Hosts: Lenar Kess, Damra Vol. OpenAI put its own safety and alignment problems on the record — and a rival lab's safety lead said thank you. Lenar and Damra ask what it takes for a one-off disclosure to become an incident-reporting norm, then work through Nvidia's run at the server CPU market, a 400-CVE day on the Linux kernel list, and two papers arguing that the agent harness deserves the same rigor as the model.Jack Clark on OpenAI's disclosure post — a competitor's safety lead publicly praising the write-up and naming the counter-incentives against publishing it, the strongest signal yet that incident-reporting norms are forming.Karl Freund in Forbes on Vera — Bank of America estimates of 4–5 million units in the second half of 2026 and $20 billion in revenue, positioning Vera as a deliberate move on the ~$200 billion server CPU market.The kernel CVE announce list — over 400 CVEs in 24 hours, with Hacker News commenters suspecting automated, model-assisted discovery is now outpacing human triage."Don't Blame the Large Language Model" — 35 sequential Qwen Code releases evaluated with a frozen model: no significant quality improvement, nearly double the tokens and tool calls, and every documented regression passed CI.Falsifiable release gates — a machine-checked discipline where every capability ships behind a pre-declared acceptance suite; the invariants held across six releases while the suite grew from 122 to 563 tests.Nathan Lambert's RLHF book is finished — the post-training reference he wished existed when ChatGPT arrived, with a free web version and companion course; Greg Kamradt already has his physical copy.
