Max H

Karma: 2,432

Most of my posts and comments are about AI and alignment. Posts I’m most proud of, which also provide a good introduction to my worldview:

I also created Forum Karma, and wrote a longer self-introduction here.

PMs and private feedback are always welcome.

NOTE: I am not Max Harms, author of Crystal Society. I’d prefer for now that my LW postings not be attached to my full name when people Google me for other reasons, but you can PM me here or on Discord (m4xed) if you want to know who I am.

Max H Apr 28, 2025, 11:24 PM
4 points
0
in reply to: Wei Dai’s comment on: Wei Dai’s Shortform
I kind of doubt that leaders at big labs would self-identify as being motivated by anything like Eliezer’s notion of heroic responsibility. If any do self-identify that way though, they’re either doing it wrong or misunderstanding. Eliezer has written tons of stuff about the need to respect deontology and also think about all of the actual consequences of your actions, even (especially when) the stakes are high:
The critical question here is: what happens if the plot successfully places the two of them in an epistemic Cooperation-Defection Dilemma, where rather than the two of them just having different goals, Carissa believes that he is mistaken about what happens...

In this case, Carissa could end up believing that to play ‘Defect’ against him would be to serve even his own goals, better than her Cooperating would serve them. Betraying him might seem like a friendly act, an act of aid.
(https://glowfic.com/replies/1874768#reply-1874768)
If he commits to a drastic action he will estimate that actual victory lies at the end of it, and his desperation and sacrifice will not have figured into that estimation process as positive factors. His deontology is not for sale at the price point of failure.
(https://glowfic.com/replies/1940939#reply-1940939)
Starting an AI lab in order to join a doomed race to superintelligence, and then engaging in a bunch of mundane squabbles for corporate control, seems like exactly the opposite of the sentiment here:
For Albus Dumbledore, as for her, the rule in extremis was to decide what was the right thing to do, and do it no matter the cost to yourself. Even if it meant breaking your bounds, or changing your role, or letting go of your picture of yourself. That was the last resort of Gryffindor.
(https://hpmor.com/chapter/93)
Also, re this example:
It also seemingly justifies or obligates Sam Altman to fight back when the OpenAI board tried to fire him, if he believed the board was interfering with his mission.

In general, it seems perfectly fine and normal for a founder-CEO to fight back against a board ouster—no need to bring heroic responsibility into it. Of course, all parties including the CEO and the board should stick to legal / above-board / ethical means of “fighting back”, but if there’s a genuine disagreement between the board and the CEO on how to best serve shareholder interests (or humanity’s interests, for a non-profit), why wouldn’t both sides vigorously defend their own positions and power?

Perhaps the intended reading of your example is that heroic responsibility would obligate or justify underhanded tactics to win control, when the dispute has existential consequences. But I think that’s a misunderstanding of the actual concept. Ordinary self-confidence and agency obligate you to defend your own interests / beliefs / power, and heroic responsibility says that you’re obligated to win without stepping outside the bounds of deontology or slipping into invalid / motivated reasoning.

Max H Apr 6, 2025, 4:36 PM
12 points
−2
on: Max H’s Shortform
Maybe the recent tariff blowup is actually just a misunderstanding due to bad terminology, and all we need to do is popularize some better terms or definitions. We’re pretty good at that around here, right?
Here’s my proposal: flip the definitions of “trade surplus” and “trade deficit.” This might cause a bit of confusion at first, and a lot of existing textbooks will need updating, but I believe these new definitions capture economic reality more accurately, and will promote clearer thinking and maybe even better policy from certain influential decision-makers, once widely adopted.
New definitions:
- Trade surplus: Country A has a bilateral “trade surplus” with Country B if Country A imports more tangible goods (cars, steel, electronics, etc.) from Country B than it exports back. In other words, Country A ends up with more real, physical items. Country B, meanwhile, ends up with more than it started with of something much less important: fiat currency (flimsy paper money) or 1s and 0s in a digital ledger (probably not even on a blockchain!).
  If you extrapolate this indefinitely in a vacuum, Country A eventually accumulates all of Country B’s tangible goods, while Country B is left with a big pile of paper. Sounds like a pretty sweet deal for Country A if you ask me.
  It’s OK if not everyone follows this explanation or believes it—they can tell it’s the good one because it has “surplus” in the name. Surely everyone wants a surplus.
- Trade deficit: Conversely, Country A has a “trade deficit” if it exports more tangible resources than it imports, and thus ends up with less goods on net. In return, it only receives worthless fiat currency from some country trying to hoard actual stuff for their own people. Terrible deal!
  Again, if you don’t totally follow, that’s OK, just pay attention to the word “deficit”. Everyone knows that deficits are bad and should be avoided.
Under the new definitions, it becomes clear that merely returning to the previous status quo of a few days ago, where the US only “wins” the trade war by several hundred billion dollars, is insufficient for the truly ambitious statesman. Instead, the US government should aggressively mint more fiat currency in order to purchase foreign goods, magnifying our trade surplus and ensuring that in the long run the United States becomes the owner of all tangible global wealth.
Addressing second order concerns: if we’re worried about a collapse in our ability to manufacture key strategic goods at home during a crisis, we can set aside part of the resulting increased surplus to subsidize domestic production in those areas. Some of the extra goods we’re suddenly importing will probably be pretty useful in getting some new factories of our own off the ground. (But of course we shouldn’t turn around and export any of that domestic production to other countries! That would only deplete our trade surplus.)

