4 ms·

The really scary thing is all the value in the AI boom is predicated on the belief that the technology is "early" and that it will improve over time. We're seei

The really scary thing is all the value in the AI boom is predicated on the belief that the technology is "early" and that it will improve over time. We're seeing the opposite. Instead, all the competing models are basically converging on the same benchmark performance numbers, as we saw yesterday with the gpt-5 debacle. This suggests that performance is actually topping out, which makes intuitive sense if advancements in LLMs are proportional to their training data. They've already used up all the data. So it very well could be what we see right now is basically as good as it gets, or at least approximately so. The market is not ready for that.

Animats 1y ago
> We're seeing the opposite. Instead, all the competing models are basically converging on the same benchmark performance numbers... That question is really important. Comments on that? Maybe LLMs are asymptotically approaching some fundamental limit for that technology. What goes on inside an LLM hasn't changed much in years. More data and more compute is thrown at that little kernel that makes it all go. Additional gimmicks are added around the core algorithm. But there's not much progress down at the bottom. There's no reaching into the net and extracting a reliable "don't know". Is there an upper limit to what that architecture can do? AGI may not be reachable by this route. Maybe someone will find a way around that limit. The field has way too much money, too much visibility, and too many people. The technology works, although it has limits.
- jcgrillo 1y ago
  Money can't (necessarily) buy a breakthrough. It could take 10 or 100 or 1000 years, we just can't know. This kind of technical risk isn't usually so prominent. It's usually more diffuse with fallbacks, offramps, less palatable but workable alternatives.. This time it seems like the technical risk is looming very, very large.

Animats 1y ago

> We're seeing the opposite. Instead, all the competing models are basically converging on the same benchmark performance numbers... That question is really important. Comments on that? Maybe LLMs are asymptotically approaching some fundamental limit for that technology. What goes on inside an LLM hasn't changed much in years. More data and more compute is thrown at that little kernel that makes it all go. Additional gimmicks are added around the core algorithm. But there's not much progress down at the bottom. There's no reaching into the net and extracting a reliable "don't know". Is there an upper limit to what that architecture can do? AGI may not be reachable by this route. Maybe someone will find a way around that limit. The field has way too much money, too much visibility, and too many people. The technology works, although it has limits.

jcgrillo 1y ago
Money can't (necessarily) buy a breakthrough. It could take 10 or 100 or 1000 years, we just can't know. This kind of technical risk isn't usually so prominent. It's usually more diffuse with fallbacks, offramps, less palatable but workable alternatives.. This time it seems like the technical risk is looming very, very large.