It isn’t clear, but researchers are finding ways to overcome old LLM limitations.

People used to say: “LLMs only think in a single pass and can’t perform long or recursive chains of reasoning.” Now LLMs are being used to produce long chains of reasoning that the models then review and extend. This has enhanced the abilities of modern AIs.

AI is a moving target. The researchers in this field can see the obstacles, and they’re doing their best to surmount them.

Other approaches may achieve superintelligence soon, even if LLMs don’t.

The field is good at overcoming obstacles, including coming up with new AI architectures and approaches. We didn’t write If Anyone Builds It, Everyone Dies to warn people about LLMs in particular. We wrote it to warn people about superintelligence.

The reason we talk about LLMs isn’t that we’re sure LLMs are the shortest path from here to superintelligence. We talk about LLMs because they represent the AI approach that is currently working, and because studying them is a fine way to understand just how little anyone knows about these new minds humanity is growing.

See also the extended discussion on why gradient descent matters.