Max H Mar 4, 2025, 6:35 AM
9 points
−3
on: Self-fulfilling misalignment data might be poisoning our AI models
Describing misaligned AIs as evil feels slightly off. Even “bad goals” makes me think there’s a missing mood somewhere. Separately, describing other peoples’ writing about misalignment this way is kind of straw.
Current AIs mostly can’t take any non-fake responsibility for their actions, even if they’re smart enough to understand them. An AI advising someone to e.g. hire a hitman to kill their husband is a bad outcome if there’s a real depressed person and a real husband who are actually harmed. An AI system would be responsible (descriptively / causally, not normatively) for that harm to the degree that it acts spontaneously and against its human deployers’ wishes, in a way that is differentially dependent on its actual circumstances (e.g. being monitored / in a lab vs. not).
Unlike current AIs, powerful, autonomous, situationally-aware AI could cause harm for strategic reasons or as a side effect of executing large-scale, transformative plans that are indifferent (rather than specifically opposed) to human flourishing. A misaligned AI that wipes out humanity in order to avoid shutdown is a tragedy, but unless the AI is specifically spiteful or punitive in how it goes about that, it seems kind of unfair to call the AI itself evil.

Max H Feb 25, 2025, 12:35 AM
4 points
0
on: Export Surplusses
The original tweets seem at least partially tongue-in-cheek? Trade has lots of benefits that don’t depend on the net balance. If Country A buys $10B of goods from Country B and sells $9B of other goods to country B, that is $19B of positive-sum transactions between individual entities in each country, presumably with all sorts of positive externalities and implications about your economy.
The fact that the net flow is $1B in one direction or the other just doesn’t matter too much. Having a large trade surplus (or large trade deficit) is a proxy for generally doing lots of trading and industry, which will tend to correlate with a lot of other things that made or will make you wealthy. But it would be weird if a country could get rich solely by running a trade surplus, while somehow avoiding reaping any of the other usual benefits of trading. “Paying other countries to discern your peoples’ ability to produce” is plausibly a benefit that you get from a trade surplus even if you try hard to avoid all the others, though.

Max H Feb 6, 2025, 1:45 AM
52 points
18
on: Wired on: “DOGE personnel with admin access to Federal Payment System”
My guess is that the IT and computer security concerns are somewhat exaggerated and probably not actually that big of a deal, nor are they likely to cause any significant or lasting damage on their own. At the very least, I wouldn’t put much stock in what a random anonymous IT person says, especially when those words are filtered through and cherry-picked by a journalist.
These are almost certainly sprawling legacy systems, not a modern enterprise cloud where you can simply have a duly authorized superadmin grant a time-limited ReadOnly IAM permission to * or whatever, along with centralized audit logging and sophisticated change management. There are probably more old-school / manual processes in place, which require going through layers of humans who aren’t inclined to be cooperative or speedy, especially at this particular moment. I think Elon (and Trump) have some justified skepticism of those processes and the people who implemented them.
Still, there’s going to be some kind of audit logging + technical change management controls, and I kind of doubt that any of Elon’s people are going to deliberately sidestep or hide from those, even if they don’t follow all the on-paper procedures and slash some red tape.
And ultimately, even sophisticated technical controls are not a substitute for actual legal authority, which they (apparently / perhaps questionably) have. I’ll be much more concerned if they start violating court orders, even temporarily. e.g. I think it would be very bad (and more plausible than IT malfeasance or negligence) if they are ordered to stop doing whatever by a lower court, but they don’t actually stop, or slow-walk on reversing everything, because they expect to win on appeal to the Supreme Court (even if they’re correct about their appeal prospects). IDK about Elon’s people specifically, but I think ignoring court orders (especially lower courts and temporary injunctions) is a more dangerous form of institutional decay that Trump is likely to usher in, especially since the legislature seems unlikely to offer any real push-back / rebuke.

Max H Feb 5, 2025, 1:04 AM
13 points
5
in reply to: evhub’s comment on: evhub’s Shortform
It seems more elegant (and perhaps less fraught) to have the reference class determination itself be a first class part of the regular CEV process.
For example, start with a rough set of ~all alive humans above a certain development threshold at a particular future moment, and then let the set contract or expand according to their extrapolated volition. Perhaps the set or process they arrive at will be like the one you describe, perhaps not. But I suspect the answer to questions about how much to weight the preferences (or extrapolated CEVs) of distant ancestors and / or “edge cases” like the ones you describe in (b) and (c) wouldn’t be affected too much by the exact starting conditions either way.
Re: the point about hackability and tyranny, humans already have plenty of mundane / naturalistic reasons to seek power / influence / spread of their own particular current values, absent any consideration about manipulating a reference class for a future CEV. Pushing more of the CEV process into the actual CEV itself minimizes the amount of further incentive to do these things specifically for CEV reasons. Whereas, if a particular powerful person or faction doesn’t like your proposed lock-in procedure, they now have (more of) an incentive to take power beforehand to manipulate or change it.

