Perhaps these can be thought of as homework questions -- when I imagine us successfully making AI go well, I imagine us building expertise such that we can answer these questions quickly and easily. Before I read the answers I'm going to think for 10min or so about each one and post my own guesses.

Useful links / background reading: The glorious EfficientZero: How it Works. Related comment. EfficientZero GitHub. LW discussion.

Some of these questions are about EfficientZero, the net trained recently; others are about EfficientZero the architecture, imagined to be suitably scaled up to AGI levels. "If we made a much bigger and longer-trained version of this (with suitable training environment) such that it was superhuman AGI..."

  1. EfficientZero vs. reward hacking and inner alignment failure:
    1. Barring inner alignment failure, it’ll eventually reward hack, right? That is, if it gets sufficiently knowledgeable and capable, it’ll realize that it can get loads of reward by hacking its reward channel, and then its core algorithm would evaluate that action/plan highly and do it. Right?
    2. But fortunately (?) maybe there would be an inner alignment failure and the part of it that predicts reward would predict low reward from that action, even in the limit of knowledge and capability? Because it’s learned to predict proxies for reward rather than reward itself, and continued to do so even as it got smarter and more capable? (Would this happen? Why would this not be corrected by further training? Has some sort of proxy crystallization set in? How?)
  2. EfficientZero approximates evidential decision theory, right?
  3. EfficientZero is a consequentialist (in the sense defined here) architecture, right? It’s not, for example, updateless or deontological. For example, it has no deontological constraints except by accident (i.e. if its predictor-net mistakenly predicted super low reward for actions of type X, always, even in cases where actually a reasonable intelligent predictor would predict high reward.) Right?
  4. What is the most complex environment AIs in the family of MuZero, EfficientZero, etc. have been trained on? Is it just some Atari game?
  5. Roughly how many parameters does EfficientZero have? If you don’t know, what about MuZero? What about the biggest net to date from that general family? The EfficientZero paper doesn't give a direct answer but it describes the architecture in enough detail that you might be able to calculate it...
  6. If we kept scaling up EfficientZero by OOMs in every way, what would happen? Would it eventually get to agenty AGI? / APS-AI? After all, it seems pretty sample-efficient already. What if its sample was an entire lifetime of diverse experiences?


New Answer
New Comment

1 Answers sorted by

Daniel Kokotajlo


OK, here are my guesses, without seeing anyone else's answers. I think I'm probably wrong, which is why I'm asking this question:

1.a. Underdetermined? It depends on what we mean by the outer objective, and what we mean when we assume it has no inner alignment problems? See e.g. this discussion. That said, yeah it totally seems possible. If the part that predicts reward gets good at generalizing, it should be able to reason/infer/guess that hacking the reward function would yield tons of reward. And then that's what the agent would do.

1.b. Yes? Even though I usually think of inner alignment failures on the context of black-box neural nets and this is a bit more transparent (because of the tree search at the heart) I think the usual arguments would apply?

2. Yes? My limited, probably faulty understanding is that the policy is trained to approximate what you get when you search the tree of possible actions and predicted consequences some depth and then look at how good or bad the resulting states seem to be. There's nothing in there about doing causal interventions or do-calculus or whatnot; so I assume that the predicted-consequences network would approximate conditionalization... however Caspar Oesterheld has a paper arguing I think that under some conditions neural nets approximate CDT in the limit, so *shrugs*. I should go read it again and see if it applies.

3. Yes? Again, if it's approximating what you get when you search the tree and pick the action that leads to the best predicted state, that sure seems pretty consequentialist. If initially there was some "never lie" heuristic, then wouldn't it disappear quickly once you encountered situations where lying led to a better predicted outcome? Sure, you could hack it by always predicting bad outcomes from lying, but that contradicts our "sufficiently knowledgeable and capable" hypothetical.

4. I don't know but from what I've seen the answer is yeah, just Atari and stuff like that.

5. I don't know but it seems EfficientZero took 7 hours on 4 GPUs to do one 2-subjective-hour training run. That gives us a sense of how many FLOPS it must run at (maybe 5 x 10^14 operations per subjective second?), but presumably it has many operations per parameter... I guess it would have made news if it had parameter count comparable to large language models, so maybe it has something like 10^8 parameters? But surely they aren't doing a million ops per parameter per subjective second? Gosh I really am just guessing.

6. I see no obstacle, it seems like a quantitative question of how much bigger we'd need to make it, how much longer we'd need to train it, and maybe what sort of data we'd have to give it. Qualitatively it seems like this would probably keep scaling to superintelligence if we poured enough into it (but this is probably true for lots of architectures, and doesn't mean much in practice since we don't have galaxy-sized computers.)

6 comments, sorted by Click to highlight new comments since:

it seems pretty sample-efficient already

Maybe I'm confused, but I'm not ready to take that for granted. I think it's a bit subtle.

Let's talk about chess. Suppose I start out knowing nothing about chess, and have a computer program that I can play that enforces the rules of chess, declares who wins, etc.

  • I play the computer program for 15 minutes, until I'm quite confident that I know all the rules of chess.
  • …Then I spend 8000 years sitting on a chair, mentally playing chess against myself.

If I understand correctly, "EfficientZero is really good at Atari after playing for only 15 minutes" is true in the same sense that "I am good after chess after playing for only 15 minutes" in the above scenario. The 8000 years doesn't count as samples because it was in my head.

