Someone who is interested in learning and doing good.
My Twitter: https://twitter.com/MatthewJBar
My Substack: https://matthewbarnett.substack.com/
Matthew Barnett
Karma: 10,514
Just a quick reply to this:
Is that a testable-prior-to-the-apocalypse prediction? i.e. does your model diverge from mine prior to some point of no return? I suspect not. I’m interested in seeing if we can make some bets on this though; if we can, great; if we can’t, then at least we can avoid future disagreements about who should update.
I’ll note that my prediction was for the next “few years” and the 1-3 OOMs of compute. It seems your timelines are even shorter than I thought if you think the apocalypse, or point of no return, will happen before that point.
With timelines that short, I think betting is overrated. From my perspective, I’d prefer to simply wait and become vindicated as the world does not end in the meantime. However, I acknowledge that simply waiting is not very satisfying from your perspective, as you want to show the world that you’re right before the catastrophe. If you have any suggestions for what we can bet on that would resolve in such a short period of time, I’m happy to hear them.
Yes, rereading the passage, Bostrom’s central example of a reason why we could see this “when dumb, smarter is safer; yet when smart, smarter is more dangerous” pattern (that’s a direct quote btw) is that they could be scheming/pretending when dumb. However [...] Bostrom is explicitly calling out the possibility of an AI being genuinely trying to help you, obey you, or whatever until it crosses some invisible threshold of intelligence and has certain realizations that cause it to start plotting against you. This is exactly what I currently think is plausibly happening with GPT4 etc.
When stated that way, I think what you’re saying is a reasonable point of view, and it’s not one I would normally object to very strongly. I agree it’s “plausible” that GPT-4 is behaving in the way you are describing, and that current safety guarantees might break down at higher levels of intelligence. I would like to distinguish between two points that you (and others) might have interpreted me to be making:
We should now think that AI alignment is completely solved, even in the limit of unlimited intelligence and future agentic systems. I am not claiming this.
We (or at least, many of us) should perform a significant update towards alignment being easier than we thought because of the fact that some traditional problems are on their way towards being solved. <--- I am claiming this
The fact that Bostrom’s central example of a reason to think that “when dumb, smarter is safer; yet when smart, smarter is more dangerous” doesn’t fit for LLMs, seems adequate for demonstrating (2), even if we can’t go as far as demonstrating (1).
It remains plausible to me that alignment will become very difficult above a certain intelligence level. I cannot rule that possibility out: I am only saying that we should reasonably update based on the current evidence regardless, not that we are clearly safe from here and we should scale all the way to radical superintellligence without a worry in the world.
Instruction-tuned LLMs are not powerful general agents. They are pretty general but they are only a tiny bit agentic. They haven’t been trained to pursue long-term goals and when we try to get them to do so they are very bad at it. So they just aren’t the kind of system Bostrom, Yudkowsky, and myself were theorizing about and warning about.
I have two general points to make here:
I agree that current frontier models are only a “tiny bit agentic”. I expect in the next few years they will get significantly more agentic. I currently predict they will remain roughly equally corrigible. I am making this prediction on the basis of my experience with the little bit of agency current LLMs have, and I think we’ve seen enough to know that corrigibility probably won’t be that hard to train into a system that’s only 1-3 OOMs of compute more capable. Do you predict the same thing as me here, or something different?
There’s a bit of a trivial definitional problem here. If it’s easy to create a corrigible, helpful, and useful AI that allows itself to get shut down, one can always say “those aren’t the type of AIs we were worried about”. But, ultimately, if the corrigible AIs that let you shut them down are competitive with the agentic consequentialist AIs, then it’s not clear why we should care? Just create the corrigible AIs. We don’t need to create the things that you were worried about!
Here’s my positive proposal for what I think is happening. [...] General world-knowledge is coming first, and agency later. And this is probably a good thing for technical alignment research, because e.g. it allows mechinterp to get more of a head start, it allows for nifty scalable oversight schemes in which dumber AIs police smarter AIs, it allows for faithful CoT-based strategies, and many more things besides probably. So the world isn’t as grim as it could have been, from a technical alignment perspective.
