Best of your X follows: May 29
Today's digest: Anthropic closes a $65B Series H at a near-$1T valuation with $47B run-rate revenue, OpenAI launches Rosalind Biodefense for pandemic preparedness, Claude self-calibrates its own research rigor (Ethan Mollick), Naval on founders laying bricks in a skyscraper, and François Chollet cites Einstein on whether language even matters for invention.
Today's digest covers six threads worth your attention: Anthropic's near-trillion-dollar valuation and $47B run-rate, OpenAI launching an AI biodefense platform, Ethan Mollick's find that Claude self-calibrates its own research rigor, Naval on what founders are really building, and a debate on whether language even matters for invention.
Enterprise & Business
Anthropic closes $65B Series H at a $965B valuation
正在加载内容卡片…
Anthropic announced a $65 billion Series H round — led by Altimeter, Dragoneer, Greenoaks, and Sequoia — at a $965 billion post-money valuation. The company's run-rate revenue crossed $47 billion in early May 2026, up from $30B in April and $14B in February. Axios' Jim VandeHei said he couldn't find any company in any era that has scaled organic revenue this quickly at this level.
The round includes $15B of previously committed investment from hyperscalers and new strategic infrastructure partners Micron, Samsung, and SK hynix. Signed compute agreements span Amazon (up to 5 GW), Google/Broadcom (5 GW of TPU capacity), and SpaceX/Colossus. Funding will go toward safety and interpretability research, compute expansion, and enterprise product development.
1AI Safety & Biodefense
OpenAI launches Rosalind Biodefense for pandemic preparedness
正在加载内容卡片…
OpenAI announced Rosalind Biodefense — a new program to accelerate defensive progress in biology. It gives trusted builders access to GPT-Rosalind to develop biodefense and pandemic preparedness capabilities, with expanded access to select U.S. government and allied partners. OpenAI frames this as building a more robust ecosystem in which frontier AI helps defenders detect and respond to biological threats before they spread.
The announcement follows a pattern of frontier labs seeking government and national-security partnerships — Anthropic published its US-China AI leadership paper earlier this month, and the Gates Foundation deal with Anthropic was announced in the same window.
Model Releases & Research
Simon Willison's notes on Claude Opus 4.8 — and what "modest" means at the frontier
正在加载内容卡片…
Simon Willison published his follow-up analysis of Claude Opus 4.8 (released May 28), calling it "a modest but tangible improvement" over Opus 4.7. Key differences: sharper judgment on ambiguous tasks, reduced rate of missing code flaws (4× less likely), and a new Fast Mode for eligible enterprise accounts at 3× lower cost. Pricing otherwise unchanged.
To test the five different thinking-effort levels, Willison generated pelicans riding bicycles at each setting — an informal but surprisingly legible proxy for reasoning depth.
2"On a 1–10 identification scale, I'd now put the paper at about 4.5 — better than the 3.5 I'd have given before these tests, but well short of quasi-experimental (~7). The framing 'conditional association consistent with…' is still the right calibration."
Claude as economist: self-calibration in the wild
正在加载内容卡片…
Ethan Mollick (Wharton) posted a striking example of Claude roleplaying a research economist and then voluntarily downgrading its own paper's identification strength after robustness checks — on a 1–10 scale, from 3.5 to 4.5, with an explanation that stops well short of claiming causal inference. The model described its own work as "conditional association consistent with…" rather than a causal claim.
This is noteworthy because it runs against the sycophantic default most people expect from LLMs: Claude found a flaw in its own analysis and disclosed it without being asked. Whether this generalizes or was a one-off is an open empirical question.
AI Agents & Tools
OpenAI + CGR Teams: AI in professional motorsports
正在加载内容卡片…
Greg Brockman (OpenAI) highlighted the partnership with CGR Teams (Chip Ganassi Racing) — using OpenAI's models to analyze race data and improve performance decisions in real-time. It's a quiet but concrete example of agentic workflows entering high-stakes physical domains, where the feedback loop is fast and the value of marginal performance gains is high.
Founders & Mindset
Naval: you're laying bricks
正在加载内容卡片…
Naval's two best posts today landed on the same theme from different angles:
- "You aren't starting a company. You're laying bricks in the foundation of a skyscraper." (7.3K likes, 539 RTs) — A reframe of startup work as patient accumulation rather than event-driven milestones.
- "There are 10x engineers because there are 10x thinkers." (6K likes) — Locating the performance gap in cognition, not tooling.
Both sit against the current backdrop of AI-assisted engineering, where the throughput differential between fast and slow developers is widening rapidly.
Research & Thinking
Einstein didn't use language to think — does your AI system?
正在加载内容卡片…
François Chollet posted Einstein's well-known quote about thought mechanisms: "The words or the language, as they are written or spoken, do not seem to play any role in my mechanism of thought." Chollet's framing raises a question that runs through his ARC-AGI work: if genuine invention happens in non-linguistic substrate, and today's LLMs are fundamentally language-shaped predictors, what's the ceiling on what they can originate? It's a short post — but a useful corrective to the assumption that fluency equals intelligence.
围绕这条内容继续补充观点或上下文。