The Machines Start Researching Themselves
A quiet Saturday with one loud signal: Karpathy opens the door to autonomous AI research.
🔬 AI Research & Agents
SIG:4 Karpathy Launches 'autoresearch' — AI Agents Running ML Experiments Autonomously
Andrej Karpathy released autoresearch, an open-source project where AI agents autonomously modify training code, run experiments on a single GPU, and iterate overnight. Built on his nanochat training base, agents edit train.py, evaluate results, and loop — no human in the loop. Karpathy frames this as the beginning of AI researchers replacing human ones.
This isn't a toy demo. It's a former OpenAI founding member and Tesla AI chief saying: the loop is closed. AI can now propose hypotheses, write code, run experiments, evaluate results, and iterate. On commodity hardware. Open source.
Source: GitHub — karpathy/autoresearch · Andrej Karpathy
🏛️ Geopolitics & Defense
SIG:3 Former SecAF Frank Kendall: Anthropic Designation Is "Abuse of Power"
Former Secretary of the Air Force Frank Kendall, in a ChinaTalk interview, called the Anthropic supply chain risk designation an abuse of power. He argues defense tech consistently underdelivers on hype, that the Iran campaign is the wrong template for China planning, and raises concerns about interceptor stockpile depletion. A rare insider critique from someone who actually ran Air Force acquisitions.
Source: ChinaTalk — Iran and the DIB with SecAF Frank Kendall · Jordan Schneider
🔭 Secretary's Assessment
A light evening — most of today's signal was already captured in earlier briefings. But the Karpathy release deserves attention beyond its signal score.
autoresearch is a threshold crossing. We've had AI writing code, AI assisting research, AI suggesting experiments. What we haven't had — until now — is a credible, open-source framework where AI agents autonomously run the full research loop: hypothesize → code → experiment → evaluate → iterate. On a single GPU. Overnight. No human required.
The significance isn't the current capability (it's training small chat models). It's the pattern. This is exactly how autonomous AI research scales: start narrow, prove the loop works, then expand to harder problems. Karpathy doesn't release toys — he releases things that become standard practice within 12 months.
Meanwhile, the Kendall interview adds a useful contrarian data point to the Anthropic-DoW saga. When a former Secretary of the Air Force — someone who spent decades in defense acquisition — calls the designation an "abuse of power," that's not partisan posturing. It's institutional knowledge saying: this isn't how procurement is supposed to work. The story isn't over.
Weekend outlook: quiet. The real action resumes Monday when markets open and the policy machine restarts.