[godelski]

It is CRITICAL that we be realistic about what fulfills the optimization objectives in the models that we train. I think there's been a significant unwillingness that objectives like "human preference" (RLHF, DPO, etc) not only help models become more accurate and sound more natural in speech, BUT ALSO optimize the models to be deceptive and convincing when they are wrong. It's easy to see, because you know what's more preferential than a lie? A lie that you don't know is a lie. You (may) prefer the truth, but if you cannot differentiate the truth from a lie you'll preference based on some other criteria. We all know that lies frequently win out here. If you doubt this, just turn on the news or talk to someone that belongs to the opposite political party of yourself.

This creates a very poorly designed tool! A good tool should fail as loudly as possible, in that it alerts the user of the failure and does its best to specify the conditions that led to this. This isn't always possible, but if you look at physical engineers you'll see that this is where they spend a significant portion of their time. Even in software I'd argue we do a lot here, but also that it is easy to brush off (we all love those compiler messages... right?). Clearly right now LLMs are in a state where we don't know how to make their failures more visible, and honestly, that is okay. But what is not okay is to pretend that this is not current reality and pretend that there are no dangers or consequences that this presents. We dismiss this because we catch some obvious errors and over-generalize the error quality, but that just means we suffer from Murray Gell-Mann Amnesia. It's REALLY hard to measure what you don't know. Importantly, we can't even begin to resolve these issues and build the tools we want (the ones we pretend these are!) if we ignore the reality of what we have. You cannot make things better if you are unwilling to recognize their limitations.

Everyone here is an engineer, researcher, or builder. This framework of thinking should be natural to us! We should also be able to understand that there's a huge difference between critiques and limitations and dismissing things. I'm an AI critic, but also very optimistic. I'm a researcher and spending my life working on this topic. It'd be insane to do such a thing if I thought it was a fruitless or evil effort. But it would be equally insane to pursue a topic with pure optimism. If I were to blind myself to limits and paint everything as a trivial to solve problem, I'd never be able to solve any of those problems. Ignoring or dismissing technical issues and limitations is the domain of the MBA managers, not engineers.