Co-founder @ Gladstone AI.
Contact: edouard@gladstone.ai
Website: eharr.is
Edouard Harris
Karma: 558
Gotcha, that makes sense!
Looks awesome! Minor correction on the cost of the GPT-4 training run: the website says $40 million, but sama confirmed publicly that it was over $100M (and several news outlets have reported the latter number as well).
Done, a few days ago. Sorry thought I’d responded to this comment.
Excellent context here, thank you. I hadn’t been aware of this caveat.
Great question. This is another place where our model is weak, in the sense that it has little to say about the imperfect information case. Recall that in our scenario, the human agent learns its policy in the absence of the AI agent; and the AI agent then learns its optimal policy conditional on the human policy being fixed.
It turns out that this setup dodges the imperfect information question from the AI side, because the AI has perfect information on all the relevant parts of the human policy during its training. And it dodges the imperfect information question from the human side, because the human never considers even the existence of the AI during its training.
This setup has the advantage that it’s more tractable and easier to reason about. But it has the disadvantage that it unfortunately fails to give a fully satisfying answer to your question. It would be interesting to see if we can remove some of the assumptions in our setup to approximate the imperfect information case.
Agreed. We think our human-AI setting is a useful model of alignment in the limit case, but not really so in the transient case. (For the reason you point out.)
I think you might have reversed the definitions of and in your comment,[1] but otherwise I think you’re exactly right.
To compute (the correlation coefficient between terminal values), naively you’d have reward functions and , that respectively assign human and AI rewards over every possible arrangement of matter . Then you’d look at every such reward function pair over your joint distribution , and ask how correlated they are over arrangements of matter. If you like, you can imagine that the human has some uncertainty around both his own reward function over houses, and also over how well aligned the AI is with his own reward function.
And to compute (the correlation coefficient between instrumental values), you’re correct that some of the arrangements of matter will be intermediate states in some construction plans. So if the human and AI both want a house with a swimming pool, they will both have high POWER for arrangements of matter that include a big hole dug in the backyard. Plot out their respective POWERs at each , and you can read the correlation right off the alignment plot!
- ^
Looking again at the write-up, it would have made more sense for us to define as the terminal goal correlation coefficient, since we introduce that one first. Alas, this didn’t occur to us. Sorry for the confusion.
- ^
Good question. Unfortunately, one weakness of our definition of multi-agent POWER is that it doesn’t have much useful to say in a case like this one.
We assume AI learning timescales vastly outstrip human learning timescales as a way of keeping our definition tractable. So the only way to structure this problem in our framework would be to imagine a human is playing chess against a superintelligent AI — a highly distorted situation compared to the case of two roughly equal opponents.
On the other hand, from other results I’ve seen anecdotally, I suspect that if you gave one of the agents a purely random policy (i.e., take a random legal action at each state) and assigned the other agent some reasonable reward function distribution over material, you’d stand a decent chance of correctly identifying high-POWER states with high-mobility board positions.
You might also be interested in this comment by David Xu, where he discusses mobility as a measure of instrumental value in chess-playing.
Thanks for you comment. These are great questions. I’ll do the best I can to answer here, feel free to ask follow-ups:
On pre-committing as a negotiating tactic: If I’ve understood correctly, this is a special case of the class of strategies where you sacrifice some of your own options (bad) to constrain those of your opponent (good). And your question is something like: which of these effects is strongest, or do they cancel each other out?
It won’t surprise you that I think the answer is highly context-dependent, and that I’m not sure which way it would actually shake out in your example with Fred and Bob and the $5000. But interestingly, we did in fact discover an instance of this class of “sacrificial” strategies in our experiments!
You can check out the example in Part 3 if you’re interested. But briefly, what happens is that when the agents get far-sighted enough, one of them realizes that there is instrumental value in having the option to bottle up the other agent in a dead-end corridor (i.e., constraining that other agent’s options). But it can only actually do this by positioning itself at the mouth of the corridor (i.e., sacrificing its own options). Here is a full-size image of both agents’ POWERs in this situation. You can see from the diagram that Agent A prefers to preserve its own options over constraining Agent H’s options in this case. But crucially, Agent A values the option of being able to constrain Agent H’s options.
In the language of your negotiating example, there is instrumental value in preserving one’s option to pre-commit. But whether actually pre-committing is instrumentally valuable or not depends on the context.
On babies being more powerful than adults: Yes, I think your reasoning is right. And it would be relatively easy to do this experiment! All you’d need would be to define a “death” state, and set your transition dynamics so that the agent gets sent to the “death” state after N turns and can never escape from it afterwards. I think this would be a very interesting experiment to run, in fact.
