OpenAI used its models to help surface ten notable advances in mathematics and theoretical CS, while Quanta and a new explorative-modeling post probe whether reasoning models are actually reasoning or just getting lucky, and ByteDance shipped Seedance 2.5 for video generation. On the infra side, engineers are wrestling with bursty LLM inference (speculative KV replication), price/performance tradeoffs (Kimi K3 on AMD's MI355X vs. Nvidia B300), and whether LLM routers are even worth building, per Manifest's own deprecation post. Security had a rough day: Tailscale published a frank account of how its overlay network didn't stop a Hugging Face intrusion, CISA flagged active targeting of water-sector PLCs, and a subtle random.bytes() bug shows how silent randomness failures slip through. Meanwhile systems folks got NetBSD 11.0, a Go 1.27 interactive tour, a RowHammer/RowPress deep-dive, and a rare public postmortem of a soundness bug in the Lean theorem prover's kernel.

• OpenAI highlights ten math/TCS advances aided by its models
• text2
• Tailscale admits its overlay didn't prevent the Hugging Face intrusion
• CISA warns of active targeting of water-sector PLCs
• Manifest deprecates its own LLM router, questioning the whole category
• NetBSD 11.0 and a Go 1.27 interactive tour land for systems folks