NOV 9 - Sutton: raw compute alone is no longer enough
In this episode
- Sutton: raw compute alone is no longer enough - The Decoder
- AI coding agents still need human cleanup - The Decoder
- Today's AI is not conscious, experts argue - The Decoder
- Karpathy builds HackerTube with Claude - The Decoder
Full transcript
Good morning. It's Monday, November 9th, and here's the AI news, from France.
A debate is running through the AI research community this week, sparked by Richard Sutton. The Canadian computer scientist and Turing Award winner wrote "The Bitter Lesson" in 2019. That essay argued that general methods powered by computing power always end up beating approaches designed by human experts. In a new post, Sutton now says that raw compute alone is not enough for intelligence. The science of AI, he writes, is far from finished. He calls for agents that keep learning from their own experience, with models of the world that today's large language models lack. Other researchers, including former OpenAI scientist Jerry Tworek, agree that the algorithm matters more as training grows.
Turning to everyday use, a team from Alibaba and several Asian universities has put coding agents to a tougher test. Their benchmark, called Co-Design Bench, is built from real requests to modify existing web apps. The results are sobering. Even GPT-6 Astra, the best of the ten models tested, gets only 26.4 percent of tasks fully right. In most cases, the agent does the bulk of the work but leaves at least one detail unfinished. Older benchmarks, where models now score above 99 percent, no longer reflect real professional projects, the authors argue.
On a more philosophical note, three leading consciousness researchers have taken a firm stand in the journal Nature. Anil Seth, Ned Block and Giulio Tononi argue that today's AI is not conscious, and probably cannot be on silicon chips. They reject the idea that running the right computations is enough to produce subjective experience. Their commentary answers figures such as Geoffrey Hinton and Richard Dawkins, who consider that AI systems may already be conscious. The question is not only academic. It shapes the rules society may want for these systems.
In brief. Andrej Karpathy, the OpenAI co-founder who now works as a researcher at Anthropic, built a curated video site in a single weekend. Called HackerTube, it draws on 43 million comments from 20 years of the Hacker News forum to surface 6,454 recommended videos. Karpathy says he let Claude handle nearly all the work, from data analysis to server setup.
That's it for today. See you tomorrow.