Frustratingly, this was all normal stuff five or six years ago: build models with the skills of a cat or two year old (ie, non-linguistic). It motivated the NetHack challenge and research into Minecraft, with the hope it'd build safer cars. LLMs killed a lot of perceptual reasoning research.

He's working, same as a lot of cog sci inspired folks, on perceptual learning and reasoning: tracking objects, predicting casual patterns, and forming counterfactual supporting models. The recent stuff tests models on infant psychology tasks: does it have object permanence?

A lot of folks don't care about desert as much as using punishment to express social condemnation of the crime and deter future wrongs. In which case, both can tolerate wrongful deaths for broadly utilitarian reasons (which is a good reason to dislike utilitarian theories of punishment).

“The deskilling, denigration, and displacement of teachers and scholars have historically been central to fascist takeovers, since educators serve as bulwarks against propaganda, anti-intellectualism, and illiteracy.” — @olivia.science

www.project-syndicate.org/commentary/a...

A life well-lived--except for a foolish, multi-decade neck-beard that just didn't work.