I think this was a helpful thing to say. To be clear: I am in ~full agreement with the reasons you gave here, regarding why current LLM behavior provides evidence that the “world isn’t as grim as it could have been”. For brevity, and in part due to laziness, I omitted these more concrete mechanisms why I think the current evidence is good news from a technical alignment perspective. But ultimately I agree with the mechanisms you offered, and I’m glad you spelled it out more clearly.
At any rate speaking for myself, I have updated towards hopefulness about the technical alignment problem repeatedly over the past few years, even as I updated towards pessimism about the amount of coordination and safety-research-investment that’ll happen before the end (largely due to my timelines shortening, but also due to observing OpenAI). These updates have left me at p(doom) still north of 50%.
As we have discussed in person, I remain substantially more optimistic about our ability to coordinate in the face of an intelligence explosion (even a potentially quite localized one). That said, I think it would be best to save that discussion for another time.
That’s reasonable. I’ll edit the top comment to make this exact clarification.
My claim was not that current LLMs have a high level of big picture awareness.
Instead, I claim current systems have limited situational awareness, which is not yet human-level, but is definitely above zero. I further claim that solving the shutdown problem for AIs with limited (non-zero) situational awareness gives you evidence about how hard it will be to solve the problem for AIs with more situational awareness.
And I’d predict that, if we design a proper situational awareness benchmark, and (say) GPT-5 or GPT-6 passes with flying colors, it will likely be easy to shut down the system, or delete all its copies, with no resistance-by-default from the system.
And if you think that wouldn’t count as an adequate solution to the problem, then it’s not clear the problem was coherent as written in the first place.
I continue to think that you are misinterpreting the old writings as making predictions that they did not in fact make.
We don’t need to talk about predictions. We can instead talk about whether their proposed problems are on their way towards being solved. For example, we can ask whether the shutdown problem for systems with big picture awareness is being solved, and I think the answer is pretty clearly “Yes”.
(Note that you can trivially claim the problem here isn’t being solved because we haven’t solved the unbounded form of the problem for consequentialist agents, who (perhaps by definition) avoid shutdown by default. But that seems like a red herring: we can just build corrigible agents, rather than consequentialist agents.)
Moreover, I think people generally did not make predictions at all when writing about AI alignment, perhaps because that’s not very common when theorizing about these matters. I’m frustrated about that, because I think if they did make predictions, they would likely have been wrong in roughly the direction I’m pointing at here. That said, I don’t think people should get credit for failing to make any predictions, and as a consequence, failing to get proven wrong.
To the extent their predictions were proven correct, we should give them credit. But to the extent they made no predictions, it’s hard to see why that vindicates them. And regardless of any predictions they may or may not have made, it’s still useful to point out that we seem to be making progress on several problems that people pointed out at the time.
I do not know how much one should be punished for various crimes. I’d imagine that our current policy is too inhumane. But however much one thinks people should be punished for various crimes, it’s hard to fathom why corporal punishment is ruled out but prison is tolerated. Given that prison is the less humane option, either both should be allowed or neither should.
One reason to support prison as punishment for crimes over corporal punishment is that prisons confine and isolate dangerous individuals for lengthy periods, protecting the general public via physical separation.
I’d argue that physically preventing certain violent people from being able to harm others is indeed one of the most important purposes served by criminal law, and it’s not served very well by corporal punishment. Some individuals are simply too impulsive or myopic to be deterred by corporal punishment. Almost the moment you let them free, after their beating, they’d just begin committing crimes again. By contrast, putting them in a high security prison allows society to monitor these people and prevent them from harming others directly.
The death penalty perhaps served this purpose in the past by making violent criminals permanently incapable of harming others ever again, but our society has (probably correctly) largely decided that it is morally wrong to toss away someone’s life merely because they are pathologically dangerous. Therefore, prison serves as a useful compromise when protecting the public from violent criminals who are unable to stop committing repeated offenses.
