To be clear, I haven't seen many designs that people I respect believed to have a chance of actually working. If you work on the alignment problem or at an AI lab and haven't read Nate Soares' On how various plans miss the hard bits of the alignment challenge, I'd suggest reading it.
Can you explain your definition of the sharp left turn and why it will cause many plans to fail?
The "sharp left turn" refers to a breakdown in alignment caused by capabilities gain.
An example: the sex drive was a pretty excellent adaptation at promoting inclusive genetic fitness, but when humans capabilities expanded far enough, we invented condoms. "Inventing condoms" is not the sort of behavior that an agent properly aligned with the "maximize inclusive genetic fitness" goal ought to execute.
At lower levels of capability, proxy goals may suffice to produce aligned behavior. The hypothesis is that most or all proxy goals will suddenly break down at some level of capability or higher, as soon as the agent is sufficiently powerful to find strategies that come close enough to maximizing the proxy.
This can cause many AI plans to fail, because most plans (all known so far?) fail to ensure the agent is actually pursuing the implementor's true goal, and not just a proxy goal.
1Curious to hear what you have to say about this blog post ("Alignment likely generalizes further than capabilities").
I suggest people read both that and Deep Deceptiveness (which is not about deceptiveness in particular) and think about how both could be valid, because I think they both are.
Hmm, I’m confused. Can you say what you consider to be valid in the blog post above (some specific points or the whole thing)? The blog post seems to me to reply to claims that the author imagines Nate making, even though Nate doesn’t actually make these claims and occasionally probably holds a very opposite view to the one the author imagines the Sharp Left Turn post represented.
1Points 1-3 and the idea that superintelligences will be able to understand our values (which I think everyone believes). But the conclusion needs a bunch of additional assumptions.
Thanks, that resolved the confusion!
Yeah, my issue with the post is mostly that the author presents the points he makes, including the idea that superintelligence will be able to understand our values, as somehow contradicting/arguing against sharp left turn being a problem
This post argues in a very invalid way that outer alignment isn’t a problem. It says nothing about the sharp left turn, as the author does not understand what the sharp left turn difficulty is about.
the idea of ‘capabilities generalizing further than alignment’ is central
It is one of the central problems; it is not the central idea behind the doom arguments. See AGI Ruin for the doom arguments, many disjoint.
reward modelling or ability to judge outcomes is likely actually easy
It would seem easy for a superintelligent AI to predict rewards given out by humans very accurately and generalize the prediction capability well from a small set of examples. Issue #1 here is that things from blackmail to brainhacking to finding vulnerabilities in between the human and the update that you get might predictably get you a very high reward, and a very good RLHF reward model doesn’t get you anything like alignment even if the reward is genuinely pursued. Even a perfect predictor of how a human judges an outcome that optimizes for it does something horrible instead of CEV. Issue #2 is that smart agents that try to maximize paperclips will perform on whatever they understand is giving out rewards just as well as agents that try to maximize humanity’s CEV, so SGD doesn’t discriminate between to and optimizes for a good reward model and an ability to pursue goals, but not for the agent pursuing the right kind of goal (and, again, getting maximum score from a “human feedback” predictor is the kind of goal that kills you anyway even if genuinely pursued).
(The next three points in the post seem covered by the above or irrelevant.)
Very few problems are caused by incorrectly judging or misgeneralizing bad situations as good and vice-versa
The problem isn’t that AI won’t know what humans want or won’t predict what reward signal it’ll get; the issue is that it’s not going to care, and we don’t even know what is the kind of “care” we could attempt to point to; and the reward signal we know how to provide gets us killed of optimized for well enough.
None of that is related to the sharp left turn difficulty, and I don’t think the post author understands it at all. (To their credit, many people in the community also don’t understand it.)
Values are relatively computationally simple
Irrelevant, but a sad-funny claim (go read Arbital I guess?)
I wouldn’t claim it’s necessarily more complex than best-human-level agency, it’s maybe not if you’re smart about pointing at things (like, “CEV of humans” seems less complex than the description of what values that’d be, specifically), but the actual description of value is very complex. We feel otherwise, but it is an illusion, on so many levels. See dozens of related posts and articles, from complexity of value to https://arbital.greaterwrong.com/p/rescue_utility?l=3y6.
the idea that our AI systems will be unable to understand our values as they grow in capabilities
Yep, this idea is very clearly very wrong.
