Wei Dai(Wei Dai)
Some evidence in favor of your explanation (being at least a correct partial explanation):
von Neuman apparently envied Einstein’s physics intuitions, while Einstein lacked von Neuman’s math skills. This seems to suggest that they were “tuned” in slightly different directions.
Neither of the two seem superhumanly accomplished in other areas (that a smart person/agent might have goals for), such as making money, moral/philosophical progress, changing culture/politics in their preferred direction.
(An alternative explanation for 2 is that they could have been superhuman in other areas but their terminal goals did not chain through instrumental goals in those areas, which in turn raises the question of what those terminal goals must have been for this explanation to be true and what that says about human values.)
I note that under your explanation, someone could surprise the world by tuning a not-particularly-advanced AI for a task nobody previously thought to tune AI for, or by inventing a better tuning method (either general or specialized), thus achieving a large capability jump in one or more domains. Not sure how worrisome this is though.
A government might model the situation as something like “the first country/coalition to open up an AI capabilities gap of size X versus everyone else wins” because it can then easily win a tech/cultural/memetic/military/economic competition against everyone else and take over the world. (Or a fuzzy version of this to take into account various uncertainties.) Seems like a very different kind of utility function.
Hmm, open models make it easier for a corporation to train closed models, but also make that activity less profitable, whereas for a government the latter consideration doesn’t apply or has much less weight, so it seems much clearer that open models increase overall incentive for AI race between nations.
I think open source models probably reduce profit incentives to race, but can increase strategic (e.g., national security) incentives to race. Consider that if you’re the Chinese government, you might think that you’re too far behind in AI and can’t hope to catch up, and therefore decide to spend your resources on other ways to mitigate the risk of a future transformative AI built by another country. But then an open model is released, and your AI researchers catch up to near state-of-the-art by learning from it, which may well change your (perceived) tradeoffs enough that you start spending a lot more on AI research.
What do you think of this post by Tammy?
It seems like someone could definitely be wrong about what they want (unless normative anti-realism is true and such a sentence has no meaning). For example consider someone who thinks it’s really important to be faithful to God and goes to church every Sunday to maintain their faith and would use a superintelligent religious AI assistant to help keep the faith if they could. Or maybe they’re just overconfident about their philosophical abilities and would fail to take various precautions that I think are important in a high-stakes reflective process.
Mostly that thing where we had a lying vs lie-detecting arms race and the liars mostly won by believing their own lies and that’s how we have things like overconfidence bias and self-serving bias and a whole bunch of other biases.
Are you imagining that the RL environment for AIs will be single-player, with no social interactions? If yes, how will they learn social skills? If no, why wouldn’t the same thing happen to them?
Unless we do a very stupid thing like reading the AI’s thoughts and RL-punish wrongthink, this seems very unlikely to happen.
We already RL-punish AIs for saying things that we don’t like (via RLHF), and in the future will probably punish them for thinking things we don’t like (via things like interpretability). Not sure how to avoid this (given current political realities) so safety plans have to somehow take this into account.
Retinoids, which is a big family of compounds but I would go with adapalene, which has better safety/side effect than anything else. It has less scientific evidence for anti-aging than other retinoids (and is not marketed for that purpose), but I’ve tried it myself (bought it for acne), and it has very obvious anti-wrinkle effects within like a week. You can get generic 0.1% adapalene gel on Amazon for 1.6oz/$12.
(I’m a little worried about long term effects, i.e. could the increased skin turnover mean faster aging in the long run, but can’t seem to find any data or discussion about it.)
I would honestly be pretty comfortable with maximizing SBF’s CEV.
Yikes, I’m not even comfortable maximizing my own CEV. One crux may be that I think a human’s values may be context-dependent. In other words, current me-living-in-a-normal-society may have different values from me-given-keys-to-the-universe and should not necessarily trust that version of myself. (Similar to how earlier idealistic Mao shouldn’t have trusted his future self.)
My own thinking around this is that we need to advance metaphilosophy and social epistemology, engineer better discussion rules/norms/mechanisms and so on, design a social process that most people can justifiably trust in (i.e., is likely to converge to moral truth or actual representative human values or something like that), then give AI a pointer to that, not any individual human’s reflection process which may be mistaken or selfish or skewed.
TLDR: Humans can be powerful and overconfident. I think this is the main source of human evil. I also think this is unlikely to naturally be learned by RL in environments that don’t incentivize irrationality (like ours did).
