tara tan

Frontier research is being disrupted as fast as SaaS startups. Is the Eureka moment obsolete? Last week: OpenAI's general model disproved an 80-year-old geometry conjecture. Two papers landed in Nature on AI science agents, from FutureHouse and Google DeepMind. I initally assumed it was brute force, or a lot of human hand-holding. It's neither. The model's chain of thought is a search that diagnoses its own failures and lets each one point to the next move. Language is thinking, and thinking is language. Maybe frontier research becomes the craft of reverse engineering the discovery that AI makes. Read the full below: open.substack.com/pub/strangerev…

→ The US Commerce Department took equity in nine quantum companies for $2B. Washington is reusing the same ownership-plus-support template it applied to rare earths to quantum, calling it a “critical frontier technology”. Read the full brief in The Strange Review, plus an op-ed on "Is the Eureka Moment obsolete?" open.substack.com/pub/strangerev…

→ Google made I/O 2026 about agents everywhere, spanning consumer, developer, and science tools. It is leaning hard on the one thing rivals can't copy: distribution across several billion users.

Things worth paying attention to this week: → Karpathy joined Anthropic to make models that improve how models are trained. This puts one of the field's best-known researchers on the bet that the next gains come from recursive training, rather than from raw compute alone.

there are 75+ neo labs and counting today nvidia was one of 90+ graphics cards startups in the 90s. do with that information what you will

@evans1vn Gochu twin

@taratan Ty for this fam 🫶

Things to pay attention to this week: → H200s approved for China, but Beijing declines. They are going all-in on building out their domestic chip stack, investing billions in local foundries.

Read the full brief in The Strange Review. open.substack.com/pub/strangerev…

→ xAI open-sources the X recommendation algorithm. Ships with a runnable inference pipeline, 3GB model checkpoint, and the ads blending system.

Anecdotal but the models seem to be responding much faster and better on weekends. i wonder if we might start seeing startups or research labs shuffling around definitions of "working hours" to get better compute

we're looking for a cracked growth engineer. forms.gle/XZjzhdDA3Hi8CU…

MagicPath is now fully agentic!