[opportune]

I don’t think there’s anything making it impossible for actual intelligence to arise from a task as simple as “predicting the next token (to model human thought/speech/writing)” because with enough compute resources, smart AI implementations, and training that task basically would be optimized by becoming a general intelligence.

But it’s clear based on current implementations that once you work backwards from the knowledge that it’s “just predicting the next token” you can easily find situations in which the AI doesn’t demonstrate general intelligence. This is most obvious when it comes to math, but it’s also apparent in hallucinations and the model not being able to reason through/synthesize ideas very well, deviate from the script (instead of just answering a question with what it has already, in some cases it should not even try to answer and instead ask more clarifying questions). To be fair, there are plenty of humans with excellent writing or speaking skills that are bad at that kind of stuff too.

[simonh]

The problem is such an approach is limited by the content of the training texts. As I mentioned elsewhere, our written texts assume huge swathes of contextual and experiential information and knowledge that LLMs don’t have. It’s possible some of it might be inferred from the texts, but not all of it by a long shot.

If somehow you could generate a training text encoding a complete and thorough understanding of the physical world, human psychology and sociology, and reasoning then that might get you quite far. But the existing it even near future human textual corpus isn’t really that. Even then I still think you’d hit the limitations of the LLM cognitive architecture pretty hard.

[giardini]

simonh says >"The problem is such an approach is limited by the content of the training texts."<

Aside: I would like to see ChatGPTs with distinct training texts, e.g., a ChatGPTs trained on the "great books" of Western philosophy and science knowledge up to the time of Victorian England.