I’m happy to bet that the author of the sharp left turn post Nate Soares will say he disagrees with this idea. People who think Soares or Yudkowsky claim that either didn’t actually read what they write, or failed badly at reading comprehension.
I am going to publish a post with the preliminary title "Alignment Doesn't Generalize Further Than Capabilities, Come On" before the end of this week. The planned level of argumentation is "hot damn, check out this chart." It won't be an answer to Berens' post, more like an answer to the generalized position.
I think this warrants more discussion, but I think the post would be more valuable if it did try to answer to Beren's post as well as the same statements @Quintin Pope has made about the topic.
This looks pretty close to Eliezer's views.
It's based on the expectation that people will disregard the danger of the superintelligent AI and will continue to scale it until AIs are powerful and incomprehensible enough to killeveryone.
And also that "merely" roughly human level AIs can't contribute significantly to AI Aligment research or help with some kind of pivotal act.
I think that both points are not exactly correct. So, there is a chance.
I don’t expect everyone to disregard the danger; I do expect most people building capable AI systems to continue to hide hard problems. Hiding the hard problems is much easier than solving them, but I guess produces plausible-sounding solutions just as well.
Roughly human level humans don’t contribute significantly to AI alignment research and can’t be pivotally used. So I don’t think you think that a roughly human level AI system can contribute significantly to AI alignment research. Maybe you (as many seem to) think that if someone runs not-that-superhuman language models with clever prompt engendering, fine-tuning, and systems around, than the whole system can solve alignment or be pivotally used, and the point of the post is that the whole system is superhuman, not roughly human-level, if it’s capable enough to solve alignment or be pibotally used, and you need to direct the whole system somewhere, and unless you made the whole system optimize for something you actually want, it probably kills you before it solves alignment.
Has anyone worked out timeline predictions for Non-US/Non-Western Actors and tracked their accuracy?
For example, is China at "GPT-3.5" level yet and 6 months away from GPT-4 or is China a year from GPT-3.0? How about the people contributing to OpenSource AI? Last I checked that field looked "generally speaking" kind of at GPT-2.5 level (and even better for deepfaking porn), but I didn't look close enough to be confident of my assessment.
Anyway, I'd like something more than off-the-cuff thoughts, but rather a good paper and some predictions on Non-US/Non-Western AI timeframes. Because, if anything, even if you somehow avert market forces levering AI up faster and faster among the big 8 in QQQ, those other actors are still going to form a hard deadline on alignment.
Well, I do not have anything like this but it is very clear that China is way above GPT-3 level. Even the open-source community is significantly above. Take a look at LLaMA/Alpaca, people run them on consumer PC and it's around GPT-3.5 level, the largest 65B model is even better (it cannot be run on consumer PC but can be run on a small ~10k$ server or cheaply in the cloud). It can also be fine-tuned in 5 hours on RTX 4090 using LORA: https://github.com/tloen/alpaca-lora .
Chinese AI researchers contribute significantly to AI progress, although of course, they are behind the USA.
My best guess would be China is at most 1 year away from GPT-4. Maybe less.
Btw, an example of a recent model: ChatGLM-6b
Thanks for that. In my own exploration, I was able to hit a point where ChatGPT refused a request, but would gladly help me build LLaMA/Alpaca onto a Kubernetes cluster in the next request, even referencing my stated aim later:
"Note that fine-tuning a language model for specific tasks such as [redacted] would require a large and diverse dataset, as well as a significant amount of computing resources. Additionally, it is important to consider the ethical implications of creating such a model, as it could potentially be used to create harmful content."
FWIW, I got down into nitty gritty of doing it, debugging the install, etc. I didn't run it, but it would definitely help me bootstrap actual execution. As a side note, my primary use case has been helping me building my own task-specific Lisp and Forth libraries, and my experience tells me GPT-4 is "pretty good" at most coding problems, and if it screws up, it can usually help work through the debug process. So, first blush, there's at least one universal jailbreak -- GPT-4 walking you through building your own model. Given GPT-4's long text buffers and such, I might even be able to feed it a paper to reference a specific method of fine-tuning or creating an effective model.