On paperclip maximizers: This is a very deep and interesting question. One way to think about this schematically might be: a superintelligent paperclip maximizer will go through a Phase One, in which it accumulates its POWER; and then a Phase Two in which it spends the POWER it’s accumulated. During the accumulation phase, the system might drive towards a state where (without loss of generality) the Planet Earth is converted into a big pile of computronium. This computronium-Earth state is high-POWER, because it’s a common “way station” state for paperclip maximizers, thumbtack maximizers, safety pin maximizers, No. 2 pencil maximizers, and so on. (Indeed, this is what high POWER means.)
Once the system has the POWER it needs to reach its final objective, it will begin to spend that POWER in ways that maximize its objective. This is the point at which the paperclip, thumbtack, safety pin, and No. 2 pencil maximizers start to diverge from one another. They will each push the universe towards sharply different terminal states, and the more progress each maximizer makes towards its particular terminal state, the fewer remaining options it leaves for itself if its goal were to suddenly change. Like a male praying mantis, a maximizer ultimately sacrifices its whole existence for the pursuit of its terminal goal. In other words: zero POWER should be the end state of a pure X-maximizer![1]
My story here is hypothetical, but this is absolutely an experiment on can do (at small scale, naturally). The way to do it would be to run several rollouts of an agent, and plot the POWER of the agent at each state it visits during the rollout. Then we can see whether most agent trajectories have the property where their POWER first goes up (as they, e.g., move to topological junction points) and then goes down (as they move from the junction points to their actual objectives).
Thanks again for your great questions. Incidentally, a big reason we’re open-sourcing our research codebase is to radically lower the cost of converting thought experiments like the above into real experiments with concrete outcomes that can support or falsify our intuitions. The ideas you’ve suggested are not only interesting and creative, they’re also cheaply testable on our existing infrastructure. That’s one reason we’re excited to release it!
- ^
Note that this assumes the maximizer is inner aligned to pursue its terminal goal, the terminal goal is stable on reflection, and all the usual similar incantations.
Yes, I think this is right. It’s been pointed out elsewhere that feature universality in neural networks could be an instance of instrumental convergence, for example. And if you think about it, to the extent that a “correct” model of the universe exists, then capturing that world-model in your reasoning should be instrumentally useful for most non-trivial terminal goals.
We’ve focused on simple gridworlds here, partly because they’re visual, but also because they’re tractable. But I suspect there’s a mapping between POWER (in the RL context) and generalizability of features in NNs (in the context of something like the circuits work linked above). This would be really interesting to investigate.
Got it. That makes sense, thanks!
This is really interesting. It’s hard to speak too definitively about theories of human values, but for what it’s worth these ideas do pass my intuitive smell test.
One intriguing aspect is that, assuming I’ve followed correctly, this theory aims to unify different cognitive concepts in a way that might be testable:
On the one hand, it seems to suggest a path to generalizing circuits-type work to the model-based RL paradigm. (With shards, which bid for outcomes on a contextually activated basis, being analogous to circuits, which contribute to prediction probabilities on a contextually activated basis.)
On the other hand, it also seems to generalize the psychological concept of classical conditioning (Pavlov’s salivating dog, etc.), which has tended to be studied over the short term for practical reasons, to arbitrarily (?) longer planning horizons. The discussion of learning in babies also puts one in mind of the unfortunate Little Albert Experiment, done in the 1920s:
For the experiment proper, by which point Albert was 11 months old, he was put on a mattress on a table in the middle of a room. A white laboratory rat was placed near Albert and he was allowed to play with it. At this point, Watson and Rayner made a loud sound behind Albert’s back by striking a suspended steel bar with a hammer each time the baby touched the rat. Albert responded to the noise by crying and showing fear. After several such pairings of the two stimuli, Albert was presented with only the rat. Upon seeing the rat, Albert became very distressed, crying and crawling away.
[...]
In further experiments, Little Albert seemed to generalize his response to the white rat. He became distressed at the sight of several other furry objects, such as a rabbit, a furry dog, and a seal-skin coat, and even a Santa Claus mask with white cotton balls in the beard.A couple more random thoughts on stories one could tell through the lens of shard theory:
As we age, if all goes well, we develop shards with longer planning horizons. Planning over longer horizons requires more cognitive capacity (all else equal), and long-horizon shards do seem to have some ability to either reinforce or dampen the influence of shorter-horizon shards. This is part of the continuing process of “internally aligning” a human mind.
Introspectively, I think there is also an energy cost involved in switching between “active” shards. Software developers understand this as context-switching, actively dislike it, and evolve strategies to minimize it in their daily work. I suspect a lot of the biases you might categorize under “resistance to change” (projection bias, sunk cost fallacy and so on) have this as a factor.
