[blamestross]

> If AI stays on the trajectory that we think it will, then amazing things will be possible. Maybe with 10 gigawatts of compute, AI can figure out how to cure cancer.

At least the statement starts with a conditional, even if it is a silly one.

If you know your growth curve is ultimately going to be a sigmoid, fitting a model with only data points before the inflection point is underdetermined.

> If AI stays on the trajectory that we think it will

Is a statement that no amount of prior evidence can support.

[jjk166]

Assuming AI stays on the trajectory they think it will doesn't mean they assume it will be infinite exponential growth. If you know for a fact it's sigmoid curve, presumably the path you think it will continue on is the sigmoid curve. The trillion dollar question is does performance plateau before or after AI can do the really exciting stuff, and while I may not agree with it myself the more optimistic position is not an unreasonable belief.

Also you can most certainly fit a sigmoid function only from past data points. Any projection will obviously have error, but your error at any given point should be smaller than for an exponential function with the same sampling.

[jennyholzer]

I think if you're critiquing AI, you should use harsher words.

AI boosters are going to spam the replies to your comment in attempts to muddy the waters.

[CuriouslyC]

I'm an AI booster, but he's right, these models are in the sigmoid elbow and we're being hard-headed trying to push frontier models, it's not sustainable. We need to take a step back and work on the engineering of the systems around the frontier models while trying to find a new architecture that scales better.

That being said the current models are transformative on their own, once the systems catch up to the models that will be glaringly obvious to everyone.