-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Shashua Talk: Can AI do abstract thinking?
- Thread starter diplomat33
- Start date
DanCar
Active Member
Amnon says that LLMs lack the ability to abstract. Without this ability they will never be generally intelligent. Some new tech adding the ability to abstract could come tomorrow , 50 years, or later.
Yep. LLMs basically work by memorizing huge datasets, but they can't actually "think". They are really faking thinking by regurgitating what they memorized. So if we extrapolate that to an end-to-end approach like FSD V12, it is not intelligently figuring out what to do, it just memorized a lot of driving situations and is repeating what it memorized. And yes, with a big enough data set, it could memorize A LOT and probably drive great like 80% of the time. That is why we see some cool E2E self-driving demos. But I am skeptical this approach will achieve 99.99999% reliability needed for safe driverless. It just seems like it will be difficult to achieve 99.99999% simply by memorizing a lot of driving cases. That is why people raise the question of whether we need some kind of AGI breakthrough to achieve L5 because they argue that we will need AVs to do abstract thinking, like humans, to be able to handle unknown cases, without memorization.
shahryaran
Member
I think it's fair to say that LLMs have gone way further than anyone ever expected an auto-complete system to go, it's like having a generalized language simulator that you can direct using English, super cool stuff, they'll be diffusing down into different industries and fields for years/decades. But folks are also overly fixated on them right now, the big gains are obviously going to come from a MuZero or CICERO-style agent that can actually do planning.
Similar threads
- Article