Max H Dec 26, 2024, 7:03 PM
43 points
0
on: How to have Polygenically Screened Children
My wife completed two cycles of IVF this year, and we had the sequence data from the preimplantation genetic testing on the resulting embryos analyzed for polygenic factors by the unnamed startup mentioned in this post.
I can personally confirm that the practical advice in this post is generally excellent.
The basic IVF + testing process is pretty straightforward (if expensive), but navigating the medical bureaucracy can be a hassle once you want to do anything unusual (like using a non-default PGT provider), and many clinics aren’t going to help you with anything to do with polygenic screening, even if they are open to it in principle. So knowing exactly what you want and what you need to ask for is key.
Since this post was written, there have been lots of other developments and related posts in this general area:
And probably many others I am forgetting. But if you’re a prospective parent looking for practical advice on how to navigate the IVF process and take advantage of the latest in genetic screening technology, this post is still the best place to start that I know of. Some of the things in the list above are more speculative, but the technology for selection is basically ready and practical now, and the effect size doesn’t have to be very large for it to beat the status quo of having an embryologist eyeball it.
I think this post is a slam dunk for a +9 and a spot in the LW canon, both for its object-level information and its exemplary embodiment of the virtue of empiricism and instrumental rationality. The rest of this review details my own experience with IVF in the U.S. in 2024.
This section of the original post basically covers it, but to recap, the two main things you’ll want to ask your prospective IVF clinic are:
- Can we use Orchid Labs or Genomic Prediction for PGT?
- Can we implant any healthy embryo of our choosing? (some clinics can have policies against sex selection, etc.)
In my experience, the best time to ask these questions is in-person at your initial consultation; it can be hard to get answers over the phone / in email before you’re at least a prospective patient, since they generally require a doctor or NP to answer.
The good news is, if you get affirmative answers to these questions, you mostly don’t need to worry about whether the clinic or your embryologist is skeptical or even outright hostile to polygenic screening, because you can simply request your sequence data from your PGT provider directly and have it analyzed on your own.
Note: Genomic Prediction offers their own polygenic screening test (LifeView), but if you’re planning to have a third party analyze the sequence data for non-disease traits, you don’t need this. You can just have your IVF clinic order PGT-A tests from GP, and then request your raw sequence data from GP directly once you get the PGT-A results. AFAIK the actual sequencing that GP does is the same regardless of what test you order from them, and they’re happy to share the raw data with you if you ask.
Another thing you’ll want to confirm is whether you can store any aneuploid embryo(s) long-term. Aneuploid embryos are typically considered not viable and most clinics won’t even try to implant them. But they’re worth keeping frozen speculatively, in case future technology allows them to be repaired, etc. Some clinics will have a policy of automatically discarding confirmed-aneuploid embryos unless you make specific arrangements to do something else with them. Usually this will be a question in a big packet of paperwork you’ll have to fill out about what you want to do with the embryos / eggs in various scenarios, e.g. death / divorce / time limit etc. so just make sure to read carefully.
On selecting a good-quality IVF clinic: the live birth metrics in this post are a good starting point, but probably confounded somewhat by the population the clinic serves, and realistically the biggest factor in determining how many embryos you get is going to be your personal health factors and age. My wife is over 30 and in pretty good shape and took a bunch of vitamins before / during the cycles (B12, Omega-3, Myo-Inositol, a prenatal vitamin) and we ended up with 21 eggs retrieved across two cycles, which is right around the expected number for her age.
These attrited down to 10 mature embryos during the fertilization process, 5 of which were screened out as aneuploid via ordinary PGT-A. We had the remaining 5 embryos polygenically screened.
We’re planning to start implanting next year, so I can’t speak to that part of the process personally yet, but overall we’re very happy with the results so far. There was a clear “winner” among the embryos we screened that will be our first choice for implantation, but it’s nice to have all the data we can on all the embryos, and depending on how things go we may end up using more than one of them down the line.
The polygenic screening wasn’t cheap, and given the number of embryos we had to select from, the maximum possible benefit is pretty mild (2.3 bits of selection if we only use one of the 5 and it successfully implants). But given the hassle and expense of IVF itself (not to mention pregnancy and raising a child...) it seems overwhelmingly worth it. We were considering IVF for fertility preservation reasons anyway, so the main question for us was the marginal cost of the extra screening on top.
I’d like to write a longer post with my own takes about having / raising kids on the eve of AI, but for now I’ll just say a few things:
- The choice about whether to have kids is always a personal one, but right now seems like as good a time as any, historically speaking, and takes like these seem crazy wrong to me. Even if you’re thinking purely in terms of the advantage you can give your kids (probably not a good idea to think this way), by far the biggest advantage is being marginally earlier. If you’re interested in having kids but worried about the costs or effectiveness of current IVF / screening methods, consider just having kids the old-fashioned way instead of delaying.
- I have short timelines and expect the default outcome from ASI being developed is swift human extinction, but I still think it’s worth having kids now, at least for me personally. My wife and I had happy childhoods and would have enjoyed being alive even it was only for a short time, and hopefully that’s at least a somewhat heritable trait. And regardless of the endpoint, I expect things to be pretty OK (even great) at least for me and my family, right up until the end, whenever that is. Despite being a “doomer”, I am long the market, and expect to be pretty well-off even if those particular bets don’t pay off and the larger world is somewhat chaotic in the short term.
- 2.3 bits of selection on a single kid realistically isn’t going to make a difference in any kind of “Manhattan project for human intelligence enhancement” and that’s not why we did it. But my sense from having gone through this process is that the barriers to some of the things that Tsvi describes here are more social and financial and scale than technical.

