Abstract Scalable oversight protocols aim to enable humans to accurately supervise superhuman AI. In this paper we study debate, where two AI's compete to convince a human judge; consultancy, where a single AI tries to convince a human judge that asks questions; and compare to a baseline of direct question-answering,...
Power-seeking is a major source of risk from advanced AI and a key element of most threat models in alignment. Some theoretical results show that most reward functions incentivize reinforcement learning agents to take power-seeking actions. This is concerning, but does not immediately imply that the agents we train will...
Scalable oversight protocols aim to enable humans to accurately supervise superhuman AI. In this paper we study debate, where two AI's compete to convince a human judge; consultancy, where a single AI tries to convince a human judge that asks questions; and compare to a baseline of direct question-answering, where the human judge just answers outright without the AI. We use large language models (LLMs) as both AI agents and as stand-ins for human judges, taking the judge models to be weaker than agent models. We benchmark on a diverse range of asymmetries between judges and agents, extending previous work on a single extractive QA task with information asymmetry, to also include... (read 1890 more words →)
Power-seeking is a major source of risk from advanced AI and a key element of most threat models in alignment. Some theoretical results show that most reward functions incentivize reinforcement learning agents to take power-seeking actions. This is concerning, but does not immediately imply that the agents we train will seek power, since the goals they learn are not chosen at random from the set of all possible rewards, but are shaped by the training process to reflect our preferences. In this work, we investigate how the training process affects power-seeking incentives and show that they are still likely to hold for trained agents under some assumptions (e.g. that the agent learns... (read 2430 more words →)
Is there a reason to think this problem is less amenable to being solved by complexity priors than other learning problems? / Might we build an unaligned agent competent enough to be problematic without solving problems similar to this one?
What is Mathematics? by Courant and Robbins is a classic exploration that goes reasonably deep into most areas of math.
This makes me think of two very different things.
One is informational containment, ie how to run an AGI in a simulated environment that reveals nothing about the system it's simulated on; this is a technical challenge, and if interpreted very strictly (via algorithmic complexity arguments about how improbable our universe is likely to be in something like a Solomonoff prior), is very constraining.
The other is futurological simulation; here I think the notion of simulation is pointing at a tool, but the idea of using this tool is a very small part of the approach relative to formulating a model with the right sort of moving parts. The latter has been tried with various simple models (eg the thing in Ch 4); more work can be done, but justifying the models&priors will be difficult.
Certainly, interventions may be available, just as for anything else; but it's not fundamentally more accessible or malleable than other things.
I'm arguing that the fuzzy-ish definition that corresponds to our everyday experience/usage is better than the crisp one that doesn't.
Re IQ and "way of thinking", I'm arguing they both affect each other, but neither is entirely under conscious control, so it's a bit of a moot point.
Apropos the original point, under my usual circumstances (not malnourished, hanging out with smart people, reading and thinking about engaging, complex things that can be analyzed and have reasonable success measures, etc), my IQ is mostly not under my control. (Perhaps if I was more focused on measurements, nootropics, and getting enough sleep, I could increase my IQ a bit; but not very much, I think.) YMMV.
I think what you're saying is that if we want a coherent, nontrivial definition of "under our control" then the most natural one is "everything that depends on the neural signals from your brain". But this definition, while relatively clean from the outside, doesn't correspond to what we ordinarily mean; for example, if you have a mental illness, this would suggest that "stop having that illness!!" is reasonable advice, because your illness is "under your control".
I don't know enough neuroscience to give this a physical backing, but there are certain conscious decisions or mental moves that feel like they're very much under my control, and I'd say the things under my control... (read more)
March 2nd isn't a Tuesday; is it Monday night or Tuesday night?
If you want to discuss the nature of reality using a similar lexicon to what philosophers use, I recommend consulting the Stanford Encyclopedia of Philosophy: http://plato.stanford.edu/
Musk has joined the advisory board of FLI and CSER, which are younger sibling orgs of FHI and MIRI. He's aware of the AI xrisk community.
Cool. Regarding bounded utility functions, I didn't mean you personally, I meant the generic you; as you can see elsewhere in the thread, some people do find it rather strange to think of modelling what you actually want as a bounded utility function.
This is where I thought you were missing the point:
Or you might say it's a suboptimal outcome because you just know that this allocation is bad, or something. Which amounts to saying that actually you know what the utility function should be and it isn't the one the analysis assumes.
Sometimes we (seem to) have stronger intuitions about allocations than about the utility function itself, and parlaying that to identify what... (read more)