Where is the longer version of this? I do want to read it. :) Specifically, what is it about the human ancestral environment that made us irrational, and why wouldn’t RL environments for AI cause the same or perhaps a different set of irrationalities?
Also, how does RL fit into QACI? Can you point me to where this is discussed?
Luckily the de-facto nominees for this position are alignment researchers, who pretty strongly self-select for having cosmopolitan altruistic values.
But we could have said the same thing of SBF, before the disaster happened.
Due to very weird selection pressure, humans ended up really smart but also really irrational. [...] An AGI (at least, one that comes from something like RL rather than being conjured in a simulation or something else weird) will probably end up with a way higher rationality:intelligence ratio, and so it will be much less likely to destroy everything we value than an empowered human.
Please explain your thinking behind this?
Dealing with moral uncertainty is just part of expected utility maximization.
It’s not, because some moral theories are not compatible with EU maximization, and of the ones that are, it’s still unclear how to handle uncertainty between them.
the inductive bias doesn’t precisely match human vision, so it has different mistakes, but as you scale both architectures they become more similar. that’s exactly what you’d expect for any approximately Bayesian setup.
I can certainly understand that as you scale both architectures, they both make less mistakes on distribution. But do they also generalize out of training distribution more similarly? If so, why? Can you explain this more? (I’m not getting your point from just “approximately Bayesian setup”.)
They needed a giant image classification dataset which I don’t think even existed 5 years ago.
This is also confusing/concerning for me. Why would it be necessary or helpful to have such a large dataset to align the shape/texture bias with humans?
Do you know if it is happening naturally from increased scale, or only correlated with scale (people are intentionally trying to correct the “misalignment” between ML and humans of shape vs texture bias by changing aspects of the ML system like its training and architecture, and simultaneously increasing scale)? I somewhat suspect the latter due the existence of a benchmark that the paper seems to target (“humans are at 96% shape / 4% texture bias and ViT-22B-384 achieves a previously unseen 87% shape bias / 13% texture bias”).
In either case, it seems kind of bad that it has taken a decade or two to get to this point from when adversarial examples were first noticed, and it’s unclear whether other adversarial examples or “misalignment” remain in the vision transformer. If the first transformative AIs don’t quite learn the right values due to having a different inductive bias from humans, it may not matter much that 10 years later the problem would be solved.
Traditionally, those techniques are focused on what the model is outputting, not what the model’s underlying motivations are. But I haven’t read all the literature. Am I missing something?
It’s confusing to me as well, perhaps because different people (or even the same person at different times) emphasize different things within the same approach, but here’s one post where someone said, “It is important that the overseer both knows which action the distilled AI wants to take as well as why it takes that action.”
Did SBF or Mao Zedong not have a pointer to the right values, or had a right pointer but made mistakes due to computational issues (i.e., would have avoided causing the disasters that they did if they were smarter and/or had more time to think)? Both seem possible to me, so I’d like to understand how the QACI approach would solve (or rule out) both of these potential problems:
If many humans don’t have pointers to right values, how to make sure QACI gets a pointer from humans who have a pointer to the right values?
How to make sure that AI will not make some catastrophic mistake while it’s not smart enough to fully understand the values we give it, while still being confident enough in its guesses of what to do in the short term to do useful things?
Moral uncertainty is an area in philosophy with ongoing research, and assuming that AI will handle it correctly by default seems unsafe, similar to assuming that AI will have the right decision theory by default.
I see that Tasmin Leake also pointed out 2 above as a potential problem, but I don’t see anything that looks like a potential solution at QACI table of contents.
Katja Grace notes that image synthesis methods have no trouble generating photorealistic human faces.
They’re terrible at hands though (which has ruined many otherwise good images for me). That post used Stable Diffusion 1.5, but even the latest SD 3.0 (with versions 2.0, 2.1, XL, Stable Cascade in between) is still terrible at it.
Don’t really know how relevant this is to your point/question about fragility of human values, but thought I’d mention it since it seems plausibly as relevant as AIs being able to generate photorealistic human faces.
Adversarial examples suggest to me that by default ML systems don’t necessarily learn what we want them to learn:
They put too much emphasis on high frequency features, suggesting a different inductive bias from humans.
They don’t handle contradictory evidence in a reasonable way, i.e., giving a confident answer when high frequency features (pixel-level details) and low frequency features (overall shape) point to different answers.
The evidence from adversarial training suggests to me that AT is merely patching symptoms (e.g., making the ML system de-emphasize certain specific features) and not fixing the underlying problem. At least this is my impression from watching this video on Adversarial Robustness, specifically the chapters on Adversarial Arms Race and Unforeseen Adversaries.