I do have a question about your claim that shards are not full subagents. I understand that in general different shards will share parameters over their world-model, so in that sense they aren’t fully distinct — is this all you mean? Or are you arguing that even a very complicated shard with a long planning horizon (e.g., “earn money in the stock market” or some such) isn’t agentic by some definition?
Anyway, great post. Looking forward to more.
Nice. Congrats on the launch! This is an extremely necessary line of effort.
Interesting. The specific idea you’re proposing here may or may not be workable, but it’s an intriguing example of a more general strategy that I’ve previously tried to articulate in another context. The idea is that it may be viable to use an AI to create a “platform” that accelerates human progress in an area of interest to existential safety, as opposed to using an AI to directly solve the problem or perform the action.
Essentially:
A “platform” for work in domain X is something that removes key constraints that would otherwise have consumed human time and effort when working in X. This allows humans to explore solutions in X they wouldn’t have previously — whether because they’d considered and rejected those solution paths, or because they’d subconsciously trained themselves not to look in places where the initial effort barrier was too high. Thus, developing an excellent platform for X allows humans to accelerate progress in domain X relative to other domains, ceteris paribus. (Every successful platform company does this. e.g., Shopify, Amazon, etc., make valuable businesses possible that wouldn’t otherwise exist.)
For certain carefully selected domains X, a platform for X may plausibly be relatively easier to secure & validate than an agent that’s targeted at some specific task x ∈ X would be. (Not easy; easier.) It’s less risky to validate the outputs of a platform and leave the really dangerous last-mile stuff to humans, than it would be to give an end-to-end trained AI agent a pivotal command in the real world (i.e., “melt all GPUs”) that necessarily takes the whole system far outside its training distribution. Fundamentally, the bet is that if humans are the ones doing the out-of-distribution part of the work, then the output that comes out the other end is less likely to have been adversarially selected against us.
(Note that platforms are tools, and tools want to be agents, so a strategy like this is unlikely to arise along the “natural” path of capabilities progress other than transiently.)
There are some obvious problems with this strategy. One is that point 1 above is no help if you can’t tell which of the solutions the humans come up with are good, and which are bad. So the approach can only work on problems that humans would otherwise have been smart enough to solve eventually, given enough time to do so (as you already pointed out in your example). If AI alignment is such a problem, then it could be a viable candidate for such an approach. Ditto for a pivotal act.
Another obvious problem is that capabilities research might benefit from the similar platforms that alignment research can. So actually implementing this in the real world might just accelerate the timeline for everything, leaving us worse off. (Absent an intervention at some higher level of coordination.)
A third concern is that point 2 above could be flat-out wrong in practice. Asking an AI to build a platform means asking for generalization, even if it is just “generalization within X”, and that’s playing a lethally dangerous game. In fact, it might well be lethal for any useful X, though that isn’t currently obvious to me. e.g., AlphaFold2 is a primitive example of a platform that that’s useful and non-dangerous, though it’s not useful enough for this.
On top of all that, there are all the steganographic considerations — AI embedding dangerous things in the tool itself, etc. — that you pointed out in your example.
But this strategy still seems like it could bring us closer to the Pareto frontier for critical domains (alignment problem, pivotal act), than it would be to directly train an AI to do the dangerous action.
Yep, I’d say I intuitively agree with all of that, though I’d add that if you want to specify the set of “outcomes” differently from the set of “goals”, then that must mean you’re implicitly defining a mapping from outcomes to goals. One analogy could be that an outcome is like a thermodynamic microstate (in the sense that it’s a complete description of all the features of the universe) while a goal is like a thermodynamic macrostate (in the sense that it’s a complete description of the features of the universe that the system can perceive).
This mapping from outcomes to goals won’t be injective for any real embedded system. But in the unrealistic limit where your system is so capable that it has a “perfect ontology” — i.e., its perception apparatus can resolve every outcome / microstate from any other — then this mapping converges to the identity function, and the system’s set of possible goals converges to its set of possible outcomes. (This is the dualistic case, e.g., AIXI and such. But plausibly, we also should expect a self-improving systems to improve its own perception apparatus such that its effective goal-set becomes finer and finer with each improvement cycle. So even this partition over goals can’t be treated as constant in the general case.)
Gotcha. I definitely agree with what you’re saying about the effectiveness of incentive structures. And to be clear, I also agree that some of the affordances in the quote reasonably fall under “alignment”: e.g., if you explicitly set a specific mission statement, that’s a good tactic for aligning your organization around that specific mission statement.
