hot events · 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.

All 5 items come from Reddit LocalLLaMA, and the titles align on PR #22673 being merged; the article body is only a 403 block, so this is community amplification, not independent confirmation. My read: MTP entering llama.cpp mainline matters because it hits decode throughput and speculative execution paths, the stuff local users actually feel at runtime. But the useful numbers are absent here: speedup, supported models, quantization behavior, backend coverage, and fallback rules. I would not treat this as a free latency win yet. llama.cpp has shipped plenty of clever optimizations that later exposed rough edges across CUDA, Metal, CPU, and mixed quant formats.

The thing that makes this worth opening: SANA-WM brings world model training down to a scale where a small lab can actually run it. One-minute 720p video generation used to mean either closed industrial systems or hundreds of GPUs. NVIDIA got it done with 64 H100s in 15 days, and inference runs on a single H100—the distilled version even hits 34 seconds for a 60-second clip on an RTX 5090. Both sources are pulling from the same NVIDIA project page, so the numbers are consistent but there's no independent verification yet. I'd hold off on the camera-control claims for now—all the demos show fixed-camera first-person views, and the paper's 6-DoF trajectory following hasn't been shown with moving cameras. The model weights are also marked "soon," so you can't test this locally yet. If the 36x throughput claim holds, the real unlock is iteration speed: long-video world model experiments that used to need a cluster can now happen on a single GPU.

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.

Four outlets covered Brockman taking product strategy, with English headlines stressing the agent race and Chinese headlines framing it as a power move. The shared hook is the same memo line: OpenAI plans to “invest in a single agentic platform.” I read this as OpenAI admitting its product surface sprawled too far. ChatGPT, Operator, Codex, and enterprise automation have each carried an agent story, and builders still lack a clean answer on which interface to bet on. Putting Greg Brockman over product says the company no longer trusts organic convergence across teams. Anthropic’s Claude Code path has been narrower and less internally noisy; OpenAI is now paying down org debt before it can sell agents as a coherent platform.

Alibaba Health launched a medical AI called Qinglingzi, pitched at China's 5 million doctors. Two tech outlets covered it, both hammering the same angle: exclusive access to 10 years of BMJ journal literature for evidence-based medicine. The coverage is nearly identical — even the '88 days, 193 logins' detail matches — which screams a single press release. One outlet frames it as 'top-tier evidence + evidence-based medicine,' the other as 'competing on evidence sources.' No real difference in angle. I'd discount this on two fronts. First, there's zero model-level detail: no base model, no parameter count, no clinical scenario benchmarks, no accuracy numbers. Second, '5 million doctors' is the addressable market, not actual adoption. BMJ access sounds impressive, but literature retrieval is a long way from clinical decision support — the real question is how the product integrates evidence into actual workflows. No pricing, no validation, no comparisons yet. Don't read this as a medical AI milestone.

Four outlets covered OpenAI connecting ChatGPT to bank accounts through Plaid, and the factual line is aligned; the split is tone: The Verge is alarmed, HN is plain, Chinese headlines swing between fear and reassurance. The disclosed facts are Plaid, bank access, and no money movement; the body does not give default permissions, retention periods, or training-exclusion terms. I don’t buy the “read-only, so safe” framing. Plaid data exposes salary, rent, debt, subscriptions, medical payments, and cash-flow stress as a continuous behavioral feed. That is denser than a Gmail summary. OpenAI has already moved toward health records, and bank feeds are the next obvious substrate for a personal agent. The sharp question is not whether ChatGPT can transfer a dollar. It is whether users can audit every read and revoke it cleanly.

All 3 headlines align, and the facts trace back to OpenAI’s own post: Maltese citizens who finish a University of Malta course get one free year of ChatGPT Plus, with phase one starting in May. I read this less as AI literacy and more as an OpenAI for Countries distribution pilot. Malta has about 500,000 people, EU status, and a small enough rollout surface to make the optics clean. The post gives the course, one-year access, and MDIA distribution, but leaves out procurement price, account-level data boundaries, and any cap on Plus seats. Compared with Estonia and Greece education partnerships, handing out Plus directly has a much sharper commercial edge.