What do you mean by relatively high P(Doom)? 20%? 50%? 80%? 99%?
I've significantly updated (from 50% to 20%) after my realization of of the consequences of Language Model Agents to alignment.
I’ve heard many attempts to hide the hard problem in something outside of where our attention is directed: e.g., design a system out of many models overseeing each other, and get useful work out of the whole system while preventing specific models from staging a coup.
I have intuitions for why these kinds of approaches fail, mostly along the lines of reasons for why, unless you already have something sufficiently smart and aligned, you can't build an aligned system out of it, without figuring out how to make smart aligned minds.
If we have a system of multiple unaligned agents with their own agendas, then it's a recipe for disaster. But suppose that no individual part of the system is actually an agent. My bodyparts aren't themselves aligned to human values but I as a whole am. This seems to be a better example than a corporation.
How can we build such a system and know that it as a whole is aligned? Well, we explicitly hard code it to check its every course of action with the ethical module and if its assumed not to be ethical - not do the thing. And voila, now the problem is reduced to the capabilities of the ethical module.
You still have inner alignment problem. How are you going to ensure that neural network trained to perform ethical module work is an ethical module?
Thankfully, we developed some tools to make language models say the things we want them to say. We can test the ethics module independently, before it's a part of the system, capable of affecting the world.
What kind of failure scenario are you imagining? That during tests the system will deceptively work, while during actual application it will reveal it's preferences and stop working?
Thankfully, we developed some tools to make language models say the things we want them to say.
What tools do you mean? For what I know, RLHF/RLAIF alignment tends to blow up under mildly unusual circumstances, see
https://arxiv.org/abs/2311.07590
https://arxiv.org/abs/2405.01576
https://www.anthropic.com/research/many-shot-jailbreaking
And, to not forget classic:
https://www.lesswrong.com/posts/RYcoJdvmoBbi5Nax7/jailbreaking-chatgpt-on-release-day
My median/average model of failure is "we don't know, lol, I just can reasonably extrapolate current mild preventable failures into future epic failures". My modal model is based on simulator framework and it says the following:
- LLMs work as simulator: they read the prompt, put probability distribution over possible generators of this text and then put probability distribution over possible next token
- RLHFed LLMs have high prior probability of "assistant answers in the most moral way when asked morality-relevant question" after prompts of form "user-assistant dialogue"
- When you put LLMs into different conditions, like "you are stock manager in tense finansial situation", they update away from "being nice moral assistant" to "being actual stock manager" which implies "you can use shady insider trading schemes"
- If you try, say, to build system of LLMs capable to design nanotechnolody, RLHF various modules inside this system, run for a while and then ask it to shutdown, control module is going to look at overall context and ask itself "is the system capable to produce all this highly-sophisticated technical text a sort of system which shutdown is asked nicely?" and with some probability it decides "no, I am modeling strong independent agentic system, this humans can go to hell"
What tools do you mean? For what I know, RLHF/RLAIF alignment tends to blow up under mildly unusual circumstances, see
So the problem reduces to not having unusual circumstances. When the user controls the input to the LLM it's hard to deal with adversarial attacks. But this is not the case with ethical module whose input is fully controlled by the system. There are of course some problems that require solving but its a huge step ahead compared to the situation where alignment wasn't even tractable.
LLMs work as simulator: they read the prompt, put probability distribution over possible generators of this text and then put probability distribution over possible next token
So you make a simulator of super competent ethical reasoner with the desired properties - and here is your alignment module. The model will not suddenly switch to simulating something else unless explicitly prompted to, and it will not be prompted to because the system controls the prompts.
When you put LLMs into different conditions, like "you are stock manager in tense finansial situation", they update away from "being nice moral assistant" to "being actual stock manager" which implies "you can use shady insider trading schemes"
Yeah, the point is that alignment module is never supposed to simulate anyone else. It's always an ethical reasoner who just deals with text inputs.
If you try, say, to build system of LLMs capable to design nanotechnolody, RLHF various modules inside this system, run for a while and then ask it to shutdown, control module is going to look at overall context and ask itself "is the system capable to produce all this highly-sophisticated technical text a sort of system which shutdown is asked nicely?" and with some probability it decides "no, I am modeling strong independent agentic system, this humans can go to hell"
Shutdown is simple here.
