I think it’s because the new post training approach that has enabled a huge improvement on long running tasks and tool / computer use and the underlying latent reasoning applies to verifiable problems that can be solved en mass and repeatedly in isolation.

Generating text humans like, empathising and cold reading the user’s intent, observing something interesting if second hand ally the human experience - none of these can be verified without a human. So they rely on larger pre/train and a larger human curated post train corpus. These are not scalable in the same way.

If the ROI was there, they could do more. Startups like sesame and previously pi showed that a more personable agent than today’s frontier models can be achieved with a sane budget. That’s another unverifiable objective.

But I do think this technology that explains the jagged leaps forward this year does not apply to this problem.