But some of the other affordances aren’t as clearly goal-dependent. For example, iterating quickly is an instrumentally effective strategy across a pretty broad set of goals a company might have. That (in my view) makes it closer to a capability technique than to an alignment technique. i.e., you could imagine a scenario where I succeeded in building a company that iterated quickly, but I failed to also align it around the mission statement I wanted it to have. In this scenario, my company was capable, but it wasn’t aligned with the goal I wanted.
Of course, this is a spectrum. Even setting a specific mission statement is an instrumentally effective strategy across all the goals that are plausible interpretations of that mission statement. And most real mission statements don’t admit a unique interpretation. So you could also argue that setting a mission statement increases the company’s capability to accomplish goals that are consistent with any interpretation of it. But as a heuristic, I tend to think of a capability as something that lowers the cost to the system of accomplishing any goal (averaged across the system’s goal-space with a reasonable prior). Whereas I tend to think of alignment as something that increases the relative cost to the system of accomplishing classes of goals that the operator doesn’t want.
I’d be interested to hear whether you have a different mental model of the difference, and if so, what it is. It’s definitely possible I’ve missed something here, since I’m really just describing an intuition.
Thanks, great post.
These include formulating and repeating a clear mission statement, setting up a system for promotions that rewards well-calibrated risk taking, and iterating quickly at the beginning of the company in order to habituate a rhythm of quick iteration cycles.
I may be misunderstanding, but wouldn’t these techniques fall more under the heading of capabilities rather than under alignment? These are tactics that should increase a company’s effectiveness in general, for most reasonable mission statements or products the company could have.
This is fantastic. Really appreciate both the detailed deep-dive in the document, and the summary here. This is also timely, given that teams working on superscale models with concerning capabilities haven’t generally been too forthcoming with compute estimates. (There are exceptions.)
As you and Alex point out in the sibling thread, the biggest remaining fudge factors seem to be:
Mixture models (or any kind of parameter-sharing, really) for the first method, which will cause you to systematically overestimate the “Operations per forward pass” factor; and
Variable effective utilization rates of custom hardware for the second method, which will cause an unknown distribution of errors in the “utilization rate” factor.
Nonetheless, my flying guess would be that your method is pretty much guaranteed to be right within an OOM, and probably within a factor of 2 or less. That seems pretty good! It’s certainly an improvement over anything I’ve seen previously along these lines. Congrats!
It’s simply because we each (myself more than her) have an inclination to apply a fair amount of adjustment in a conservative direction, for generic “burden of proof” reasons, rather than go with the timelines that seem most reasonable based on the report in a vacuum.
While one can sympathize with the view that the burden of proof ought to lie with advocates of shorter timelines when it comes to the pure inference problem (“When will AGI occur?”), it’s worth observing that in the decision problem (“What should we do about it?”) this situation is reversed. The burden of proof in the decision problem probably ought instead to lie with advocates of non-action: when one’s timelines are >1 generation, it is a bit too easy to kick the can down the road in various ways — leaving one unprepared if the future turns out to move faster than we expected. Conversely someone whose timelines are relatively short may take actions today that will leave us in a better position in the future, even if that future arrives more slowly than they believed originally.
(I don’t think OpenPhil is confusing these two, just that in a conversation like this it is particularly worth emphasizing the difference.)
Thanks very much for writing this. We appreciate all the feedback across the board, and I think this a well done and in-depth write up.
On the specific numerical thresholds in the report (i.e., your Key Proposal section), I do need to make one correction that also applies to most of Brooks’s commentary. All the numerical thresholds mentioned in the report, and particularly in that subsection, are solely examples and not actual recommendations. They are there only to show how one can calculate self-consistent licensing thresholds under the principles we recommend. They are not themselves recommendations. We had to do it this way for the same reason we propose granting fairly broad rule-setting flexibility to the regulatory entity. The field is changing so quickly that any concrete threshold risks being out of date (for one reason or the other) in very short order. We would have liked to do otherwise, but that is not a realistic expectation for a report that we expect to be digested over the course of several months.
To avoid precisely this misunderstanding, the report states in several places that those very numbers are, in fact, only examples for illustration. A few screencaps of those disclaimers are below, but there are several others. Of course we could have included even more, but beyond a certain point one is simply adding more length to what you correctly point out is already quite a sizeable document. Note that the Time article, in the excerpt you quoted, does correctly note and acknowledge that the Tier 3 AIMD threshold is there as an example (emphasis added):
Apart from this, I do think overall you’ve done a good and accurate job of summarizing the document and offering sensible and welcome views, emphasis, and pushback. It’s certainly a long report, so this is a service to anyone who’s looking to go one or two levels deeper than the Executive Summary. We do appreciate you giving it a look and writing it up.