Max H Oct 18, 2024, 1:58 AM
2 points
1
in reply to: Matthew Barnett’s comment on: The Hidden Complexity of Wishes
My main point was that I thought recent progress in LLMs had demonstrated progress at the problem of building such a function, and solving the value identification problem, and that this progress goes beyond the problem of getting an AI to understand or predict human values.
I want to push back on this a bit. I suspect that “demonstrated progress” is doing a lot of work here, and smuggling an assumption that current trends with LLMs will continue and can be extrapolated straightforwardly.
It’s true that LLMs have some nice properties for encapsulating fuzzy and complex concepts like human values, but I wouldn’t actually want to use any current LLMs as a referent or in a rating system like the one you propose, for obvious reasons.
Maybe future LLMs will retain all the nice properties of current LLMs while also solving various issues with jailbreaking, hallucination, robustness, reasoning about edge cases, etc. but declaring victory already (even on a particular and narrow point about value identification) seems premature to me.
Separately, I think some of the nice properties you list don’t actually buy you that much in practice, even if LLM progress does continue straightforwardly.
A lot of the properties you list follow from the fact that LLMs are pure functions of their input (at least with a temperature of 0).
Functional purity is a very nice property, and traditional software that encapsulates complex logic in pure functions is often easier to reason about, debug, and formally verify vs. software that uses lots of global mutable state and / or interacts with the outside world through a complex I/O interface. But when the function in question is 100s of GB of opaque floats, I think it’s a bit of a stretch to call it transparent and legible just because it can be evaluated outside of the IO monad.
Aside from purity, I don’t think your point about an LLM being a “particular function” that can be “hooked up to the AI directly” is doing much work - input() (i.e. asking actual humans) seems just as direct and particular as llm(). If you want your AI system to actually do something in the messy real world, you have to break down the nice theoretical boundary and guarantees you get from functional purity somewhere.
More concretely, given your proposed rating system, simply replace any LLM calls with a call that just asks actual humans to rate a world state given some description, and it seems like you get something that is at least as legible and transparent (in an informal sense) as the LLM version. The main advantage with using an LLM here is that you could potentially get lots of such ratings cheaply and quickly. Replay-ability, determinism and the relative ease of interpretability vs. doing neuroscience on the human raters are also nice, but none of these properties are very reassuring or helpful if the ratings themselves aren’t all that good. (Also, if you’re doing something with such low sample efficiency that you can’t just use actual humans, you’re probably on the wrong track anyway.)

Max H Oct 8, 2024, 1:51 AM
2 points
−7
in reply to: Raemon’s comment on: “Slow” takeoff is a terrible term for “maybe even faster takeoff, actually”
Economic / atomic takeoff

Max H Oct 8, 2024, 1:49 AM
2 points
0
on: “Slow” takeoff is a terrible term for “maybe even faster takeoff, actually”
For specifically discussing the takeoff models in the original Yudkowsky / Christiano discussion, what about:
Economic vs. atomic takeoff
Economic takeoff because Paul’s model implies rapid and transformative economic growth prior to the point at which AIs can just take over completely. Whereas Eliezer’s model is that rapid economic growth prior to takeover is not particularly necessary—a sufficiently capable AI could act quickly or amass resources while keeping a low profile, such that from the perspective of almost all humanity, takeover is extremely sudden.
Note: “atomic” here doesn’t necessarily mean “nanobots”—the goal of the term is to connote that an AI does something physically transformative, e.g. releasing a super virus, hacking / melting all uncontrolled GPUs, constructing a Dyson sphere, etc. A distinguishing feature of Eliezer’s model is that those kinds of things could happen prior to the underlying AI capabilities that enable them having more widespread economic effects.

