Posts
Comments
The results seem to be cherry picked or else perhaps I am using the code incorrectly. I'm trying to use the VAE for a separate project and the encoded vectors don't steer generations very well (or reconstruct -- which is what I was hoping to use this for).
In addition to seeing more AI behavioral psychology work, I would be excited about seeing more AI developmental psychology -- i.e. studying how varying properties of training or architecture affect AI behavior. Shard theory is an example of this.
I've written a bit about the motivations for AI developmental psychology here.
I agree with this norm, though I think it would be better to say that the "burden of evidence" should be on labs. When I first read the title, I thought you wanted labs to somehow prove the safety of their system in a conclusive way. What this probably looks like in practice is "we put x resources into red teaming and didn't find any problems." I would be surprised if 'proof' was ever an appropriate term.
The analogy between AI safety and math or physics is assumed it in a lot of your writing and I think it is a source of major disagreement with other thinkers. ML capabilities clearly isn’t the kind of field that requires building representations over the course of decades.
I think it’s possible that AI safety requires more conceptual depth than AI capabilities; but in these worlds, I struggle to see how the current ML paradigm coincides with conceptual ‘solutions’ that can’t be found via iteration at the end. In those worlds, we are probably doomed so I’m betting on the worlds in which you are wrong and we must operate within the current empirical ML paradigm. It’s odd to me that you and Eliezer seem to think the current situation is very intractable, and yet you are confident enough in your beliefs to where you won’t operate on the assumption that you are wrong about something in order to bet on a more tractable world.
+1. As a toy model, consider how the expected maximum of a sample from a heavy tailed distribution is affected by sample size. I simulated this once and the relationship was approximately linear. But Soares’ point still holds if any individual bet requires a minimum amount of time to pay off. You can scalably benefit from parallelism while still requiring a minimum amount of serial time.
At its core, the argument appears to be "reward maximizing consequentialists will necessarily get the most reward." Here's a counter example to this claim: if you trained a Go playing AI with RL, you are unlikely to get a reward maximizing consequentialist. Why? There's no reason for the Go playing AI to think about how to take over the world or hack the computer that is running the game. Thinking this way would be a waste of computation. AIs that think about how to win within the boundaries of the rules therefore do better.
In the same way, if you could robustly enforce rules like "turn off when the humans tell you to" or "change your goal when humans tell you to" etc, perhaps you end up with agents that follow these rules rather than agents that think "hmmm... can I get away with being disobedient?"
Both achieve the same reward if the rules are consistently enforced during training and I think there are weak reasons to expect deontological agents to be more likely.
Also, AI startups make AI safety resources more likely to scale with AI capabilities.
It hasn't been canceled.
I don't mean to distract from your overall point though which I take to be "a philosopher said a smart thing about AI alignment despite not having much exposure." That's useful data.
I don't know why this is a critique of RLHF. You can use RLHF to train a model to ask you questions when it's confused about your preferences. To the extent to which you can easily identify when a system has this behavior and when it doesn't, you can use RLHF to produce this behavior just like you can use RLHF to produce many other desired behaviors.
Thoughts on John's comment: this is a problem with any method for detecting deception that isn't 100% accurate. I agree that finding a 100% accurate method would be nice, but good luck.
Also, you can somewhat get around this by holding some deception detecting methods out (i.e. not optimizing against them). When you finish training and the held out methods tell you that your AI is deceptive, you start over. Then you have to try to think of another approach that is more likely to actually discourage deception than fool your held out detectors. This is the difference between gradient descent search and human design search, which I think is an important distinction.
Also, FWIW, I doubt that trojans are currently a good microcosm for detecting deception. Right now, it is too easy to search for the trigger using brute force optimization. If you ported this over to sequential-decision-making land where triggers can be long and complicated, that would help a lot. I see a lot of current trojan detection research as laying the groundwork for future research that will be more relevant.
In general, it seems better to me to evaluate research by asking "where is this taking the field/what follow-up research is this motivating?" rather than "how are the words in this paper directly useful if we had to build AGI right now?" Eventually, the second one is what matters, but until we have systems that look more like agents that plan and achieve goals in the real world, I'm pretty skeptical of a lot of the direct value of empirical research.
From what I understand, Dan plans to add more object-level arguments soon.
(opinions are my own)
I think this is a good review. Some points that resonated with me:
1. "The concepts of systemic safety, monitoring, robustness, and alignment seem rather fuzzy." I don't think the difference between objective and capabilities robustness is discussed but this distinction seems important. Also, I agree that Truthful AI could easily go into monitoring.
2. "Lack of concrete threat models." At the beginning of the course, there are a few broad arguments for why AI might be dangerous but not a lot of concrete failure modes. Adding more failure modes here would better motivate the material.