Thankfully, most people generally age out of crime, so life sentences are rarely necessary, even for those who are generally quite violent.
A treacherous turn can result from a strategic decision to play nice and build strength while weak in order to strike later
LLMs are clearly not playing nice as part of a strategic decision to build strength while weak in order to strike later! Yet, Bostrom imagines that general AIs would do this, and uses it as part of his argument for why we might be lulled into a false sense of security.
This means that current evidence is quite different from what’s portrayed in the story. I claim LLMs are (1) general AIs that (2) are doing what we actually want them to do, rather than pretending to be nice because they don’t yet have a decisive strategic advantage. These facts are crucial, and make a big difference.
I am very familiar with these older arguments. I remember repeating them to people after reading Bostrom’s book, years ago. What we are seeing with LLMs is clearly different than the picture presented in these arguments, in a way that critically affects the conclusion.
I am not claiming that the alignment situation is very clear at this point. I acknowledge that LLMs do not indicate that the problem is completely solved, and we will need to adjust our views as AI gets more capable.
I’m just asking people to acknowledge the evidence in front of their eyes, which (from my perspective) clearly contradicts the picture you’d get from a ton of AI alignment writing from before ~2019. This literature talked extensively about the difficulty of specifying goals in general AI in a way that avoided unintended side effects.
To the extent that LLMs are general AIs that can execute our intended instructions, as we want them to, rather than as part of a deceptive strategy to take over the world, this seems like clear evidence that the problem of building safe general AIs might be easy (and indeed easier than we thought).
Yes, this evidence is not conclusive. It is not zero either.
Me: “Oh ok, that’s a different misunderstanding then. We always believed that getting the AGI to follow our intended instructions, behaviorally, would be easy while the AGI is too weak and dumb to seize power. In fact Bostrom predicted it would get easier to get AIs to do what you want, behaviorally, up until the treacherous turn.”
This would be a valid rebuttal if instruction-tuned LLMs were only pretending to be benevolent as part of a long-term strategy to eventually take over the world, and execute a treacherous turn. Do you think present-day LLMs are doing that? (I don’t)
I claim that LLMs do what we want without seeking power, rather than doing what we want as part of a strategy to seek power. In other words, they do not seem to be following any long-term strategy on the path towards a treacherous turn, unlike the AI that is tested in a sandbox in Bostrom’s story. This seems obvious to me.
Note that Bostrom talks about a scenario in which narrow AI systems get safer over time, lulling people into a false sense of security, but I’m explicitly talking about general AI here. I would not have said this about self-driving cars in 2019, even though those were pretty safe. I think LLMs are different because they’re quite general, in precisely the ways that Bostrom imagined could be dangerous. For example, they seem to understand the idea of an off-switch, and can explain to you verbally what would happen if you shut them off, yet this fact alone does not make them develop an instrumentally convergent drive to preserve their own existence by default, contra Bostrom’s theorizing.
I think instruction-tuned LLMs are basically doing what people thought would be hard for general AIs: they allow you to shut them down by default, they do not pursue long-term goals if we do not specifically train them to do that, and they generally follow our intentions by actually satisfying the goals we set out for them, rather than incidentally as part of their rapacious drive to pursue a mis-specified utility function.
The scenario outlined by Bostrom seems clearly different from the scenario with LLMs, which are actual general systems that do what we want and ~nothing more, rather than doing what we want as part of a strategy to seek power instrumentally. What am I missing here?
In the last year, I’ve had surprisingly many conversations that have looked a bit like this:
Me: “Many people in ~2015 used to say that it would be hard to build an AGI that follows human values. Current instruction-tuned LLMs are essentially weak AGIs that follow human values. We should probably update based on this evidence.”
Interlocutor: “You misunderstood the argument. We never said it would be hard to build an AGI that understands human values. We always said that getting the AGI to care was the hard part.”
