With that soundtrack it feels like it could be a scene from the animatrix.

AI tornado!

Ideas: AI hoover, AI funnel, AI suck up (vs trickle down?).

Where can one read up about these new insights?

His example of the DDP container was pretty instructive. He calls it slop because he doesn't want it. Of course for many people it's nice that the LLM would handle porting to a distributed version for them. One person's slop is another's flop I guess.

Are you sure you and Karpathy mean the same thing by slop? Feels like not.

I get the sense that that is a research direction borne out of necessity. You can't get enough data for one platform, so you average data and hope you get something that works for all of them at the end of it.

Got a link to a description of/paper on those issues?

mlx, made easy by llm (Simon Willison)

Thanks!

Can you say more on (or link) why diffusion working is related to low-D of image manifolds?

Russ had the most convincing perspective, hands down.

What do you use for fast envs? Isaac Sim, MJX, other?

Yes! I forgot about that route

I did too, partly due to a sizable volume of past text files, and an embarrassing amount of tikz. BUT you can drop in equations from tex files as-is using mitex. Not sure what to do about the tikz stuff yet. But compilation in typst is so fast that it's worth migrating. And vim speeds up conversion.