Glaneur
Glaneur AI

OCT 26 - Self-improving AI by 2027, says Jack Clark

Monday, October 26th · 2 min 44

In this episode

Full transcript

Good morning. It's Monday, October 26th, and here's the AI news, from France.

Jack Clark, a co-founder of Anthropic, has moved forward his prediction for self-improving AI. He now expects it as early as 2027. His update rests on a new benchmark called TasteVal. It measures what researchers call research taste, the ability to design experiments and interpret the results. The best model tested, Anthropic's Opus 5.5, matched a panel of 24 human experts from labs like OpenAI, Google DeepMind and Nvidia, using half the computing power. The benchmark covers only a narrow slice of science, its authors caution.

Staying with Anthropic, the company now says a single AI agent often beats a fleet of specialized subagents. Newer models can handle complex plans in one loop, loading tools and context only when needed. In one comparison, the single agent ran about twice as fast. Anthropic warns against what it calls the Frankenprompt, where teams pile up contradictory rules. It has published a migration kit to help companies make the switch.

Meanwhile, former Google chief executive Eric Schmidt compares OpenAI and Anthropic to the colonial trading companies of past centuries. In an essay for the journal Foreign Affairs, written with historian Niall Ferguson and researcher Lisa Klaassen, he argues that America and China have locked up the global AI market. Washington trades closed systems for political loyalty, they write, while Beijing gives away open models to spread its standards. They see a conflict over Taiwan as the greatest threat to the whole supply chain.

The price war is also heating up. According to the tech outlet The Information, OpenAI is offering discounts of up to 90 percent to lure Anthropic's biggest customers. The payroll company Rippling raised OpenAI's share of its coding work from 16 to 52 percent.

In brief. Elon Musk's Grok has released the first seven minutes of an AI-generated version of Homer's Odyssey. A TechCrunch reviewer found it less ugly than expected, with moments that look almost real.

And Nvidia has open-sourced Flora, a system that labels robot training videos. It says one hour of video costs 72 cents, against 88 dollars with OpenAI's GPT-6 Astra.

That's it for today. See you tomorrow.

Subscribe

RSS feed Apple Podcasts, Spotify