TL;DR * On 18 real tasks from two large open-source repositories, early-2025 AI agents often implement functionally correct code that cannot be easily used as-is, because of issues with test coverage, formatting/linting, or general code quality. * This suggests that automatic scoring used by many benchmarks may overestimate AI agent...
TL;DR We designed an ambitious scalable oversight experimental setup, where we had people with no coding/programming experience try to answer coding questions (“Which of these two outputs is the correct output of the given function on this input?”), using LLMs that debate or are arguing for the correct/incorrect answer 50%...
Hmm, I actually kind of lean towards it being rational, and labs just underspending on labor vs. capital for contigent historical/cultural reasons. I do think a lot of the talent juice is in "banal" progress like efficiently running lots of experiments, and iterating on existing ideas straightforwardly (as opposed to something like "only a few people have the deep brilliance/insight to make progress"), but that doesn't change the upshot IMO.
Salaries have indeed now gotten pretty high—it seems like they're within an OOM of compute spend (at least at Meta).
That's indeed what I meant!
We designed an ambitious scalable oversight experimental setup, where we had people with no coding/programming experience try to answer coding questions (“Which of these two outputs is the correct output of the given function on this input?”), using LLMs that debate or are arguing for the correct/incorrect answer 50% of the time. The goal is to simulate the case when we have significantly superhuman AI systems that deeply understand things even human experts don’t.
We ran two pilots with human non-coder judges, but we’re dropping this project in its current form, chiefly because we weren’t able to coax the RLHF/RLAIF’d models like GPT-4T and Claude 3 Opus to argue competently enough for incorrect... (read 2742 more words →)
the existence of predicates on the world that are easier to evaluate than generate examples of (in the same way that verifying the answer to a problem in NP can be easier than generating it) guarantees that the model should be better at distinguishing between evaluation and deployment than any evaluator can be at tricking it into thinking it's in deployment
Where does the guarantee come from? Why do we know that for this specific problem (generating vs. evaluating whether the model is deployed) it's easier to evaluate? For many problems it's equally difficult, right?
Given that the judge that selects the best argument for BoN is the same as the one that chooses the winner, what is your main takeaway from the fact that ELO increases as you increase N? I see this as mainly a sanity check, but want to check if I'm missing something.
Another author here! Regarding specifically the 74% vs. 84% numbers - a key takeaway that our error analysis is intended to communicate is that we think a large fraction of the errors judges made in debates were pretty easily solvable with more careful judges, whereas this didn't feel like it was the case with consultancy.
For example, Julian and I both had 100% accuracy as judges on human debates for the 36 human debates we judged, which was ~20% of all correct human debate judgments. So I'd guess that more careful judges overall could increase debate accuracy to at least 90%, maybe higher, although at that point we start hitting measurement limits from the questions themselves being noisy.
The issue with early finetuning is that there’s not much that humans can actually select on, because the models aren’t capable enough - it’s really hard for me to say that one string of gibberish is better/worse.
I think the issue with the more general “neocortex prosthesis” is that if AI safety/alignment researchers make this and start using it, every other AI capabilities person will also start using it.
- unreasonable ^5
I think there's a typographical error - this doesn't link to any footnote for me, and there doesn't appear to be a fifth footnote at the end of the post
Has anyone thought about the best ways of intentionally inducing the most likely/worst kinds of misalignment in models, so we can test out alignment strategies on them? I think red teaming kinda fits this, but that’s more focused on eliciting bad behavior, instead of causing a more general misalignment. I’m thinking about something along the lines of “train with RLHF so the model reliably/robustly does bad things, and then we can try to fix that and make the model good/non-harmful”, especially in the sandwiching context where the model is more capable than the overseer.
This is especially relevant for Debate, where we currently do self-play with a helpful assistant-style RLHF'd model, where one of the models is prompted to argue for an incorrect answer. But prompting the model to argue for an incorrect answer is a very simple/rough way of inducing misalignment, which is (at least partially) what we're trying to design Debate to be robust against.
Is there any other reason to think that scalable oversight is possible at all in principle, other the standard complexity theory analogy? I feel like this is forming the basis of a lot of our (and other’s) work in safety, but I haven’t seen work that tries to understand/conceptualize this analogy concretely.
Is anyone thinking about how to scale up human feedback collection by several orders of magnitudes? A lot of alignment proposals aren’t focused on the social choice theory questions, which I’m okay with, but I’m worried that there may be large constant factors in the scalability of human feedback strategies like amplification/debate, such that there could be big differences between collecting 50k trajectories versus say 50-500M. Obviously cost/logistics are a giant bottleneck here, but I’m wondering about what other big challenges might be (especially if we could make intellectual progress on this before we may need to)
How does shard theory differ from the Olah-style interpretability agenda? Why is there any reason to believe we can learn about "shards" without interpretability?
When doing sandwiching experiments, a key property of your "amplification strategy" (i.e. the method you use to help the human complete the task) should only help the person complete the task correctly.
For example, lets say you have a language model give arguments for why a certain answer to a question is correct. This is fine, but we don't want it to be the case that the system is also capable of convincing the person of an incorrect answer. In this example, you can easily evaluate this, by prompting or finetuning the model to argue for incorrect answers, and seeing if people also believe the incorrect arguments. This description of the problem/task naturally... (read more)
If people read at 250 words/minute, and a page of text has 500 words, and we could get 10M people to read AI outputs full-time, we could have human-evaluation of about 100B pages of text per year, which actually feels surprisingly high
Has anyone tried debate experiments where the judge can interject/ask questions throughout the debate?
It’s so easy to get caught up in meta-thinking - I want to try to remember to not spend more than maybe 10% of my time generally doing meta-reflection, process optimization, etc., and spend at least 90% of my time working directly on the concrete goal in front of me (LM alignment research, right now)
Epistemic status: I’m somewhat confident this is a useful axis to describe/consider alignment strategies/perspectives, but I’m pretty uncertain which is better. I could be missing important considerations, or weighing the considerations listed inaccurately.
When thinking about technical alignment strategy, I’m unsure whether it’s better to focus on trying to align systems that already exist (and are misaligned), or to focus on procedures that train aligned models from the start.
The first case is harder, which means focusing on it is closer to minimax optimization (minimize the badness of the worst-case outcome), which is often a good metric to optimize. Plus, it could be argued that it’s more realistic, because the current AI paradigm of... (read more)
'The goal of alignment research should be to get us into "alignment escape velocity", which is where the rate of alignment progress (which will largely come from AI as we progress) is fast enough to prevent doom for enough time to buy even more time.'
^ the above argument only works if you think that there will be a relatively slow takeoff. If there is a fast takeoff, the only way to buy more time is to delay that takeoff, because alignment won't scale as quickly as capabilities under a period of significant and rapid recursive self-improvement.
Important to caveat that these results are pretty small—I wouldn't take the absolute numbers too seriously beyond the general "algorithmic scoring may often overestimate software capabilities".