Someone can correct me if I'm wrong…

(Even if that's right, it doesn't mean that EfficientZero isn't an important technological advance—after all, there are domains like robotics where simulation-in-your-head is much cheaper than querying the environment. But it would maybe deflate some of the broader conclusions that people are drawing from EfficientZero, like how AI compares to human brains…)

It's a little bit less dramatic than that: the model-based simulation playing is interleaved with the groundtruth environment. It's more like you spend a year playing games in your head, then you play 1 30s bullet chess match with Magnus Carlsen (madeup ratio), then go back to playing in your head for another year. Or maybe we should say, "you clone yourself a thousand times, and play yourself at correspondence chess timescales for 1 game per pair in a training montage, and then go back for a rematch".

(The scenario where you play for 15 minutes at the beginning, and then pass a few subjective eons trying to master the game with no further real data, would correspond to sample-efficient "offline reinforcement learning" (review), eg / , . Very important in its own right, of course, but poses challenges of its own related to quality of those 15 minutes - what if your 15 minute sample doesn't include any games which happen to use en passant, say? How could you ever learn that should be part of your model of chess? When you interleave and learn online, Carlsen might surprise you with an en passant a few games in, and then all your subsequent imagined games can include that possible move and you use it yourself.).

But it would maybe deflate some of the broader conclusions that people are drawing from EfficientZero, like how AI compares to human brains…

I don't think it does. If humans can't learn efficiently by imagining hypothetical games like machines can, so much the worse for them. The goal is to win.

Like Gwern said, the goal is to win. If EfficientZero gets superhuman data-efficiency by "cheating," well it still got superhuman data-efficiency...

I think a relevant comparison here would be total cost. EfficientZero took 7 hours on 4 GPUs to master each particular Atari game. Equipment probably cost around $10,000 last I checked. How long is the lifespan of a GPU? Two years? OK, so that's something like 20,000 hours of 4 GPU's time for $10K, so $0.50 for one hour, so $3.50 for the training run? Eh, maybe it's a bit more than that due to energy costs or something. But still it seems probable that it beats human wages.

For computational expenditure cost... I think 4 GPUs would be doing something like 10^14 FLOPS, which is one OOM less than the human brain?

All that being said, I do take your point that if matters if EfficientZero is getting results via an importantly different method than the human method, because that should give us a shred of hope that what EfficientZero is doing, while equally viable for this task, would be less viable for really important or dangerous tasks. However it could also be more viable. (For example, maybe the human method is "draw upon decades of pre-training and background knowledge. Also benefit from the fact that the game was literally designed for humans." Plausibly, that's our secret sauce. In which case EfficientZero is just... qualitatively smarter than us in some sense. Less dollars, less FLOPs, etc. Compensates for the lack of secret sauce by thinking really hard about its experiences, in a way we simply can't because our brains aren't fast enough.)

Hmm, I think my comment came across as setting up a horse-race between EfficientZero and human brains, in a way that I didn't intend. Sorry for bad choice of words. In particular, when I wrote "how AI compares to human brains", I meant in the sense of "In what ways are they similar vs different? What are their relative strengths and weaknesses? Etc.", but I guess it sounded like I was saying "human brain algorithms are better and EfficientZero is worse". Sorry.

I could write a "human brain algorithms are fundamentally more powerful than EfficientZero" argument, but I wasn't trying to, and such an argument sure as heck wouldn't fit in a comment. :-)

If EfficientZero gets superhuman data-efficiency by "cheating," well it still got superhuman data-efficiency...

Sure. If Atari sample efficiency is what we ultimately care about, then the results speak for themselves. For my part, I was using sample efficiency as a hint about other topics that are not themselves sample efficiency. For example, I think that if somebody wants to understand AlphaZero, the fact that it trained on 40,000,000 games of self-play is a highly relevant and interesting datapoint. Suppose you were to then say "…but of those 40,000,000 games, fundamentally it really only needed 100 games with the external simulator to learn the rules. The other 39,999,900 games might as well have been 'in its head'. This was proven in follow-up work.". I would reply: "Oh. OK. That's interesting too. But I still care about the 40,000,000 number. I still see that number as a very important part of understanding the nature of AlphaZero and similar systems."

(I'm not sure we're disagreeing about anything…)

One thing to note is that you don't know how many games humans are playing in their head in some sense. We don't have access to that kind of information about our own algorithms. Even if you think we don't because we don't consciously experience/remember them, that's obviously wrong. Every time you have a thought pop out of nowhere or an eureka! moment from the incubation effect, or every time you have a Tetris effect dream (or all the experience-replay hippocampus neuroscience), you see how it feels to have powerful subconscious algorithms churning away on difficult problems without you having any awareness of it: nothing. But they still take wallclock years to reach levels of performance that something like AlphaZero does in hours...

That's an interesting thought. My hunch is that hippocampal replay can't happen unconsciously because if the hippocampus broadcasts a memory at all, it broadcasts it broadly to the cortex including GNW. That's just my current opinion, I'm not sure if there's neuroscience consensus on that question.

Here I'm sneaking in an assumption that "activity in the GNW" = "activity that you're conscious of". Edge-cases include times when there's stuff happening in the GNW, but it's not remembered after the fact (at least, not as a first-person episodic memory). Are you "conscious" during a dream that you forget afterwards? Are you "conscious" when you're 'blacked out' from drinking too much? I guess I'd say "yes" to both, but that's a philosophy question, or maybe just terminology.

If we want more reasons that human-vs-EfficientZero comparisons are not straightforward, there's also the obvious fact that humans benefit from transfer-learning whereas EfficientZero starts with random weights.