IIUC, both Eliezer and Paul agree that you get atomic takeoff of some kind eventually, so one of the main disagreements between Paul and Eliezer could be framed as their answer to the question: “Will economic takeoff precede atomic takeoff?” (Paul says probably yes, Eliezer says maybe.)
Separately, an issue I have with smooth / gradual vs. sharp / abrupt (the current top-voted terms) is that they’ve become a bit overloaded and conflated with a bunch of stuff related to recent AI progress, namely scaling laws and incremental / iterative improvements to chatbots and agents. IMO, these aren’t actually closely related nor particularly suggestive of Christiano-style takeoff—if anything it seems more like the opposite:
- Scaling laws and the current pace of algorithmic improvement imply that labs can continue improving the underlying cognitive abilities of AI systems faster than those systems can actually be deployed into the world to generate useful economic growth. e.g. o1 is already “PhD level” in many domains, but doesn’t seem to be on pace to replace a significant amount of human labor or knowledge work before it is obsoleted by Opus 3.5 or whatever.
- Smooth scaling of underlying cognition doesn’t imply smooth takeoff. Predictable, steady improvements on a benchmark via larger models or more compute don’t tell you which point on the graph you get something economically or technologically transformative.

Max H Jun 19, 2024, 11:15 PM
6 points
0
on: My AI Model Delta Compared To Christiano
I’m curious what you think of Paul’s points (2) and (3) here:
- Eliezer often talks about AI systems that are able to easily build nanotech and overpower humans decisively, and describes a vision of a rapidly unfolding doom from a single failure. This is what would happen if you were magically given an extraordinarily powerful AI and then failed to aligned it, but I think it’s very unlikely what will happen in the real world. By the time we have AI systems that can overpower humans decisively with nanotech, we have other AI systems that will either kill humans in more boring ways or else radically advanced the state of human R&D. More generally, the cinematic universe of Eliezer’s stories of doom doesn’t seem to me like it holds together, and I can’t tell if there is a more realistic picture of AI development under the surface.
- One important factor seems to be that Eliezer often imagines scenarios in which AI systems avoid making major technical contributions, or revealing the extent of their capabilities, because they are lying in wait to cause trouble later. But if we are constantly training AI systems to do things that look impressive, then SGD will be aggressively selecting against any AI systems who don’t do impressive-looking stuff. So by the time we have AI systems who can develop molecular nanotech, we will definitely have had systems that did something slightly-less-impressive-looking.
And specifically to what degree you think future AI systems will make “major technical contributions” that are legible to their human overseers before they’re powerful enough to take over completely.
You write:
I expect that, shortly after AIs are able to autonomously develop, analyze and code numerical algorithms better than humans, there’s going to be some pretty big (like, multiple OOMs) progress in AI algorithmic efficiency (even ignoring a likely shift in ML/AI paradigm once AIs start doing the AI research). That’s the sort of thing which leads to a relatively discontinuous takeoff.
But how likely do you think it is that these OOM jumps happen before vs. after a decisive loss of control?
My own take: I think there will probably be enough selection pressure and sophistication in primarily human-driven R&D processes alone to get to uncontrollable AI. Weak AGIs might speed the process along in various ways, but by the time an AI itself can actually drive the research process autonomously (and possibly make discontinuous progress), the AI will already also be capable of escaping or deceiving its operators pretty easily, and deception / escape seems likely to happen first for instrumental reasons.
But my own view isn’t based on the difficulty of verification vs. generation, and I’m not specifically skeptical of bureaucracies / delegation. Doing bad / fake R&D that your overseers can’t reliably check does seem somewhat easier than doing real / good R&D, but not always, and as a strategy seems like it would usually be dominated by “just escape first and do your own thing”.

Max H Jun 17, 2024, 5:24 AM
4 points
1
in reply to: Matthew Barnett’s comment on: Matthew Barnett’s Shortform
That sounds like a frustrating dynamic. I think hypothetical dialogues like this can be helpful in resolving disagreements or at least identifying cruxes when fleshed out though. As someone who has views that are probably more aligned with your interlocutors, I’ll try articulating my own views in a way that might steer this conversation down a new path. (Points below are intended to spur discussion rather than win an argument, and are somewhat scattered / half-baked.)
My own view is that the behavior of current LLMs is not much evidence either way about the behavior of future, more powerful AI systems, in part because current LLMs aren’t very impressive in a mundane-utility sense.
Current LLMs look to me like they’re just barely capable enough to be useful at all—it’s not that they “actually do what we want”, rather, it’s that they’re just good enough at following simple instructions when placed in the right setup / context (i.e. carefully human-designed chatbot interfaces, hooked up to the right APIs, outputs monitored and used appropriately, etc.) to be somewhat / sometimes useful for a range of relatively simple tasks.
So the absence of more exotic / dangerous failure modes can be explained mostly as a lack of capabilities, and there’s just not that much else to explain or update on once the current capability level is accounted for.
I can sort of imagine possible worlds where current-generation LLMs all stubbornly behave like Sydney Bing, and / or fall into even weirder failure modes that are very resistant to RLHF and the like. But I think it would also be wrong to update much in the other direction in a “stubborn Sydney” world.
Do you mind giving some concrete examples of what you mean by “actually do what we want” that you think are most relevant, and / or what it would have looked like concretely to observe evidence in the other direction?
A somewhat different reason I think current AIs shouldn’t be a big update about future AIs is that current AIs lack the ability to bargain realistically. GPT-4 may behaviorally do what the user or developer wants when placed in the right context, but without the ability to bargain in a real way, I don’t see much reason to treat this observation very differently from the fact that my washing machine does what I want when I press the right buttons. The novelty of GPT-4 vs. a washing machine is in its generality and how it works internally, not the literal sense in which it does what the user and / or developer wants, which is a common feature of pretty much all useful technology.
I can imagine worlds in which the observation of AI system behavior at roughly similar capability levels to the LLMs we actually have would cause me to update differently and particularly towards your views, but in those worlds the AI systems themselves would look very different.
For example, suppose someone built an AI system with ~GPT-4 level verbal intelligence, but as a natural side effect of something in the architecture, training process, or setup (as opposed to deliberate design by the developers), the system also happened to want resources of some kind (energy, hardware, compute cycles, input tokens, etc.) for itself, and could bargain for or be incentivized by those resources in the way that humans and animals can often be incentivized by money or treats.
In the world we’re actually in, you can sometimes get better performance out of GPT-4 at inference time by promising to pay it money or threatening it in various ways, but all of those threats and promises are extremely fake—you couldn’t follow through even if you wanted to, and GPT-4 has no way of perceiving your follow-through or lack thereof anyway. In some ways, GPT-4 is much smarter than a dog or a young child, but you can bargain with dogs and children in very real ways, and if you tried to fake out a dog or a child by pretending to give them a treat without following through, they would quickly notice and learn not to trust you.
(I realize there are some ways in which you could analogize various aspects of real AI training processes to bargaining processes, but I would find optimistic analogies between AI training and human child-rearing more compelling in worlds where AI systems at around GPT-4 level were already possible to bargain with or incentivize realistically at runtime, in ways more directly analogous to how we can directly bargain with natural intelligences of roughly comparable level or lower already.)