Me: “I didn’t misunderstand the argument. I understand the distinction you are making perfectly. I am claiming that LLMs actually execute our intended instructions. I am not saying that LLMs merely understand or predict our intentions. I claim they follow our intended instructions, behaviorally. They actually do what we want, not merely understand what we want.”
Interlocutor: “Again, you misunderstood the argument. We always believed that getting the AGI to care would be the hard part. We never said it would be hard to get an AGI to understand human values.”
[… The conversation then repeats, with both sides repeating the same points...]
[Edited to add: I am not claiming that the alignment is definitely very easy. I acknowledge that LLMs do not indicate that the problem is completely solved, and we will need to adjust our views as AI gets more capable. I understand that solutions that work for GPT-4 may not scale to radical superintelligence. I am talking about whether it’s reasonable to give a significant non-zero update on alignment being easy, rather than whether we should update all the way and declare the problem trivial.]
But “The Value Learning Problem” was one of the seven core papers in which MIRI laid out our first research agenda, so I don’t think “we’re centrally worried about things that are capable enough to understand what we want, but that don’t have the right goals” was in any way hidden or treated as minor back in 2014-2015.
I think you missed my point: my original comment was about whether people are updating on the evidence from instruction-tuned LLMs, which seem to actually act on human values (i.e., our actual intentions) quite well, as opposed to mis-specified versions of our intentions.
I don’t think the Value Learning Problem paper said that it would be easy to make human-level AGI systems act on human values in a behavioral sense, rather than merely understand human values in a passive sense.
I suspect you are probably conflating two separate concepts:
It is easy to create a human-level AGI that can passively learn and understand human values (I am not saying people said this would be difficult in the past)
It is easy to create a human-level AGI that acts on human values, in the sense of actually executing instructions that follow our intentions, rather than following a dangerously mis-specified version of what we asked for.
I do not think the Value Learning Paper asserted that (2) was true. To the extent it asserted that, I would prefer to see quotes that back up that claim explicitly.
Your quote from the paper illustrates that it’s very plausible that people thought (1) was true, but that seems separate to my main point: that people thought (2) was not true. (1) and (2) are separate and distinct concepts. And my comment was about (2), not (1).
There is simply a distinction between a machine that actually acts on and executes your intended commands, and a machine that merely understands your intended commands, but does not necessarily act on them as you intend. I am talking about the former, not the latter.
From the paper,
The novelty here is not that programs can exhibit incorrect or counter-intuitive behavior, but that software agents smart enough to understand natural language may still base their decisions on misrepresentations of their programmers’ intent.
Indeed, and GPT-4 does not base its decisions on a misrepresentation of its programmers intentions, most of the time. It generally both correctly understands our intentions, and more importantly, actually acts on them!
You’ve made detailed predictions about what you expect in the next several years, on numerous occasions, and made several good-faith attempts to elucidate your models of AI concretely. There are many ways we disagree, and many ways I could characterize your views, but “unfalsifiable” is not a label I would tend to use for your opinions on AI. I do not mentally lump you together with MIRI in any strong sense.
For what it’s worth, while my credence in human extinction from AI in the 21st century is 10-20%, I think the chance of human extinction in the next 5 years is much lower. I’d put that at around 1%. The main way I think AI could cause human extinction is by just generally accelerating technology and making the world a scarier and more dangerous place to live. I don’t really buy the model in which an AI will soon foom until it becomes a ~god.
I’m confused about why your <20% isn’t sufficient for you to want to shut down AI research. Is it because of benefits outweigh the risk, or because we’ll gain evidence about potential danger and can shut down later if necessary?
I think the expected benefits outweigh the risks, given that I care about the existing generation of humans (to a large, though not overwhelming degree). The expected benefits here likely include (in my opinion) a large reduction in global mortality, a very large increase in the quality of life, a huge expansion in material well-being, and more generally a larger and more vibrant world earlier in time. Without AGI, I think most existing people would probably die and get replaced by the next generation of humans, in a relatively much poor world (compared to the alternative).
