than to align ASI, I’d be interested in discussing that. So maybe give some pointers to why you might think that (or tell me to start).
I don’t think that. See the bottom part of the comment you’re replying to. (The part after “Here’s what I would say instead:”)
I think it’s 1:1, because I think the primary bottleneck to dangerous ASI is the ability to develop coherent and correct understandings of arbitrary complex domains and systems (further details), which basically amounts to anti-slop.
If you think the primary bottleneck to dangerous ASI is not that, but rather something else, then what do you think it is? (or it’s fine if you don’t want to state it publicly)
Right, so one possibility is that you are doing something that is “speeding up the development of AIS-helpful capabilities” by 1 day, but you are also simultaneously speeding up “dangerous capabilities” by 1 day, because they are the same thing.
If that’s what you’re doing, then that’s bad. You shouldn’t do it. Like, if AI alignment researchers want AI that produces less slop and is more helpful for AIS, we could all just hibernate for six months and then get back to work. But obviously, that won’t help the situation.
And a second possibility is, there are ways to make AI more helpful for AI safety that are not simultaneously directly addressing the primary bottlenecks to AI danger. And we should do those things.
The second possibility is surely true to some extent—for example, the LessWrong JargonBot is marginally helpful for speeding up AI safety but infinitesimally likely to speed up AI danger.
I think this OP is kinda assuming that “anti-slop” is the second possibility and not the first possibility, without justification. Whereas I would guess the opposite.
I don’t think your model hangs together, basically because I think “AI that produces slop” is almost synonymous with “AI that doesn’t work very well”, whereas you’re kinda treating AI power and slop as orthogonal axes.
For example, from comments:
Two years later, GPT7 comes up with superhumanly-convincing safety measures XYZ. These inadequate standards become the dominant safety paradigm. At this point if you try to publish “belief propagation” it gets drowned out in the noise anyway.
Some relatively short time later, there are no humans.
I think that, if there are no humans, then slop must not be too bad. AIs that produce incoherent superficially-appealing slop are not successfully accomplishing ambitious nontrivial goals right?
(Or maybe you’re treating it as a “capabilities elicitation” issue? Like, the AI knows all sorts of things, but when we ask, we get sycophantic slop answers? But then we should just say that the AI is mediocre in effect. Even if there’s secretly a super-powerful AI hidden inside, who cares? Unless the AI starts scheming, but I thought AI scheming was out-of-scope for this post.)
Anti-slop AI helps everybody make less mistakes. Sloppy AI convinces lots of people to make more mistakes.
I would have said “More powerful AI (if aligned) helps everybody make less mistakes. Less powerful AI convinces lots of people to make more mistakes.” Right?
And here’s a John Wentworth excerpt:
So the lab implements the non-solution, turns up the self-improvement dial, and by the time anybody realizes they haven’t actually solved the superintelligence alignment problem (if anybody even realizes at all), it’s already too late.
If the AI is producing slop, then why is there a self-improvement dial? Why wouldn’t its self-improvement ideas be things that sound good but don’t actually work, just as its safety ideas are?
Really, I think John Wentworth’s post that you’re citing has a bad framing. It says: the concern is that early transformative AIs produce slop.
Here’s what I would say instead:
Figuring out how to build aligned ASI is a harder technical problem than just building any old ASI, for lots of reasons, e.g. the latter allows trial-and-error. So we will become capable of building ASI sooner than we’ll have a plan to build aligned ASI.
Whether the “we” in that sentence is just humans, versus humans with the help of early transformative AI assistance, hardly matters.
But if we do have early transformative AI assistants, then the default expectation is that they will fail to solve the ASI alignment problem until it’s too late. Maybe those AIs will fail to solve the problem by outputting convincing-but-wrong slop, or maybe they’ll fail to solve it by outputting “I don’t know”, or maybe they’ll fail to solve it by being misaligned, a.k.a. a failure of “capabilities elicitation”. Who cares? What matters is that they fail to solve it. Because people (and/or the early transformative AI assistants) will build ASI anyway.
For example, Yann LeCun doesn’t need superhumanly-convincing AI-produced slop, in order to mistakenly believe that he has solved the alignment problem. He already mistakenly believes that he has solved the alignment problem! Human-level slop was enough. :)
In other words, suppose we’re in a scenario with “early transformative AIs” that are up to the task of producing more powerful AIs, but not up to the task of solving ASI alignment. You would say to yourself: “if only they produced less slop”. But to my ears, that’s basically the same as saying “we should creep down the RSI curve, while hoping that the ability to solve ASI alignment emerges earlier than the breakdown of our control and alignment measures and/or ability to take over”.
…Having said all that, I’m certainly in favor of thinking about how to get epistemological help from weak AIs that doesn’t give a trivial affordance for turning the weak AIs into very dangerous AIs. For for that matter, I’m in favor of thinking about how to get epistemological help from any method, whether AI or not. :)
Yeah, I’ve written about that in §2.7.3 here.
