I think the exponential view is largely a mistake of looking at “quantitative to qualitative” changes and mistaking the qualitative for something that is common and inevitable, rather than something that requires more quantitative buildup again before it can reach a qualitative change again.

Attention Is All You Need, the paper that could be pointed at as the qualitative turning point for AI transforming from niche tinkering to mass marketing bubble buildout, laid out the transformer architecture that people are still exploring and testing the limitations of; the architecture that powers generative AI in general. As far as I’m aware, there is not yet any alternative/better architecture created yet. There may be theorized experiments, but nothing practical. It’s possible the research will hit too many walls too consistently with the current base architecture and require another equivalent breakthrough to go much further than it has, and if that’s the case, it’s uncertain whether such a breakthrough can or will occur within our lifetimes.

I’m blanking on the specifics, but I want to say there is some stuff in computing, perhaps it was to do with AI, that was actually theorized entire decades ago and only became more viable to make real because of advances in hardware. I mention this to emphasize that these things are not always as rapid as they may seem, nor as inevitable. But the capitalists need the investor class to believe that improvement is continuous and guaranteed to keep the money rolling in.

This is not to argue that AI will stagnate, but more to be sober about the reality that it very well may. Like you say, it’s hard to tell where the plateau is.