AI digest: Reality check week
Big infrastructure moves, massive valuations, and benchmarks showing AI still can't think about physics properly.
The gap between AI hype and reality stayed wide this week, even as the money kept flowing.
Anthropic overtakes OpenAI with $900bn valuation
Anthropic raised another $30bn just three months after their last mega-round, pushing their valuation to $900bn and past OpenAI for the first time. Revenue hit $45bn annualised, a fivefold jump since their last funding. The speed of these numbers feels unsustainable, but Claude’s clearly winning enterprise customers.
Video models look brilliant but think like toddlers
New benchmarks confirm what we suspected: AI video generators are stunning but can’t reason about basic physics. WorldReasonBench tests logical plausibility rather than visual quality, and models fail at things like understanding gravity. Meanwhile, NVIDIA’s SANA-WM generates minute-long 720p videos on a single GPU, which is genuinely impressive progress for open models.
Infrastructure gets serious about production agents
LiteLLM launched their Agent Platform, a Kubernetes-based system for running AI agents in production with proper isolation and session management. This tackles the real problem: getting agents from demo scripts to enterprise deployment. Plus one founder is spending $1.3m monthly on OpenAI APIs to run 100 coding agents, which shows someone’s taking the agent future seriously.
Attention gets faster for long contexts
Nous Research’s Lighthouse Attention delivers 1.4-1.7x speedup for long context training by pooling queries, keys and values across a multi-resolution pyramid. The clever bit is it’s training-only, so you can remove it afterward and still get the benefits.