Glaneur
Glaneur AI

NOV 3 - Anthropic warns it is close to self-improving AI

Tuesday, November 3rd · 2 min 43

In this episode

Full transcript

Good morning. It's Tuesday, November 3rd, and here's the AI news, from France.

Anthropic, the maker of the Claude models, has slipped a striking warning into its revised stock market filing. The document, filed on Monday ahead of its initial public offering, suggests the company is close to recursive self-improvement, known as RSI. That is the point where AI gets good enough to help build better AI, which could sharply accelerate progress. It warns this could pose a considerable safety risk. It says its Responsible Scaling Policy sets out how it will monitor and limit that risk.

The same filing details the company's finances. According to Bloomberg, quarterly revenue reached nearly 17 billion dollars, twelve times more than a year earlier. But operating costs hit 17.5 billion, and the net loss came to 23.3 billion dollars for the quarter. Anthropic blames 5 billion dollars of short-term computing capacity, and 2 billion in matched employee donations. In total, it has committed 895 billion dollars to computing power and data center leases. The public offering could raise around 100 billion dollars, according to The New York Times.

Elsewhere, Garry Tan, the head of the startup accelerator Y Combinator, says he spends 100,000 to 200,000 dollars a year on AI tokens. He told Walter Isaacson's podcast that his agents handle coding, research and much of his personal life. They have read his email, his files, and even twenty years of therapy notes. Tan has also shared a set of 23 Claude Code skills on GitHub, to show others what agents can do.

On a different note, Goldman Sachs has cut most of its staff off from AI coding tools, such as Claude Code. According to Bloomberg, only engineers can still use them, under internal guidelines from October. Traders and bankers had spent months using the tools. In July, the bank's technology chief had praised so-called vibe coding by non-technical staff. A spokesman cited productivity and risk management.

In brief. South Korea's court administration wants a legal basis to fine fake citations generated by AI in court filings. In a survey of 360 judges, about 80 percent of those handling civil, family and administrative cases said they had seen false or nonexistent rulings cited.

That's it for today. See you tomorrow.

Subscribe

RSS feed Apple Podcasts, Spotify