Clément

We all agree that we need to arrive at a consensus on a number of questions. I agree with @geoffreyhinton that LLM have *some* level of understanding and that it is misleading to say they are "just statistics." However, their understanding of the world is very superficial, in large part because they are trained purely on text. Systems that would learn how the world works from vision would have a much deeper understanding of reality. Second, auto-regressive LLM have very limited reasoning and planning abilities. I do not believe we can get anywhere close to human-level AI (even cat-level AI) without (1) learning world models from sensory inputs like video, (2) an architecture that can reason and plan (not just auto-regress). Now, if we have architectures that can plan, they will be *objective driven*: their planning will work by optimizing a set of objectives at inference time (not just training time). These objectives can include guardrails that will make those system safe and subservient *even* if they end up having much better world models that humans. Then, the problem becomes to design (or train) good objectives functions that will guarantee safety and efficiency. It's a hard engineering problem, but not as hard as some have made it to be.