curated · 2026-05-16

Eric Jang went on Dwarkesh's podcast to walk through his sabbatical project: rebuilding AlphaGo from scratch with modern tools. Both sources covering this are pulling from the same episode, so there's no independent reporting or third-party takes — the signal here is entirely what Jang chose to lay out. The sharpest part is his comparison between AlphaGo's MCTS and the policy-gradient RL used to train LLMs today. In LLM RL, the model has to guess which of 100k+ tokens in a trajectory actually led to the right answer. MCTS sidesteps this entirely by suggesting a strictly better move at every step. Jang argues human learning is closer to the MCTS pattern. That's a concrete structural critique of current RLHF pipelines, not just a history lesson. He also tested an automated research loop with LLMs and found they're decent at execution — running experiments, tuning hyperparameters — but bad at picking which question to investigate next and escaping dead ends. That's useful ground truth for the intelligence-explosion debate, backed by hands-on tinkering rather than extrapolation. What's missing: I haven't seen the actual cost breakdown or detailed repo numbers yet. The GitHub link is out there, but the compute bill isn't spelled out in the coverage.

Two outlets are running the same story: a researcher used Anthropic's Mythos tool to find and exploit two macOS kernel vulnerabilities on Apple's M5 chip, bypassing memory integrity protections, all within five to six days. The headlines agree, but they're both pulling from the same RSS snippet — no original advisory, no technical write-up, no Apple statement. I'd discount the confidence until we see more. The interesting part is Mythos itself. Anthropic has pitched it as AI-assisted security research, and if it genuinely helped surface kernel-level bugs on brand-new hardware this fast, that's a real step toward practical automated vulnerability discovery. What's missing: the exploit type, whether Apple had a heads-up, and how much heavy lifting Mythos actually did versus the human researcher. Don't read this as 'AI breaks chip security' until those details land.