3. Give more clarity on how the various ML safety techniques address the alignment problem, and how they can potentially scale to solve bigger problems of a similar nature as AIs scale in capabilities
4. Give an assessment on the most pressing issues that should be addressed by the ML community and the potential work that can be done to contribute to the ML safety field
You can read more about how these technical problems relate to AGI failure modes and how they rank on importance, tractability, and crowdedness in Pragmatic AI Safety 5. I think the creators included this content in a separate forum post for a reason.
The course is intended for to two audiences: people who are already worried about AI X-risk and people who are only interested in the technical content. The second group doesn't necessarily care about why each research direction relates to reducing X-risk.
Putting a lot of emphasis on this might just turn them off. It could give them the impression that you have to buy X-risk arguments in order to work on these problems (which I don't think is true) or it could make them less likely to recommend the course to others, causing fewer people to engage with the X-risk material overall.
These are good points. Maybe we'll align these things enough to where they'll give us a little hamster tank to run around in.
PAIS #5 might be helpful here. It explains how a variety of empirical directions are related to X-Risk and probably includes many of the ones that academics are working on.
I agree that I did not justify this claim and it is controversial in the ML community. I'll try to explain why I think this.
First of all, when I say an AI 'wants things' or is 'an agent' I just mean it robustly and autonomously brings about specific outcomes. I think all of the arguments made above apply with this definition and don't rely on anthropomorphism.
Why are we likely to build superintelligent agents?
1. It will be possible to build superintelligent agents
I'm just going to take it as a given here for the sake of time. Note: I'm not claiming it will be easy, but I don't think anyone can really say with confidence that it will be impossible to build such agents within the century.
2. There will be strong incentives to build superintelligent agents
I'm generally pretty skeptical of claims that if you just train an AI system for long enough it becomes a scary consequentialist. Instead, I think it is likely someone will build a system like this because humans want things and building a superintelligent agent that wants to do what you want pretty much solves all your problems. For example, they might want to remain technologically relevant/be able to defend themselves from other countries/extraterrestrial civilizations, expand human civilization, etc. Building a thing that 'robustly and autonomously brings about specific outcomes' would be helpful for all this stuff.
In order to use current systems ('tool AIs' as some people call them) to actually get things done in the world, a human needs to be in the loop, which is pretty uncompetitive with fully autonomous agents.
I wouldn't be super surprised if humans didn't build superintelligent agents for a while -- even if they could. Like I mentioned in the post, most people prob want to stay in control. But I'd put >50% credence on it happening pretty soon after it is possible because of coordination difficulty and the unilateralist curse.
I don't doubt this. I was more reporting on how the branding came across to me.
Your emphasis on how 'fun' your organization is is kind of off-putting (I noticed this on your website as well). I think it gives me the impression that you are not very serious about what you are doing. Maybe it's just me though.
This is because longer runs will be outcompeted by runs that start later and therefore use better hardware and better algorithms.
Wouldn't companies port their partially-trained models to new hardware? I guess the assumption here is that when more compute is available, actors will want to train larger models. I don't think this is obviously true because:
1. Data may be the bigger bottleneck. There was some discussion of this here. Making models larger doesn't help very much after a certain point compared with training them with more data.
2. If training runs are happening over months, there will be strong incentives to make use of previously trained models -- especially in a world where people are racing to build AGI. This could look like anything from slapping on more layers to developing algorithms that expand the model in all relevant dimensions as it is being trained. Here's a paper about progressive learning for vision transformers. I didn't find anything for NLP, but I also haven't looked very hard.
Claim 1: there is an AI system that (1) performs well ... (2) generalizes far outside of its training distribution.
Don't humans provide an existence proof of this? The point about there being a 'core' of general intelligence seems unnecessary.
Thanks!
I know that this is a common argument against amplification, but I've never found it super compelling.
People often point to evil corporations to show that unaligned behavior can emerge from aligned humans, but I don't think this analogy is very strong. Humans in fact do not share the same goals and are generally competing with each other over resources and power, which seems like the main source of inadequate equilibria to me.
If everyone in the world was a copy of Eliezer, I don't think we would have a coordination problem around building AGI. They would probably have an Eliezer government that is constantly looking out for emergent misalignment and suggesting organizational changes to squash it. Since everyone in this world is optimizing for making AGI go well and not for profit or status among their Eliezer peers, all you have to do is tell them what the problem is and what they need to do to fix it. You don't have to threaten them with jail time or worry that they will exploit loopholes in Eliezer law.
I think it is quite likely that I am missing something here and it would be great if you could flush this argument out a little more or direct me towards a post that does.
That's a good point. I guess I don't expect this to be a big problem because:
1. I think 1,000,000 copies of myself could still get a heck of a lot done.
2. The first human-level AGI might be way more creative than your average human. It would probably be trained on data from billions of humans, so all of those different ways of thinking could be latent in the model.