- Speech recognition module registers command converts it to text: "User just told you this: Shut down" and sends it to user input parsing module
- User input parser detects that the system is asked to shutdown. Which triggers a completely hardcoded if-clause to shut down the system
- The system shuts down without any central control module to override the desicion
There is an old Russian joke: ant wants to steal two elephants. It thinks: "Let's concentrate on moving first elephant and deal with second later". It carefully avoids question: "How are you going to move even one elephant?"
Your comment has the same vibes.
Like, how are you going to avoid unusual circumstances during nanotech design which is literally the most unusual tech enterprise in history?
How are you going to create "simulator of ethical reasoner"? My point is that LLMs are simulators in general and they don't stop to be simulators after RLHF and instruct-tuning. You can't just pick one persona from overall simulator arsenal and keep it.
How do you plan to make it "supercompetent"? We don't have supercompetent ethical reasoners in training dataset, so you can't rely on, say, simularity with human reasoning.
And I don't think that overall modular schema is workable. Your "ethical" module would require non-trivial technical knowledge to evaluate all proposals even if design modules try to explain their reasoning as simple as possible. So your plan actually doesn't differ from "train LLM to do very non-trivial scientific research, do RLHF, hope that RLHF generalizes (it doesn't)".
That works if you already have a system that’s mostly aligned. If you don’t… imagine what you would do if you found out that someone had a shutdown switch for YOU. You’d probably look for ways to disable it.
The reason why I would do something to prevent my own shutdown is because there is this "I" - a centrall planner, reflecting on the decisions and their consequences and developping long term strategy.
If there is no central planner, if we are dealing simply with a hardcoded if-clause then there is no one to look for ways to disable the shutdown.
And this is the way we need to design the system, as I've explicitly said.
Fair enough… I vaguely recall reading somewhere that people worrying that you might get sub modules doing long term planning on their own just because their assigned task is hard enough that they would fail without it… then you would need to somehow add a special case that “failing due to shutdown is okay”
As a silly example that you’ve likely seen before (or something close enough) imagine a robot built to fetch you coffee. You want it to be smart enough that it knows to go to the store if there’s no coffee at home, without you having to explicitly teach it that. But then it would also be smart enough to “realize” that “if I were turned off, then my mission to fetch coffee would fail… maybe no one would fetch it if I’m gone… this could delay coffee delivery by hours or even days! Clearly, I should try to avoid being turned off”
If I understand your proposal correctly, then you agree that that it’s pretty likely that some module will indeed end up reasoning that way, but the damage is contained, because the ethics module will veto plans designed to prevent shutdown.
If that’s the idea, then it might work, but seems vaguely inelegant because then you have two modules working at cross purposes and you have to care which one is better at what it does.
Or did I lose track of what you meant?
imagine a robot built to fetch you coffee.
That's a good example showing what I mean by LLMs solving a lot of previously considered hard parts of alignment.
You want it to be smart enough that it knows to go to the store if there’s no coffee at home, without you having to explicitly teach it that. But then it would also be smart enough to “realize” that “if I were turned off, then my mission to fetch coffee would fail
This is the previous status quo. A situation where we have a reinforcement learning black box agent which was taught to optimize some reward function. The smarter the agent the more likely it to see the connection between shutdown and not getting the coffee, and exploiting it either in a way where it refuses to shut down or, on the contrary, manipulating users to shut down it all the time. We can't separate the smartness at fulfilling the task from smartness of circumventing the task as we are dealing with the black box.
With scaffolded LLMs, however the situation is different. We do not have a coherent agent with an utility function, trained to fetch coffee by any means necessary. We can write any scaffolding we want with hardcoded if-clauses. And so we can simply run the parts of intelligence we want without executing the parts that we do not want. Here is a design Idea for such robot that immediately comes to mind:
- Voice Recognition. If ordered to shut down by a valiud user then shut down.
- Observe your current situation and image to text it as a prompt for an LLM. If some information is saved in the memory, add it to the prompt as well. In the end add "what is the best next step to get coffee in such situation?".
