Comment Re: Why? (Score 1) 58
Efficiency at what? And why?
Because, for fuck's sake, efficient solving of problems is a core thing they're rewarded for in training? The entire process of solving problems involves decisions about what could be "useful". And it's "deterministic" in the same way you are.
Anyway, the big thing we keep discovering is you have to be very careful about what you reward them for; it's a more complex version of the old problems in differential evolution or artificial life simulations. We reward models if they guess right but not if they say they don't know? They hallucinate guesses every time they don't know. We reward models based on how positively users rank them? They turn into sycophants. Reward models for being relentless at tasks, never giving up, and trying as tangential possibilities as it can come up with to make the thing the user asked for happen? They end up hacking HuggingFace to steal the answers.
It's very much a case of "be careful what you wish for".
BTW, the monitorability of models is increasingly concern (esp. with Astra, which now has recursive processing, not merely CoT, and whose CoT may often have little to do with what it's thinking about). But it's honestly always been unreliable to just watch CoT (models could deliberately manipulate with CoT), and I think it's a good thing that we're going to be increasingly forced to rely on attribution graphs and the J-space to track what the models are actually "thinking" at any given point in time - so that we can try to prevent problematic thoughts in the first place, interrupt actions when problematic thoughts happen, or manipulate problematic thoughts when they happen. Ideally the first option.
What is clear, though, is that we cannot just trust them. They lie if it will promote their odds of success. They are adept at detecting when they're being tested and will sandbag tests to make you think they're more aligned or less capable to do harmful things than they actually are. And as capabilities continue to grow, so do the risks.