oh my god san francisco, stop pretending boomer gilded age food (prime rib) is good in any way

Any startup interested in HPC engineer who worked on 2 open architectures and has experience debugging model behavior? Almost 3 but I turned down the job for [redacted] reason. Also llama.cpp maintainer / infra runner on Gemini protocol. I need to find myself a new job.

Planetary VC founder @ramez reveals that throwing 100 Claude Opus 5.5 agents at a problem was no better than using just 10, exposing a major limitation of AI swarms: "Test time compute scales worse than pre-training. Agents scale worse than test time compute. On the Opus 5.5 model card, it is like the worst scaling we've ever seen of anything." "The success of one agent was like 70%. It gained three points when you went to four agents, and then it gained two points when you went from four agents to 16. The Opus 5.5 model just topped out after 10 agents. It didn't get any better with 100 agents." "If you took the population of Columbus, Ohio, and said, 'Create general relativity,' would they do as well as Einstein on his own? There are many categories of problems where the peak intelligence of one entity is much more important than having a lot of entities at a lower level collaborating."

Day 3 (encore)/ We silently re-shipped codex cloud. It’s pretty good now

For those who missed it yesterday, I wrote an explainer on diffusion policies in robot learning, as I'd have wanted them taught to me. It covers: - Why VAEs can still blend demos, and how hierarchical VAEs can help - Diffusion as a special case of a hierarchical VAE, and why it works - How Diffusion Policy brings this to policy learning, and my results reimplementing and tweaking it (video) - How DDIM takes fewer steps, and how flow matching straightens its path into a simpler recipe (and reframes diffusion) Now on to sim-to-real RL with a physical SO-101. Will be sharing that here as I go.