Trying to Make It to Space in a Propeller Plane

A bigger propeller doesn’t get you orbit.

“From what I know, LLM is just training data at a large scale. It’s like how many pictures of cats do you have to show the computer before it knows what a cat looks like.”

That’s the whole thing, right there. A large language model is a machine that has seen so many pictures of cats — so many words about cats, so many everything — that it can produce a convincing cat on demand. Nobody disputes that it’s impressive. The question nobody in charge wants to answer is whether showing it more pictures gets you anything other than a better cat.

The best way to understand the problem is the idea from Benedict Evans, when he said we’re “trying to make it to space in a propeller plane.” That’s his analogy, not mine, and it’s perfect: the propeller plane is a real achievement. It flies. It just doesn’t go to space, and a bigger propeller doesn’t change that. More training data is a bigger propeller. More compute is a bigger propeller. You can admire the engineering and still notice the destination isn’t getting closer.

“The people who make AI are trying to make AGI, but the question is what happens if we get more advanced machine learning that can listen better and solve frequent problems.”

That’s the honest version of the future, and notice how different it sounds from the sales pitch. Not artificial general intelligence — not a mind in a box. Just better machine learning: software that listens better, that handles the frequent…