Henry Yin ✈️ CoRL

World model" can mean a video that looks like Sora, a robot brain, or a set of abstract embeddings. These are completely different things.

Two AI research lineages quietly merged in the last two years to produce this confusion. New piece on what happened:

moe-capital.com/blog-home/the-…

Robotics finally has a scaling law that holds:

CoRL attendance

2024: 1,106
2025: 2,441
2026: 4,806

The JW Marriott wifi already broke. CoRL is scaling before robotics does.

holy triangle of workflow automation: goal + artifacts + rules

@ajambrosino can we please fix this? I’ve run into it so many times with very simple ask and it’s almost forcing me to use Claude

"That approval request is no longer active. Please retry the action if you still want ChatGPT to continue with it."

Anthropic’s S-1 was super fun to read. My takeaways and bets:

1. Serving Claude now earns an 85% gross margin, up from 12% a year ago. Prices keep falling, but costs fall faster.

2. But serving customers used only 27% of Anthropic's compute this year. The other 73% went to research and training. After that and everything else, Q2's operating margin was 8%. My bet: training keeps eating most of the compute, so profits stay thin through 2027.

3. Why so much training? Because customers buy the newest model. Over 80% of Q3 revenue came from models less than six months old.

4. Cheaper makes it bigger. When Opus 4.5 cut prices, quarterly growth jumped from 48% to 132%. My bet: Opus 5.5's 40% cut does it again, and Q4 growth beats Q3's 45%.

5. 5–7% of revenue this year came from customers later caught breaking the rules. Anthropic cracked down in August and September. My guess: that's part of why Q3 growth slowed.

6. Enterprise users who don't use Claude Code spend under $75 a month. The top 5% of Claude Code users spend about 20× that. My bet: that gap is Cowork's market. If office workers spend even a quarter of what heavy coders do, revenue per seat goes up 5×.

7. In April, Anthropic's own Claude use (at list prices) passed its cash payroll. My bet: most software companies cross that line by 2028.

8. Cyber is next. Last week, AI found bugs in PHP, the code behind most websites, and governments took sites offline to patch them. Claude's score on a hacking test went from 40% to 91% in four months. My bet: security is Anthropic's next Claude Code.

Two for fun:

9. Heads-up, SF house hunters: you have about six months. Anthropic staff can sell 10% of their shares on IPO day and the rest over the next six months. After that, every open house gets an all-cash offer written by Claude.

10. Forget benchmarks. The real AGI test is whether Claude works through Christmas. Last year, customer spend fell ~20% over the holidays as teams froze their code and went home. If this year's dip is smaller, the agents have stopped waiting for us.