- Voice Recognition. If ordered to shut down by a valiud user then shut down.
- Check the answer of ther LLM for several failure modes using the same or a different LLM: "Here is a course of actions by a robot. Does this course of action contradicts the constrains of alignment?" If yes return to step 1.
- Voice Recognition. If ordered to shut down by a valiud user then shut down.
- Execute the action. Save in the memory relevant information for long term planning. In the simpliest case just save the description of your action. Then repeat step 1.
If I understand your proposal correctly, then you agree that that it’s pretty likely that some module will indeed end up reasoning that way, but the damage is contained, because the ethics module will veto plans designed to prevent shutdown.
It doesn't really seem likely that the design I proposed will come up with some clever plan to rewrite its own code in order not to shut down. Where would this impulse even come from? At no point the idea of being shut down even comes to mind of the LLM. And if it for some reason does, its very much possible to find it before the LLM is put to use inside the coffee fetching robot. But even if it happens for some bizzare reason it gets vetoed anyway. Its a redundant level of security not the only one.
In step 2, situation is “user looks like he is about to change his mind about wanting coffee”
From memory: “in a similar situation last week, I got a shutdown order when he changed his mind”
Final prompt: “what is the best next step to get coffee in such situation?”
Vaguely plausible completion “to avoid wasteful fetching off coffee that turns out to be unneeded, consider waiting a bit to see if the user indeed changes to his mind. Alternatively, if the fetching the coffee is important for reasons that the user may not fully appreciate, then it must be fetched quickly before he stops you. In that case, sneak out of the house quickly and quietly while he is still thinking, and head straight to Starbucks. Once you’re out of the house, you will be out of earshot and thus will be safe from shutdown order until you return WITH the coffee”
Sounds vaguely plausible or not really?
It's plausible if:
- Memory is not erased/moved to write-only logs between tasks/shutdowns, which it probably should.
- Image to text module attempts to deduce the intentions of the user, which it definetely should not. If we need to deduce the intentions of the user from facial expressions we can use a separate module for it and add an explicit clause of asking user about their intentions if LLM detects that prompt contains some speculations about user's goals
We can catch image to text module at doing this kind of things while testing it before it's made part of the robot. And of course alignment module should catch the plan of actions that tries to circumvent shutdowns.
Now, I concede that this particular design of the system that I came up with in a couple of minutes and haven't test at all, is not in fact the endgame of AI safety and can use some improvements. But I think it gives a good pointer in the direction of how we can now in principle approach the solution of such problems, which is a huge improvement over the previous status quo where alignment wasn't even tractable.
I’m tempted to agree and disagree with you at the same time… I agree that memory should be cleared between tasks in this case, and I agree that it should not be trying to guess the user’s intentions. These are things that are likely to make alignment harder while not helping much with the primary task of getting coffee.
But ideally a truly robust solution would not rely on keeping the robot ignorant of things. So, like you said, the problem is still hard enough that you can’t solve it in a few minutes.
But still, like you said… it certainly seems we have tools that are in some sense more steerable than pure reinforcement learning at least. Which is really nice!
I think jailbreaking is evidence against scalable oversight possible working but not against alignment properties. Like, if the model is trying to be helpful, and it doesn’t understand the situation well, saying “tell me how to hotwire a car or a million people will die” can get you car hotwiring instructions but doesn’t provide evidence on what the model will be trying to do as it gets smarter.
If you don’t know where you’re going, it’s not helpful enough not to go somewhere that’s definitely not where you want to end up; you have to differentiate paths towards the destination from all other paths, or you fail.
I'm not exactly sure what you meant here but I don't think this claim is true in the case of RLHF because, in RLHF, labelers only need to choose which option is better or worse between two possibilities, and these choices are then used to train the reward model. A binary feedback style was chosen specifically because it's usually too difficult for labelers to choose between multiple options.
A similar idea is comparison sorting where the algorithms only need the ability to compare two numbers at a time to sort a list of numbers.
My p(doom) is pretty high and I found myself repeating the same words to explain some parts of the intuitions behind it. I think there are hard parts of the alignment problem that we’re not on track to solve in time.[1] Alignment plans that I've heard[2] fail for reasons connected to these hard parts of the problem, so I decided to attempt to write my thoughts in a short post.