2. The copies can potentially diverge. I'm expecting the first transformative model to be stateful and be able to meta-learn. This could be as simple as giving a transformer read and write access to an external memory and training it over longer time horizons. The copies could meta-learn on different data and different sub-problems and bring different perspectives to the table.
Wait... I'm quite confused. In the decision rule, how is the set of environments 'E' determined? If it contains every possible environment, then this means I should behave like I am in the worst possible world, which would cause me to do some crazy things.
Also, when you say that an infra-bayesian agent models the world with a set of probability distributions, what does this mean? Does the set contain every distribution that would be consistent with the agent's observations? But isn't this almost all probability distributions? Some distributions match the data better than others, so do you weigh them according to P(observation | data generating distribution)? But then what would you do with these weights?
Sorry if I am missing something obvious. I guess this would have been clearer for me if you explained the infra-bayesian framework a little more before introducing the decision rule.
Interesting post! I'm not sure if I understand the connection between infra-bayesianism and Newcomb's paradox very well. The decision procedure you outlined in the first example seems equivalent to an evidential decision theorist placing 0 credence on worlds where Omega makes an incorrect prediction. What is the infra-bayesianism framework doing differently? It just looks like the credence distribution over worlds is disguised by the 'Nirvana trick.'
It's great that you are trying to develop a more detailed understanding of inner alignment. I noticed that you didn't talk about deception much. In particular, the statement below is false:
Generalization <=> accurate priors + diverse data
You have to worry about what John Wentworth calls 'manipulation of imperfect search.' You can have accurate priors and diverse data and (unless you have infinite data) the training process could produce a deceptive agent that is able to maintain its misalignment.
I'm guessing that you are referring to this:
Another strategy is to use intermittent oversight – i.e. get an amplified version of the current aligned model to (somehow) determine whether the upgraded model has the same objective before proceeding.
The intermittent oversight strategy does depend on some level of transparency. This is only one of the ideas I mentioned though (and it is not original). The post in general does not assume anything about our transparency capabilities.
I'm not sure I understand. We might not be on the same page.
Here's the concern I'm addressing:
Let's say we build a fully aligned human-level AGI, but we want to scale it up to superintelligence. This seems much harder to do safely than to train the human-level AGI since you need a training signal that's better than human feedback/imitation.
Here's the point I am making about that concern:
It might actually be quite easy to scale an already aligned AGI up to superintelligence -- even if you don't have a scalable outer-aligned training signal -- because the AGI will be motivated to crystallize its aligned objective.
Thanks for the thoughtful review! I think this is overall a good read of what I was saying. I agree now that redundancy would not work.
One clarification:
The mesaobjective that was aligned to our base objective in the original setting is no longer aligned in the new setting
When I said that the 'human-level' AGI is assumed to be aligned, I meant that it has an aligned mesa-objective (corrigibly or internally) -- not that it has an objective that was functionally aligned on the training distribution, but may not remain aligned under distribution shift. I thought that internally/corrigibly aligned mesa-objectives are intent-aligned on all (plausible) distributions by definition...
Adding some thoughts that came out of a conversation with Thomas Kwa:
Gradient hacking seems difficult. Humans have pretty weak introspective access to their goals. I have a hard time determining whether my goals have changed or if I have gained information about what they are. There isn't a good reason to believe that the AIs we build will be different.
Safety and value alignment are generally toxic words, currently. Safety is becoming more normalized due to its associations with uncertainty, adversarial robustness, and reliability, which are thought respectable. Discussions of superintelligence are often derided as “not serious”, “not grounded,” or “science fiction.”
Here's a relevant question in the 2016 survey of AI researchers:
These numbers seem to conflict with what you said but maybe I'm misinterpreting you. If there is a conflict here, do you think that if this survey was done again, the results would be different? Or do you think these responses do not provide an accurate impression of how researchers actually feel/felt (maybe because of agreement bias or something)?
I have an objection to the point about how AI models will be more efficient because they don't need to do massive parallelization:
Massive parallelization is useful for AI models too and for somewhat similar reasons. Parallel computation allows the model to spit out a result more quickly. In the biological setting, this is great because it means you can move out of the way when a tiger jumps toward you. In the ML setting, this is great because it allows the gradient to be computed more quickly. The disadvantage of parallelization is that it means that more hardware is required. In the biological setting, this means bigger brains. Big brains are costly. They use up a lot of energy and make childbearing more difficult as the skulls need to fit through the birth canal.
In the ML setting, however, big brains are not as costly. We don't need to fit our computers in a skull. So, it is not obvious to me that ML models will do fewer computations in parallel than biological brains.
Some relevant information:
- According to Scaling Laws for Neural Language Models, model performance depends strongly on model size but very weakly on shape (depth vs width).