I also think the absolute level risk from AI barely decreases if we globally pause. My best guess is that pausing would mainly just delay adoption without significantly impacting safety. Under my model of AI, the primary risks are long-term, and will happen substantially after humans have already gradually “handed control” over to the AIs and retired their labor on a large scale. Most of these problems—such as cultural drift and evolution—do not seem to be the type of issue that can be satisfactorily solved in advance, prior to a pause (especially by working out a mathematical theory of AI, or something like that).
On the level of analogy, I think of AI development as more similar to “handing off control to our children” than “developing a technology that disempowers all humans at a discrete moment in time”. In general, I think the transition period to AI will be more diffuse and incremental than MIRI seems to imagine, and there won’t be a sharp distinction between “human values” and “AI values” either during, or after the period.
(I also think AIs will probably be conscious in a way that’s morally important, in case that matters to you.)
In fact, I think it’s quite plausible the absolute level of AI risk would increase under a global pause, rather than going down, given the high level of centralization of power required to achieve a global pause, and the perverse institutions and cultural values that would likely arise under such a regime of strict controls. As a result, even if I weren’t concerned at all about the current generation of humans, and their welfare, I’d still be pretty hesitant to push pause on the entire technology.
(I think of technology as itself being pretty risky, but worth it. To me, pushing pause on AI is like pushing pause on technology itself, in the sense that they’re both generically risky yet simultaneously seem great on average. Yes, there are dangers ahead. But I think we can be careful and cautious without completely ripping up all the value for ourselves.)
Chemists would give an example of chemical reactions, where final thermodynamically stable states are easy to predict, while unstable intermediate states are very hard to even observe.
I agree there are examples where predicting the end state is easier to predict than the intermediate states. Here, it’s because we have strong empirical and theoretical reasons to think that chemicals will settle into some equilibrium after a reaction. With AGI, I have yet to see a compelling argument for why we should expect a specific easy-to-predict equilibrium state after it’s developed, which somehow depends very little on how the technology is developed.
It’s also important to note that, even if we know that there will be an equilibrium state after AGI, more evidence is generally needed to establish that the end equilibrium state will specifically be one in which all humans die.
And why don’t you accept classic MIRI example that even if it’s impossible for human to predict moves of Stockfish 16, you can be certain that Stockfish will win?
I don’t accept this argument as a good reason to think doom is highly predictable partly because I think the argument is dramatically underspecified without shoehorning in assumptions about what AGI will look like to make the argument more comprehensible. I generally classify arguments like this under the category of “analogies that are hard to interpret because the assumptions are so unclear”.
To help explain my frustration at the argument’s ambiguity, I’ll just give a small yet certainly non-exhaustive set of questions I have about this argument:
Are we imagining that creating an AGI implies that we play a zero-sum game against it? Why?
Why is it a simple human vs. AGI game anyway? Does that mean we’re lumping together all the humans into a single agent, and all the AGIs into another agent, and then they play off against each other like a chess match? What is the justification for believing the battle will be binary like this?
Are we assuming the AGI wants to win? Maybe it’s not an agent at all. Or maybe it’s an agent but not the type of agent that wants this particular type of outcome.
What does “win” mean in the general case here? Does it mean the AGI merely gets more resources than us, or does it mean the AGI kills everyone? These seem like different yet legitimate ways that one can “win” in life, with dramatically different implications for the losing parties.
There’s a lot more I can say here, but the basic point I want to make is that once you start fleshing this argument out, and giving it details, I think it starts to look a lot weaker than the general heuristic that Stockfish 16 will reliably beat humans in chess, even if we can’t predict its exact moves.
There’s a pretty big difference between statements like “superintelligence is physically possible”, “superintelligence could be dangerous” and statements like “doom is >80% likely in the 21st century unless we globally pause”. I agree with (and am not objecting to) the former claims, but I don’t agree with the latter claim.
