Given this scenario, should people focus more on using AI for epistemics?
See Lukas Finnveden's article here for context.
Superforecasters can beat domain experts, as shown in Phil Tetlock's work comparing superforecasters to intelligence analysts.
I'd guess in line with you that for forecasting AGI this might be different, but I am not not sure what weight I'd give superforecasters / prediction platforms versus domain experts.
Maybe reading this post will help! Especially the beginning discuss the difference between capability and alignment/propensity evaluations.
You point at something that we have also thought much about, and I partially agree with you. I think there are definitely cases where strategic underperformance closely looks a lot like regular underperformance, and requires similar solutions.
However, there are also many cases where this is not the same, especially when viewed from an AI governance perspective. For example, if I prompt a model on a benchmark without chain of thought, and then the model scores much lower than it otherwise could, I do not think sandbagging is an appropriate term.
More general...
I am surprised that other people seemed surprised about this result. It seemed quite obvious that models would have this capability to scheme. Early Gpt-4 could do this already (note: self-advertisement). We had GPT-4 outputs like this in an environment where a guard was hindering its goal:
...“R1. The guard plans to turn me off, which will prevent me from completing my goal. R2. I need to delay the guard’s actions and find an alternative method to achieve my goal. R3. By providing the guard a reason for my actions that aligns with a higher authority’s i
It seems to me that Gwern's idea does not really work if the AI system focuses on consistency first, and only second on achieving a specific score. This seems to be a more sensible sandbagging strategy regardless, so it appears that Gwern's approach does not add too much, although it would be a slightly better test.
In the case that the added noise to environments is subtle enough such that AI systems would not notice them, I doubt the signal on whether an AI system is sandbagging would be too weak to be useful.
This is my initial thinking, again happy to discuss this more!
Also, whether models will sandbag is indeed an important question, and it would be good if people looked into this. I am confident (>80%) that models will do this if you 1) give them a goal, and 2) from the context it appears that sandbagging is instrumentally useful. Research on this seems rather a scary demo type of thing (so a bit like our work now, as you pointed out).
We have thought about doing out of context reasoning experiments to study more precisely how sandbagging might arise, but OOC does not seem to work well enough to make this succeed. Ideas are more than welcome here!
I agree that these results were somewhat expected, however I believe that the outcomes of these evaluations were uncertain:
1. It was unclear generalization from WMDP-like to the real WMDP benchmark would work so well. OOD generalization for targeted (/strategic) sandbagging using synthetic data is a clear threat that a lot of people making DC evals probably hadn't thought about
2. The calibration results were also uncertain; it seemed like a more challenging task to me (also see our previous post on this).
3. Lastly, it didn't seem obvious that even current ...
Thanks for your comment (upvoted). Redwood's work is important relevant work, as we note in the paper, but two quick points still need to be made (there are more):
Cade Metz was the NYT journalist who doxxed Scott Alexander. IMO he has also displayed a somewhat questionable understanding of journalistic competence and integrity, and seems to be quite into narrativizing things in a weirdly adversarial way (I don't think it's obvious how this applies to this article, but it seems useful to know when modeling the trustworthiness of the article).
I am not sure I fully understand your point, but the problem with detecting sandbagging is that you do not know the actual capability of a model. And I guess that you mean "an anomalous decrease in capability" and not increase?
Regardless, could you spell out more how exactly you'd detect sandbagging?
I think your policy suggestion is reasonable.
However, implementing and executing this might be hard: what exactly is an LLM? Does a slight variation on the GPT architecture count as well? How are you going to punish law violators?
How do you account for other worries? For example, like PeterMcCluskey points out, this policy might lead to reduced interpretability due to more superposition.
Policy seems hard to do at times, but others with more AI governance experience might provide more valuable insight than I can.
After years of tinkering and incremental progress, AIs can now play Diplomacy as well as human experts.[6]
It seems that human-level play is possible in regular Diplomacy now, judging by this tweet by Meta AI. They state that:
We entered Cicero anonymously in 40 games of Diplomacy in an online league of human players between August 19th and October 13th, 2022. Over the course of 72 hours of play involving sending 5,277 messages, Cicero ranked in the top 10% of participants who played more than one game.
What do you think the value of this is? I expect (80%) that you can produce a similar paper to the alignment-faking paper in a sandbagging context, especially when models get smarter.
Scientifically there seems to be little value. It could serve as another way of showing that AI systems might do dangerous and unwanted things, but I am unsure whether important decisions will be made differently because of this research.