Zooming out a bit, “not being able to bargain realistically at runtime” is just one of the ways that LLMs appear to be not like known natural intelligence once you look below surface-level behavior. There’s a minimum level of niceness / humanlikeness / “do what we want” ability that any system necessarily has to have in order to be useful to humans at all, and for tasks that can be formulated as text completion problems, the minimum amount seems to be something like “follows basic instructions, most of the time”. But I have not personally seen a strong argument for why current LLMs have much more than the minimum amount of humanlike-ness / niceness, nor why we should expect future LLMs to have more.

Max H Jun 12, 2024, 10:00 PM
11 points
0
in reply to: Richard_Ngo’s comment on: ricraz’s Shortform
Suppose we think of ourselves as having many different subagents that focus on understanding the world in different ways—e.g. studying different disciplines, using different styles of reasoning, etc. The subagent that thinks about AI from first principles might come to a very strong opinion. But this doesn’t mean that the other subagents should fully defer to it (just as having one very confident expert in a room of humans shouldn’t cause all the other humans to elect them as the dictator). E.g. maybe there’s an economics subagent who will remain skeptical unless the AI arguments can be formulated in ways that are consistent with their knowledge of economics, or the AI subagent can provide evidence that is legible even to those other subagents (e.g. advance predictions).

Do “subagents” in this paragraph refer to different people, or different reasoning modes / perspectives within a single person? (I think it’s the latter, since otherwise they would just be “agents” rather than subagents.)
Either way, I think this is a neat way of modeling disagreement and reasoning processes, but for me it leads to a different conclusion on the object-level question of AI doom.

A big part of why I find Eliezer’s arguments about AI compelling is that they cohere with my own understanding of diverse subjects (economics, biology, engineering, philosophy, etc.) that are not directly related to AI—my subagents for these fields are convinced and in agreement.

Conversely, I find many of the strongest skeptical arguments about AI doom to be unconvincing precisely because they seem overly reliant on a “current-paradigm ML subagent” that their proponents feel should be dominant, or at least more heavily weighted than I think is justified.
That will push P(doom) lower because most frames from most disciplines, and most styles of reasoning, don’t predict doom.
This might be true and useful for getting some kind of initial outside-view estimate, but I think you need some kind of weighting rule to make this work as reasoning strategy even at a meta level. Otherwise, aren’t you vulnerable to other people inventing lots of new frames and disciplines? I think the answer in geometric rationality terms is that some subagents will perform poorly and quickly lose their Nash bargaining resources, and then their contribution to future decision-making / conclusion-making will be down-weighted. But I don’t think the only way for a subagent to “perform” for the purposes of deciding on a weight is by making externally legible advance predictions.

Max H Mar 26, 2024, 8:49 PM
7 points
11
on: Modern Transformers are AGI, and Human-Level
Maybe a better question than “time to AGI” is time to mundanely transformative AGI. I think a lot of people have a model of the near future in which a lot of current knowledge work (and other work) is fully or almost-fully automated, but at least as of right this moment, that hasn’t actually happened yet (despite all the hype).

For example, one of the things current A(G)Is are supposedly strongest at is writing code, but I would still rather hire a (good) junior software developer than rely on currently available AI products for just about any real programming task, and it’s not a particularly close call. I do think there’s a pretty high likelihood that this will change imminently as products like Devin improve and get more widely deployed, but it seems worth noting (and finding a term for) the fact that this kind of automation so far (mostly) hasn’t actually happened yet, aside from certain customer support and copyediting jobs.