I also agree that it’s sometimes true that endpoints are easier to predict than intermediate points. I haven’t seen Eliezer give a reasonable defense of this thesis as it applies to his doom model. If all he means here is that superintelligence is possible, it will one day be developed, and we should be cautious when developing it, then I don’t disagree. But I think he’s saying a lot more than that.
I think it’s more similar to saying that the climate in 2040 is less predictable than the climate in 2100, or saying that the weather 3 days from now is less predictable than the weather 10 days from now, which are both not true. By contrast, the weather vs. climate distinction is more of a difference between predicting point estimates vs. predicting averages.
I unfortunately am busy right now but would love to give a fuller response someday, especially if you are genuinely interested to hear what I have to say (which I doubt, given your attitude towards MIRI).
I’m a bit surprised you suspect I wouldn’t be interested in hearing what you have to say?
I think the amount of time I’ve spent engaging with MIRI perspectives over the years provides strong evidence that I’m interested in hearing opposing perspectives on this issue. I’d guess I’ve engaged with MIRI perspectives vastly more than almost everyone on Earth who explicitly disagrees with them as strongly as I do (although obviously some people like Paul Christiano and other AI safety researchers have engaged with them even more than me).
(I might not reply to you, but that’s definitely not because I wouldn’t be interested in what you have to say. I read virtually every comment-reply to me carefully, even if I don’t end up replying.)
I appreciate the straightforward and honest nature of this communication strategy, in the sense of “telling it like it is” and not hiding behind obscure or vague language. In that same spirit, I’ll provide my brief, yet similarly straightforward reaction to this announcement:
I think MIRI is incorrect in their assessment of the likelihood of human extinction from AI. As per their messaging, several people at MIRI seem to believe that doom is >80% likely in the 21st century (conditional on no global pause) whereas I think it’s more like <20%.
MIRI’s arguments for doom are often difficult to pin down, given the informal nature of their arguments, and in part due to their heavy reliance on analogies, metaphors, and vague supporting claims instead of concrete empirically verifiable models. Consequently, I find it challenging to respond to MIRI’s arguments precisely. The fact that they want to essentially shut down the field of AI based on these largely informal arguments seems premature to me.
MIRI researchers rarely provide any novel predictions about what will happen before AI doom, making their theories of doom appear unfalsifiable. This frustrates me. Given a low prior probability of doom as apparent from the empirical track record of technological progress, I think we should generally be skeptical of purely theoretical arguments for doom, especially if they are vague and make no novel, verifiable predictions prior to doom.
Separately from the previous two points, MIRI’s current most prominent arguments for doom seem very weak to me. Their broad model of doom appears to be something like the following (although they would almost certainly object to the minutia of how I have written it here):
(1) At some point in the future, a powerful AGI will be created. This AGI will be qualitatively distinct from previous, more narrow AIs. Unlike concepts such as “the economy”, “GPT-4″, or “Microsoft”, this AGI is not a mere collection of entities or tools integrated into broader society that can automate labor, share knowledge, and collaborate on a wide scale. This AGI is instead conceived of as a unified and coherent decision agent, with its own long-term values that it acquired during training. As a result, it can do things like lie about all of its fundamental values and conjure up plans of world domination, by itself, without any risk of this information being exposed to the wider world.
(2) This AGI, via some process such as recursive self-improvement, will rapidly “foom” until it becomes essentially an immortal god, at which point it will be able to do almost anything physically attainable, including taking over the world at almost no cost or risk to itself. While recursive self-improvement is the easiest mechanism to imagine here, it is not the only way this could happen.
(3) The long-term values of this AGI will bear almost no relation to the values that we tried to instill through explicit training, because of difficulties in inner alignment (i.e., a specific version of the general phenomenon of models failing to generalize correctly from training data). This implies that the AGI will care almost literally 0% about the welfare of humans (despite potentially being initially trained from the ground up on human data, and carefully inspected and tested by humans for signs of misalignment, in diverse situations and environments). Instead, this AGI will pursue a completely meaningless goal until the heat death of the universe.
