OK, Chat GPT4 is now out.

It’s a incrementally smarter version of Chat-GPT3, shock horror. 😂

It is very capable at referencing information that is commonly referenced. eg law, book keeping, medicine, etc. It struggles with more esoteric niche things.

It has broad and deep understanding of mainstream science and technology as found across industry.

But it has no real understanding of patents other than popular patent history. It does not understand prior art, state of the art or the frontier.

It is not capable of innovation. It doesn’t even understand how to do innovation.

It is almost entirely retrospective.

This is largely down to the inescapable fact that AI requires data on which to train. For tasks and domains where lots of data exists, AI will surpass humans very quickly and will outcompete all workers on memory intense tasks.

Where AI seems to be very limited (just now at least) is at the fringe and the frontier of knowledge.

AI will achieve super human competence at many things that are already done today, but it is far from obvious to me that AI will leapfrog humans working at the state of the art.

It’s also difficult for me to imagine AI self innovating as a new vanguard beyond the frontier of human invention.

I can imagine how this might be achieved, but it would require AI to amass large datasets at the frontier and incrementally bridge off the state of the art into the abyss generating new data and retraining as it does so.

This is very different to how humans innovate. It would be vastly more laborious and difficult for AI than for a human inventor hypothesising new concepts sans data.