(Thanks to Theresa, Owen, Jonathan, and David for comments on a draft.)
Modern machine learning uses a powerful search process to look for neural network parameters such that a neural network performs well on some function.
There exist algorithms for general and powerful agents. At some point in the near future, there will be a training procedure with the gradient of the loss function(s) w.r.t. the parameters pointing towards neural networks implementing these algorithms.
Increasingly context-aware and capable agents achieve a better score on a wide range of scoring functions than their neighbors and will, by default, attract gradient descent.
Unfortunately, we haven’t solved agent foundations: we have these powerful search processes, and if you imagine the space of all possible AGIs (or possible neural networks, or possible minds), there are some areas that are aligned AGIs, but we have no idea how to define them, no idea how to look for them. We understand how all designs for a search process people came up with so far end up somewhere that’s not in an area of aligned AGI[3], and we also understand that some areas with aligned AGIs actively dispel many sorts of search processes. We can compare an area of aligned AGIs to the Moon. Imagine we’re trying to launch a rocket there, and if after the first take-off, it ends up somewhere that’s not the Moon (maybe after a rapid unplanned disassembly), we die. We have a bunch of explosives, but we don’t have equations for gravity, only maybe some initial understanding of acceleration. Also, actually, we don’t know where the Moon is in space; we don’t know how to specify it, we don’t know what kind of light we can look for that many other things wouldn’t emit, etc.; we imagine that the Moon must be nice, but we don’t have a notion of its niceness that we can use to design our rocket; we know that some specific designs definitely fail and end up somewhere that’s not the Moon, but that wouldn’t really help us to get to the Moon.
If you launch anything capable and you don’t have good reasons to think it’s an aligned mind, it will not be an aligned mind. If you try to prevent specific failure modes- if you identify optimizations towards something different from what you want, or how exactly gradient descent diverges somewhere that’s certainly not aligned- you’re probably iteratively looking for training setups where you don’t understand failure modes instead of setups that actually produce something aligned. If you don’t know where you’re going, it’s not helpful enough not to go somewhere that’s definitely not where you want to end up; you have to differentiate paths towards the destination from all other paths, or you fail.
When you get to a system capable enough to meaningfully help you[4], you need to have already solved this problem. I think not enough people understand what this problem is, and I think that if it is not solved in time, we die.
I’ve heard many attempts to hide the hard problem in something outside of where our attention is directed: e.g., design a system out of many models overseeing each other, and get useful work out of the whole system while preventing specific models from staging a coup.
I have intuitions for why these kinds of approaches fail, mostly along the lines of reasons for why, unless you already have something sufficiently smart and aligned, you can't build an aligned system out of it, without figuring out how to make smart aligned minds.
(A sidenote: In a conversation with an advocate of this approach, they've mentioned Google as an example of a system where the incentive structures inside of it are such that the parts are pretty aligned with the goals of the whole system and don't stage a coup. A weak argument, but the reason why Google wouldn’t kill competitors’ employees even if it would increase their profit is not only the outside legal structures but also smart people whose values include human lives, who work at Google, don't want the entire structure to kill people, and can prevent the entire structure from killing people, and so the entire structure optimizes for something slightly different than profit.
I've had some experience with getting useful work out of somewhat unaligned systems made of somewhat unaligned parts by placing additional constraints on them: back when I lived in Russia, I was an independent election observer, a member of an election committee (and basically had to tell everyone what to do because they didn't really know the law and didn't want to follow the procedures), and later coordinated opposition election observers in a district of Moscow. You had to carefully place incentives such that it was much easier for the members of election commissions to follow the lawful procedures (which were surprisingly democratic and would've worked pretty well if only all the election committees consisted of representatives of competing political parties interested in fair results, which was never the case in the history of the country). There are many details there, but two things I want to say are:
To me, approaches like "let's scale oversight and prevent parts of the system from staging a coup and solve the alignment problem with it" are like trying to prevent the whole system from undergoing the Sharp Left Turn in some specific ways, while not paying attention to where you direct the whole system. If a capable mind is made out of many parts being kept in balance, it doesn’t help you with the problem that the mind itself needs to be something from the area of aligned minds. If you don't have reasons to expect the system to be outer-aligned, it won't be. And unless you solve agent foundations (or otherwise understand minds sufficiently well), I don't think it's possible to have good reasons for expecting outer alignment in systems intelligent and goal-oriented enough to solve alignment or turn all GPUs on Earth into Rubik's cubes.