I think when someone asks “what is your time to AGI”, they’re usually asking about when you expect either (a) AI to radically transform the economy and potentially usher in a golden age of prosperity and post-scarcity or (b) the world to end.

And maybe I am misremembering history or confused about what you are referring to, but in my mind, the promise of the “AGI community” has always been (implicitly or explicitly) that if you call something “human-level AGI”, it should be able to get you to (a), or at least have a bigger economic and societal impact than currently-deployed AI systems have actually had so far. (Rightly or wrongly, the ballooning stock prices of AI and semiconductor companies seem to be mostly an expectation of earnings and impact from in-development and future products, rather than expected future revenues from wider rollout of any existing products in their current form.)

Max H Mar 5, 2024, 4:26 AM
6 points
3
in reply to: TurnTrout’s comment on: TurnTrout’s shortform feed
I actually agree that a lot of reasoning about e.g. the specific pathways by which neural networks trained via SGD will produce consequentialists with catastrophically misaligned goals is often pretty weak and speculative, including in highly-upvoted posts like Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover.
But to expand on my first comment, when I look around and see any kind of large effect on the world, good or bad (e.g. a moral catastrophe, a successful business, strong optimization around a MacGuffin), I can trace the causality through a path that is invariably well-modeled by applying concepts like expected utility theory (or geometric rationality, if you prefer), consequentialism, deception, Goodharting, maximization, etc. to the humans involved.
I read Humans provide an untapped wealth of evidence about alignment and much of your other writing as disagreeing with the (somewhat vague / general) claim that these concepts are really so fundamental, and that you think wielding them to speculate about future AI systems is privileging the hypothesis or otherwise frequently leads people astray. (Roughly accurate summary of your own views?)
Regardless of how well this describes your actual views or not, I think differing answers to the question of how fundamental this family of concepts is, and what kind of reasoning mistakes people typically make when they apply them to AI, is not really a disagreement about neural networks specifically or even AI generally.

Max H Mar 4, 2024, 11:11 PM
LW: 16 AF: 7
8
AF
in reply to: TurnTrout’s comment on: TurnTrout’s shortform feed

one can just meditate on abstract properties of “advanced systems” and come to good conclusions about unknown results “in the limit of ML training”

I think this is a pretty straw characterization of the opposing viewpoint (or at least my own view), which is that intuitions about advanced AI systems should come from a wide variety of empirical domains and sources, and a focus on current-paradigm ML research is overly narrow.

Research and lessons from fields like game theory, economics, computer security, distributed systems, cognitive psychology, business, history, and more seem highly relevant to questions about what advanced AI systems will look like. I think the original Sequences and much of the best agent foundations research is an attempt to synthesize the lessons from these fields into a somewhat unified (but often informal) theory of the effects that intelligent, autonomous systems have on the world around us, through the lens of rationality, reductionism, empiricism, etc.

And whether or not you think they succeeded at that synthesis at all, humans are still the sole example of systems capable of having truly consequential and valuable effects of any kind. So I think it makes sense for the figure of merit for such theories and worldviews to be based on how well they explain these effects, rather than focusing solely or even mostly on how well they explain relatively narrow results about current ML systems.

Max H Mar 1, 2024, 2:51 AM
17 points
−4
on: The Parable Of The Fallen Pendulum—Part 1
There are a bunch of ways to “win the argument” or just clear up the students’ object-level confusion about mechanics:
- Ask them to predict what happens if the experiment is repeated with the stand held more firmly in place.
- Ask them to work the problems in their textbook, using whatever method or theory they prefer. If they get the wrong answer (according to the answer key) for any of them, that suggests opportunities for further experiments (which the professor should take care to set up more carefully).
- Point out the specific place in the original on-paper calculation where the model of the pendulum system was erroneously over-simplified, and show that using a more precise model results in a calculation that agrees with the experimental results. Note that the location of the error is only in the model (and perhaps the students’ understanding); the words in the textbook describing the theory itself remain fixed.
- Write a rigid body physics simulator which can model the pendulum system in enough detail to accurately simulate the experimental result for both the case that the stand is held in place and the case that it falls over. Reveal that the source code for the simulator uses only the principles of Newtonian mechanics.
- Ask the students to pass the ITT of a more experienced physicist. (e.g. ask a physicist to make up some standard physics problems with an answer key, and then challenge the students to accurately predict the contents of the answer key, regardless of whether the students themselves believe those answers would make good experimental predictions.)
These options require that the students and professor spend some time and effort to clear up the students’ confusion about Newtonian mechanics, which may not be feasible if the lecture is ending soon. But the bigger issue is that clearing up the object-level confusion about physics doesn’t necessarily clear up the more fundamental mistakes the students are making about valid reasoning under uncertainty.
I wrote a post recently on Bayesian updating in real life that the students might be interested in, but in short I would say that their biggest mistake is that they don’t have a detailed enough understanding of their own hypotheses. Having failed to predict the outcome of their own experiment, they have strong evidence that they themselves do not possess an understanding of any theory of physics in enough mechanistic detail to make accurate predictions. However, strong evidence of their own ignorance is not strong evidence that any particular theory which they don’t understand is actually false.
The students should also consider alternatives to the “everyone else throughout history has been rationalizing away problems with Newtonian mechanics” hypothesis. That hypothesis may indeed be one possible valid explanation of the students’ own observations given everything else that they (don’t) know, but are they willing to write down some odds ratios between that hypothesis and some others they can come up with? Some alternative hypotheses they could consider:
- they are mistaken about what the theory of Newtonian mechanics actually says
- they or their professor made a calculation or modelling error
- their professor is somehow trolling them
- they themselves are trolls inside of a fictional thought experiment
They probably won’t think of the last one on their own (unless the rest of the dialogue gets very weird), which just goes to show how often the true hypothesis lies entirely outside of one’s consideration.
(Aside: the last bit of dialog from the students reminds me of the beginner computer programmer whose code isn’t working for some unknown-to-them reason, and quickly concludes that it must be the compiler or operating system that is bugged. In real life, sometimes, it really is the compiler. But it’s usually not, especially if you’re a beginner just getting started with “Hello world”. And even if you’re more experienced, you probably shouldn’t bet on it being the compiler at very large odds, unless you already have a very detailed model of the compiler, the OS, and your own code.)

