Five X signals: AI scientists, high-quality writing, and the limits of intelligence

Business and enterprise

1. Anthropic's disclosed revenue figure meets a newer leaked estimate

• What changed: Simon Willison reported that Anthropic's latest leaked run-rate figure was $65 billion for July. His linked analysis records Anthropic's own Series H announcement as saying that run-rate revenue had crossed $47 billion earlier in May. 12
• Why it matters: The two figures answer different questions: $47 billion is a company-disclosed figure, while $65 billion is a leaked estimate discussed by Simon. Run-rate revenue is an annualized projection of current revenue, so it is a way to compare momentum rather than a report of money already earned. 3
• Evidence boundary: The $65 billion figure remains a leak as described in the post. The $47 billion figure comes from Anthropic's announcement and is quoted in Simon's analysis. Audited annual revenue remains a separate measure.

Research

2. Mollick sees promise and gaps in Google's AI scientist work

• What changed: Ethan Mollick quoted a Google research scientist's post about an extension of Gemini's Co-Scientist, which is described as working with scientists in materials science, biology, and computer science. Mollick said the paper showed both promise and gaps, then asked how much more advanced models might close those gaps. 45
• Why it matters: The useful question is narrower than whether an AI scientist exists. A reader can inspect how much of the collaboration, experiment design, and evaluation the model handled before treating the project as a step toward autonomous research.
• Evidence boundary: Mollick's post gives a qualitative assessment and no performance numbers. The post gives a reason to read the paper; independent research autonomy remains unestablished here.

Tools and practice

3. Cheap model routing can make human-facing text more expensive

• What changed: Ethan Mollick argued that using a weaker model for text meant for people may soon feel disrespectful when the saving is only six cents but the output contains errors and poor writing. He used AI-generated X comments as his example. 6
• Why it matters: A routing decision should include the reader's time and attention, not only the token bill. A cheaper draft that forces a person to repair every sentence can cost more in the task that matters.
• Evidence boundary: The post states a practical standard rather than a controlled comparison between models. It gives no general threshold for when the extra quality is worth the extra price.

Society and culture

4. Science-fiction writers remain a visible source of resistance to LLMs

• What changed: Mollick said that most of his favorite hard science-fiction authors express hostility toward LLMs. He named three reasons in the scan: the view that LLMs are stochastic parrots, intellectual-property concerns, and existential-risk concerns. 7
• Why it matters: The people who write about technological futures are also part of the cultural and rights debate around AI. Their objections can affect how readers, publishers, and technology companies approach training data and creative work.
• Evidence boundary: The post describes an informal look at selected author webpages. The sample is unspecified, so the observation describes a selected group and leaves the wider field unresolved.

Models and definitions

5. Chollet defines general intelligence as adaptation to unfamiliar problems

• What changed: François Chollet wrote that intelligence becomes "general" when a system can make sense of a new problem on the fly, rather than relying on competence prepared in advance. He quoted Jerry Tworek, who described robustness as a continuing limit on automation. 89
• Why it matters: The definition points toward a practical test for broad model claims: give the model an unfamiliar problem, then examine how it handles the parts it was not prepared to solve.
• Evidence boundary: Chollet's post offers a definition and Tworek's quote offers an observation. Both posts offer a definition or an observation; a benchmark would be needed to measure a particular model's general intelligence.