I kinda want to say that there are many possible future outcomes that we should feel happy about. It’s true that many of those possible outcomes would judge others of those possible outcomes to be a huge missed opportunity, and that we’ll be picking from this set somewhat arbitrarily (if all goes well), but oh well, there’s just some irreducible arbitrariness is the nature of goodness itself.
For things like solving coordination problems, or societal resilience against violent takeover, I think it can be important that most people, or even virtually all people, are making good foresighted decisions. For example, if we’re worried about a race-to-the-bottom on AI oversight, and half of relevant decisionmakers allow their AI assistants to negotiate a treaty to stop that race on their behalf, but the other half think that’s stupid and don’t participate, then that’s not good enough, there will still be a race-to-the-bottom on AI oversight. Or if 50% of USA government bureaucrats ask their AIs if there’s a way to NOT outlaw testing people for COVID during the early phases of the pandemic, but the other 50% ask their AIs how best to follow the letter of the law and not get embarrassed, then the result may well be that testing is still outlawed.
For example, in this comment, Paul suggests that if all firms are “aligned” with their human shareholders, then the aligned CEOs will recognize if things are going in a long-term bad direction for humans, and they will coordinate to avoid that. That doesn’t work unless EITHER the human shareholders—all of them, not just a few—are also wise enough to be choosing long-term preferences and true beliefs over short-term preferences and motivated reasoning, when those conflict. OR unless the aligned CEOs—again, all of them, not just a few—are injecting the wisdom into the system, putting their thumbs on the scale, by choosing, even over the objections of the shareholders, their long-term preferences and true beliefs over short-term preferences and motivated reasoning.
I don’t think the average person would be asking AI what are the best solutions for preventing existential risks. As evidence, just look around:
There are already people with lots of money and smart human research assistants. How many of those people are asking those smart human research assistants for solutions to prevent existential risks? Approximately zero.
Here’s another: The USA NSF and NIH are funding many of the best scientists in the world. Are they asking those scientists for solutions to prevent existential risk? Nope.
Demis Hassabis is the boss of a bunch of world-leading AI experts, with an ability to ask them to do almost arbitrary science projects. Is he asking them to do science projects that reduce existential risk? Well, there’s a DeepMind AI alignment group, which is great, but other than that, basically no. Instead he’s asking his employees to cure diseases (cf Isomorphic Labs), and to optimize chips, and do cool demos, and most of all to make lots of money for Alphabet.
You think Sam Altman would tell his future powerful AIs to spend their cycles solving x-risk instead of making money or curing cancer? If so, how do you explain everything that he’s been saying and doing for the past few years? How about Mark Zuckerberg and Yann LeCun? How about random mid-level employees in OpenAI? I am skeptical.
Also, even if the person asked the AI that question, then the AI would (we’re presuming) respond: “preventing existential risks is very hard and fraught, but hey, what if I do a global mass persuasion campaign…”. And then I expect the person would reply “wtf no, don’t you dare, I’ve seen what happens in sci-fi movies when people say yes to those kinds of proposals.” And then the AI would say “Well I could try something much more low-key and norm-following but it probably won’t work”, and the person would say “Yeah do that, we’ll hope for the best.” (More such examples in §1 here.)
I’m not sure if this is what you’re looking for, but here’s a fun little thing that came up recently I was when writing this post:
Summary: “Thinking really hard for five seconds” probably involves less primary metabolic energy expenditure than scratching your nose. (Some people might find this obvious, but other people are under a mistaken impression that getting mentally tired and getting physically tired are both part of the same energy-preservation drive. My belief, see here, is that the latter comes from an “innate drive to minimize voluntary motor control”, the former from an unrelated but parallel “innate drive to minimize voluntary attention control”.)
Model: The net extra primary metabolic energy expenditure required to think really hard for five seconds, compared to daydreaming for five seconds, may well be zero. For an upper bound, Raichle & Gusnard 2002 says “These changes are very small relative to the ongoing hemodynamic and metabolic activity of the brain. Attempts to measure whole brain changes in blood flow and metabolism during intense mental activity have failed to demonstrate any change. This finding is not entirely surprising considering both the accuracy of the methods and the small size of the observed changes. For example, local changes in blood flow measured with PET during most cognitive tasks are often 5% or less.” So it seems fair to assume it’s <<5% of the ≈20 W total, which gives <<1 W × 5 s = 5 J. Next, for comparison, what is the primary metabolic energy expenditure from scratching your nose? Well, for one thing, you need to lift your arm, which gives mgh ≈ 0.2 kg × 9.8 m/s² × 0.4 m ≈ 0.8 J of mechanical work. Divide by maybe 25% muscle efficiency to get 3.2 J. Plus more for holding your arm up, moving your finger, etc., so the total is almost definitely higher than the “thinking really hard”, which again is probably very much less than 5 J.
Technique: As it happened, I asked Claude to do the first-pass scratching-your-nose calculation. It did a great job!
Yup, sounds right.
I think it can represent multiple possibilities to a nonzero but quite limited extent; I think the superposition can only be kinda local to a particular subregion of the cortex and a fraction of a second. I talk about that a bit in §2.3.
I wrote “your brain can wind up settling on either of [the two generative models]”, not both at once.
…Not sure if I answered your question.