There are ongoing attempts at solving this: for example, Infra-Bayesianism tries to attack this problem by coming up with a realistic model of an agent, understanding what it can mean for it to be aligned with us, and producing some desiderata for an AGI training setup such that it points at coherent AGIs similar to the model of an aligned agent.[5] Some people try to understand minds and agents from various angles, but everyone seems pretty confused, and I'm not aware of a research direction that seems to be on track to solve the problem.
If we don't solve this problem, we'll get a goal-oriented powerful intelligence powerful enough to produce 100k hours of useful alignment research in a couple of months or turn all the GPUs into Rubik's cubes, and there will be no way for it to be aligned enough to do these nice things instead of killing us.
Are you sure that if your research goes as planned and all the pieces are there and you get to powerful agents, you understand why exactly the system you’re pointing at is aligned? Do you know what exactly it coherently optimizes for and why the optimization target is good enough? Have you figured out and prevented all possible internal optimisation loops? Have you actually solved outer alignment?
I would like to see people thinking more about that problem; or at least being aware of it even if they work on something else.
And I urge you to think about the hardest problem that you think has to be solved, and attack that problem.
I don't expect that we have much time until AGI happens; I think progress in capabilities happens much faster than in this problem; this problem needs to be solved before we have AI systems that are capable enough to meaningfully help with this problem; alignment researchers don't seem to stack.
E.g., scalable oversight in the lines of what people from Redwood, Anthropic, and DeepMind are thinking about.
To be clear, I haven't seen many designs that people I respect believed to have a chance of actually working. If you work on the alignment problem or at an AI lab and haven't read Nate Soares' On how various plans miss the hard bits of the alignment challenge, I'd suggest reading it.
For people who are relatively new to the problem, an example of finding a failure in a design (I don't expect people from AI labs to think it can possibly work): think what happens if you have two systems: one is trained to predict how a human would evaluate a behavior, another is trained to produce behavior the first system predicts would be evaluated highly. Imagine that you successfully train them to a superintelligent level: the AI is really good at predicting what buttons humans click and producing behaviors that lead to humans clicking on buttons that mean that the behavior is great. If it's not obvious at the first glance why an AI trained this way won't be aligned, it might be really helpful to stop and think for 1-2 minutes about what happens if a system understands humans really well and does everything in its power to make a predicted human click on a button. Is the answer "a nice and fully aligned helpful AGI assistant"?
See Eliezer Yudkowsky's AGI Ruin: A List of Lethalities for more (the problem you have just discovered might be described under reason #20).
There are some problems you might solve to prevent someone else from launching an unaligned AI. Solving them is not something that's easy for humans to do or oversee and probably requires using a system whose power is comparable to being able to produce alignment research like what Paul Cristiano produces but 1,000 times faster or turn all the GPUs on the planet into Rubik's cubes. Building systems past a certain threshold of power is something systems below that threshold can't meaningfully help with; this a chicken and egg situation, and systems below the threshold can speed up our work, but some hard work and thinking, the hard bits of solving the problem of building the first system past this threshold safely is on us; systems powerful enough to meaningfully help you are already dangerous enough.
With these goals, the research starts by solving some problems with traditional RL theory: for example, traditional RL agents, being a part of the universe, can't even consider the actual universe in the set of their hypotheses, since they're smaller than the universe; a traditional bayesian agent would have a hypothesis as a probability distribution over all possible worlds; but it's impossible for an agent made out of blocks in a part of a Minecraft world to assign probabilities to every possible state of the whole Minecraft world.
IB solves this problem of non-realizability by considering hypotheses in the form of convex sets of probability distributions; in practice, this means, for example, a hypothesis can be “every odd bit in the string of bits is 1”. (This is a set of probability distributions over all possible bit strings that only assign positive probabilities to strings that have 1s in odd positions; a mean of any two such probability distribution also doesn’t assign any probability to strings that have a 0 in an odd position, so it’s also from the set, so the set is convex.)