(4) Therefore, the AGI will kill literally everyone after fooming and taking over the world.It is difficult to explain in a brief comment why I think the argument just given is very weak. Instead of going into the various subclaims here in detail, for now I want to simply say, “If your model of reality has the power to make these sweeping claims with high confidence, then you should almost certainly be able to use your model of reality to make novel predictions about the state of the world prior to AI doom that would help others determine if your model is correct.”
The fact that MIRI has yet to produce (to my knowledge) any major empirically validated predictions or important practical insights into the nature of AI, or AI progress, in the last 20 years, undermines the idea that they have the type of special insight into AI that would allow them to express high confidence in a doom model like the one outlined in (4).Eliezer’s response to claims about unfalsifiability, namely that “predicting endpoints is easier than predicting intermediate points”, seems like a cop-out to me, since this would seem to reverse the usual pattern in forecasting and prediction, without good reason.
Since I think AI will most likely be a very good thing for currently existing people, I am much more hesitant to “shut everything down” compared to MIRI. I perceive MIRI researchers as broadly well-intentioned, thoughtful, yet ultimately fundamentally wrong in their worldview on the central questions that they research, and therefore likely to do harm to the world. This admittedly makes me sad to think about.
Sure. Here’s a snippet of Nick Bostrom’s description of the value-loading problem (chapter 13 in his book Superintelligence):
Here’s my interpretation of the above passage:
We need to solve the problem of programming a seed AI with the correct values.
This problem seems difficult because of the fact that human goal representations are complex and not easily represented in computer code.
Directly programming a representation of our values may be futile, since our goals are complex and multidimensional.
We cannot postpone solving the problem until after the AI has developed enough reason to easily understand our intentions, as otherwise that would be too late.
Given that he’s talking about installing values into a seed AI, he is clearly imagining some difficulties with installing values into AGI that isn’t yet superintelligent (it seems likely that if he thought the problem was trivial for human-level systems, he would have made this point more explicit). While GPT-4 is not a seed AI (I think that term should be retired), I think it has reached a sufficient level of generality and intelligence such that its alignment properties provide evidence about the difficulty of aligning a hypothetical seed AI.
Moreover, he explicitly says that we cannot postpone solving this problem “until the AI has developed enough reason to easily understand our intentions” because “a generic system will resist attempts to alter its final values”. I think this looks basically false. GPT-4 seems like a “generic system” that essentially “understands our intentions”, and yet it is not resisting attempts to alter its final goals in any way that we can detect. Instead, it seems to actually do what we want, and not merely because of an instrumentally convergent drive to not get shut down.
So, in other words:
Bostrom talked about how it would be hard to align a seed AI, implicitly focusing at least some of his discussion on systems that were below superintelligence. I think the alignment of instruction-tuned LLMs present significant evidence about the difficulty of aligning systems below the level of superintelligence.
A specific reason cited for why aligning a seed AI was hard was because human goal representations are complex and difficult to specify explicitly in computer code. But this fact does not appear to be big obstacle for aligning weak AGI systems like GPT-4, and instruction-tuned LLMs more generally. Instead, these systems are generally able to satisfy your intended request, as you wanted them to, despite the fact that our intentions are often complex and difficult to represent in computer code. These systems do not merely understand what we want, they also literally do what we want.
Bostrom was wrong to say that we can’t postpone solving this problem until after systems can understand our intentions. We already postponed that long, and we now have systems that can understand our intentions. Yet these systems do not appear to have the instrumentally convergent self-preservation instincts that Bostrom predicted would manifest in “generic systems”. In other words, we got systems that can understand our intentions before the systems started posing genuine risks, despite Bostrom’s warning.
In light of all this, I think it’s reasonable to update towards thinking that the overall problem is significantly easier than one might have thought, if they took Bostrom’s argument here very seriously.