Aside from this, it’s also unclear how to apply AT to your original motivation:
A function that tells your AI system whether an action looks good and is right virtually all of the time on natural inputs isn’t safe if you use it to drive an enormous search for unnatural (highly optimized) inputs on which it might behave very differently.
because in order to apply AT we need a model of what “attacks” the adversary is allowed to do (in this case the “attacker” is a superintelligence trying to optimize the universe, so we have to model it as being allowed to do anything?) and also ground-truth training labels.
For this purpose, I don’t think we can use the standard AT practice of assuming that any data point within a certain distance of a human-labeled instance, according to some metric, has the same label as that instance. Suppose we instead let the training process query humans directly for training labels (i.e., how good some situation is) on arbitrary data points, well that’s slow/costly if the process isn’t very sample efficient (which modern ML isn’t), and also scary if human implementations of human values may already have adversarial examples. (The “perceptual wormholes” work and other evidence suggest that humans also aren’t 100% adversarially robust.)
My own thinking is that we probably need to go beyond adversarial training for this, along the lines of solving metaphilosophy and then using that solution to find/fix existing adversarial examples and correctly generalize human values out of distribution.
I’m confused about how heterogeneity in data quality interacts with scaling. Surely training a LM on scientific papers would give different results from training it on web spam, but data quality is not an input to the scaling law… This makes me wonder whether your proposed forecasting method might have some kind of blind spot in this regard, for example failing to take into account that AI labs have probably already fed all the scientific papers they can into their training processes. If future LMs train on additional data that have little to do with science, could that keep reducing overall cross-entropy loss (as scientific papers become a smaller fraction of the overall corpus) but fail to increase scientific ability?
Thank you for detailing your thoughts. Some differences for me:
I’m also worried about unaligned AIs as a competitor to aligned AIs/civilizations in the acausal economy/society. For example, suppose there are vulnerable AIs “out there” that can be manipulated/taken over via acausal means, unaligned AI could compete with us (and with others with better values from our perspective) in the race to manipulate them.
I’m perhaps less optimistic than you about commitment races.
I have some credence on max good and max bad being not close to balanced, that additionally pushes me towards the “unaligned AI is bad” direction.
ETA: Here’s a more detailed argument for 1, that I don’t think I’ve written down before. Our universe is small enough that it seems plausible (maybe even likely) that most of the value or disvalue created by a human-descended civilization comes from its acausal influence on the rest of the multiverse. An aligned AI/civilization would likely influence the rest of the multiverse in a positive direction, whereas an unaligned AI/civilization would probably influence the rest of the multiverse in a negative direction. This effect may outweigh what happens in our own universe/lightcone so much that the positive value from unaligned AI doing valuable things in our universe as a result of acausal trade is totally swamped by the disvalue created by its negative acausal influence.
Perhaps half of the value of misaligned AI control is from acausal trade and half from the AI itself being valuable.
Why do you think these values are positive? I’ve been pointing out, and I see that Daniel Kokotajlo also pointed out in 2018 that these values could well be negative. I’m very uncertain but my own best guess is that the expected value of misaligned AI controlling the universe is negative, in part because I put some weight on suffering-focused ethics.
If something is both a vanguard and limited, then it seemingly can’t stay a vanguard for long. I see a few different scenarios going forward:
We pause AI development while LLMs are still the vanguard.
The data limitation is overcome with something like IDA or Debate.
LLMs are overtaken by another AI technology, perhaps based on RL.
In terms of relative safety, it’s probably 1 > 2 > 3. Given that 2 might not happen in time, might not be safe if it does, or might still be ultimately outcompeted by something else like RL, I’m not getting very optimistic about AI safety just yet.
The argument is that with 1970′s tech the soviet union collapsed, however with 2020 computer tech (not needing GenAI) it would not.
I note that China is still doing market economics, and nobody is trying (or even advocating, AFAIK) some very ambitious centrally planned economy using modern computers, so this seems like pure speculation? Has someone actually made a detailed argument about this, or at least has the agreement of some people with reasonable economics intuitions?
I think it is considered a constraint by some because they think that it would be easier/safer to use a superintelligent AI to do simpler actions, while alignment is not yet fully solved. In other words, if alignment was fully solved, then you could use it to do complicated things like what you suggest, but there could be an intermediate stage of alignment progress where you could safely use SI to do something simple like “melt GPUs” but not to achieve more complex goals.