We normals are hearing in the media that the large language model approach to AI has run its course, that further scaling it up is prohibitive in terms of energy, and that there is an AI-hype-driven financial bubble ready to pop any moment. According to at least one blogger though, the big breakthrough happening right now is having these models “reason” internally before they give an answer.

Two of those leading engineers are: Julian Schrittwieser who helped teach AlphaGo how to play Go at a level never witnessed in human history and is now a lead researcher at Anthropic. And Łukasz Kaiser, who whilst at Google Brain, co‑authored the paper that launched the architecture now driving every major released model on “Attention is all you need”…

Kaiser, for his part, corrects time horizons. The category of work that still belongs unquestioned to humans is shrinking. He states, with a deep belief, that these AI systems will be able to do any labor task currently performed on a computer within a timeframe of five years! …

The question is not whether machines will pass some imagined threshold in the future, but what it means that they have already crossed thresholds we still debate as hypothetical. A society reacts to what it believes is true, not to what is true. When the prevailing public understanding is delayed by years, institutions are, by definition, operating in a prior decade.

We can model technological progress as a series of sequential, overlaid S-curves that have to overlap in just such a way to produce continuous exponential growth. At least some insiders are still thinking in terms of keeping this S-curve going, in a competition between companies and countries. And when we see a new technology break through into widespread public, commercial use, it has already been going in the lab for awhile. That used to be measured in decades, now it is months if these optimist insider voices are to be believed.