Max H Feb 28, 2024, 5:16 AM
8 points
5
in reply to: Nora Belrose’s comment on: Counting arguments provide no evidence for AI doom

It’s that they’re biased strongly against scheming, and they’re not going to learn it unless the training data primarily consists of examples of humans scheming against one another, or something.

I’m saying if they’re biased strongly against scheming, that implies they are also biased against usefulness to some degree.

As a concrete example, it is demonstrably much easier to create a fake blood testing company and scam investors and patients for $billions than it is to actually revolutionize blood testing. I claim that there is something like a core of general intelligence required to execute on things like the latter, which necessarily implies possession of most or all of the capabilities needed to pull off the former.

Max H Feb 28, 2024, 2:17 AM
15 points
6
on: Counting arguments provide no evidence for AI doom
Joe also discusses simplicity arguments for scheming, which suppose that schemers may be “simpler” than non-schemers, and therefore more likely to be produced by SGD.
I’m not familiar with the details of Joe’s arguments, but to me the strongest argument from simplicity is not that schemers are simpler than non-schemers, it’s that scheming itself is conceptually simple and instrumentally useful. So any system capable of doing useful and general cognitive work will necessarily have to at least be capable of scheming.
We will address this question in greater detail in a future post. However, we believe that current evidence about inductive biases points against scheming for a variety of reasons. Very briefly:
- Modern deep neural networks are ensembles of shallower networks. Scheming seems to involve chains of if-then reasoning which would be hard to implement in shallow networks.
- Networks have a bias toward low frequency functions— that is, functions whose outputs change little as their inputs change. But scheming requires the AI to change its behavior dramatically (executing a treacherous turn) in response to subtle cues indicating it is not in a sandbox, and could successfully escape.
- There’s no plausible account of inductive biases that does support scheming. The current literature on scheming appears to have been inspired by Paul Christiano’s speculations about malign intelligences in Solomonoff induction, a purely theoretical model of probabilistic reasoning which is provably unrealizable in the real world.^[16] Neural nets look nothing like this.
- In contrast, points of comparison that are more relevant to neural network training, such as isolated brain cortices, don’t scheme. Your linguistic cortex is not “instrumentally pretending to model linguistic data in pursuit of some hidden objective.”
Also, don’t these counterpoints prove too much? If networks trained via SGD can’t learn scheming, why should we expect models trained via SGD to be capable of learning or using any high-level concepts, even desirable ones?
These bullets seem like plausible reasons for why you probably won’t get scheming within a single forward pass of a current-paradigm DL model, but are already inapplicable to the real-world AI systems in which these models are deployed.
LLM-based systems are already capable of long chains of if-then reasoning, and can change their behavior dramatically given a different initial prompt, often in surprising ways.

If the most relevant point of comparison to NN training is an isolated brain cortex, then that’s just saying that NN training will never be useful in isolation, since an isolated brain cortex can’t do much (good or bad) unless it is actually hooked up to a body, or at least the rest of a brain.

Max H Feb 18, 2024, 6:35 PM
11 points
10
in reply to: Shankar Sivarajan’s comment on: Intuition for 1 + 2 + 3 + … = −1/12
My point is that there is a conflict for divergent series though, which is why 1 + 2 + 3 + … = −1/12 is confusing in the first place. People (wrongly) expect the extension of + and = to infinite series to imply stuff about approximations of partial sums and limits even when the series diverges.

My own suggestion for clearing up this confusion is that we should actually use less overloaded / extended notation even for convergent sums, e.g. $(1, \frac{1}{2}, \frac{1}{4}, . . .) \sim 2$ seems just as readable as the usual $l i m \to$ and $+ . . . +$ notation.