- An explanation for the above is that deep residual networks have been observed to behave like Ensembles of shallow networks.
- GPT-3 uses 96 layers (decoder blocks). That isn't very many serial computations. If a matrix multiplication, softmax, relu, or vector addition count as atomic computations, then there are 11 serial computations per layer, so that's only 1056 serial computations. It is unclear how to compare this to biological neurons as each neuron may require a number of these serial computations to properly simulate.
- PALM has 3 times more parameters than GPT-3 but only has 118 layers.
Here's another milestone in AI development that I expect to happen in the next few years which could be worth noting:
I don't think any of the large language models that currently exist write anything to an external memory. You can get a chatbot to hold a conversation and 'remember' what was said by appending the dialogue to its next input, but I'd imagine this would get unwieldy if you want your language model to keep track of details over a large number of interactions.
Fine-tuning a language model so that it makes use of a memory could lead to:
1. More consistent behavior
2. 'Mesa-learning' (it could learn things about the world from its inputs instead of just by gradient decent)
This seems relevant from a safety perspective because I can imagine 'mesa-learning' turning into 'mesa-agency.'
I'm pretty confused about the plan to use ELK to solve outer alignment. If Cakey is not actually trained, how are amplified humans accessing its world model?
"To avoid this fate, we hope to find some way to directly learn whatever skills and knowledge Cakey would have developed over the course of training without actually training a cake-optimizing AI...
- Use imitative generalization combined with amplification to search over some space of instructions we could give an amplified human that would let them make cakes just as delicious as Cakey’s would have been.
- Avoid the problem of the most helpful instructions being opaque (e.g. “Run this physics simulation, it’s great”) by solving ELK — i.e., finding a mapping from whatever possibly-opaque model of the world happens to be most useful for making superhumanly delicious cakes to concepts humans care about like “people” being “alive.”
- Spell out a procedure for scoring predicted futures that could be followed by an amplified human who has access to a) Cakey’s great world model, and b) the correspondence between it and human concepts of interest. We think this procedure should choose scores using some heuristic along the lines of “make sure humans are safe, preserve option value, and ultimately defer to future humans about what outcomes to achieve in the world” (we go into much more detail in Appendix: indirect normativity).
- Distill their scores into a reward model that we use to train Hopefully-Aligned-Cakey, which hopefully uses its powers to help humans build the utopia we want."
I don't think I agree that this undermines my argument. I showed that the utility function of person 1 is of the form h(x + y) where h is monotonic increasing. This respects the fact that the utility function is not unique. 2(x + y) + 1 would qualify, as would 3 log(x + y), etc.
Showing that the utility function must have this form is enough to prove total utilitarianism in this case since when you compare h(x + y) to h(x'+ y'), h becomes irrelevant. It is the same as comparing x + y to x' + y'.
This is a much more agreeable assumption. When I get a chance, I'll make sure it can replace the fairness one and add it to the proof and give you credit.
I am defining it as you said. They are like movie frames that haven't been projected yet. I agree that the pre-arranged nature of the snapshots is irrelevant -- that was the point of the example (sorry that this wasn't clear).
The purpose of the example was to falsify the following hypothesis:
"In order for a simulation to produce conscious experiences, it must compute the next state based on the previous state. It can't just 'play the simulation from memory'"
Maybe what you are getting at is that this hypothesis doesn't do justice to the intuitions that inspired it. Something complex is happening inside the brain that is analogous to 'computation on-demand.' This differentiates the computer-brain system from the stack of papers that are being moved around.
This seems legit... I just would like to have a more precise understanding of what this 'computation on-demand' property is.
- If they are simulating both, then I think they are simulating a lot of things. Could they be simulating Miami on a Monday morning? If the simulation states are represented with N bits and states of Miami can be expressed with <= N bits then I think they could be. You just need to define a one-to-one map between the sequence of bit arrays to sensible sequences of states of Miami, which is guaranteed to exist (every bit array is unique and every state of Miami is unique). Extending this argument implies we are almost certainly in an accidental simulation.
- Then what determines what 'their time' is? Does the order of the pages matter? Why would their time need to correspond to our space?
This looks great. I'll check it out thanks!
This is great. I'd love to see more stuff like this.
Is anyone aware of articles like Chris Olah's views on AI Safety but for other prominent researchers?
Also, it would be great to see a breakdown of how the approaches of ARC, MIRI, Anthropic, Redwood, etc differ. Does this exist somewhere?
I agree that the claims are doing all of the work and that this is not a convincing argument for utilitarianism. I often hear arguments for moral philosophies that make a ton of implicit assumptions. I think that once you make them explicit and actually try to be rigorous the argument always seems less impressive, and less convincing.
Good point, I overlooked this.