Hosts: Lenar Kess, Damra Vol. A day of performance claims made by the companies whose products are being measured, and what it takes to check any of them.SemiAnalysis put OpenAI's Jalapeño inference chip through the InferenceX benchmark against Nvidia, AMD and Google parts on open-weight models. OpenAI claims 1.5–1.9x better performance per watt and 1.7–3.6x lower latency. The Nvidia half is reproducible by anyone with the hardware; the Jalapeño half is not, because you can't rent one.Z.ai confirmed the stealth model Ox Alpha is a new GLM iteration, with weights shipping tonight after it topped OpenRouter's leaderboard at zero cost.Reuters reports Moonshot AI is in early talks with Microsoft, Amazon and Google over revenue sharing to host Kimi K3, seeking up to a 30% share — an open-weights lab asking for a cut of inference revenue it has no legal claim to.Meta drew up plans to cut many teams by as much as 60% to become "AI native", then pulled back after staff pushed and internal data showed the agents were ineffective. The plan preceded the evidence.Dylan Patel's per-megawatt arithmetic: base infrastructure at $10–15M per megawatt annually against frontier models generating up to $50M, with ASML tools and Zeiss mirrors as the binding physical constraint.Joshua Penman's Semantic Overlays paper adds an out-of-band annotation channel to a frozen model's residual stream. An overlay asserting a snippet is in a different programming language makes the model rewrite it faithfully in that language — the same power an attacker wants, held by the serving stack.Paritok-4B compresses coding-agent context to 25.7% of its size with an extractive adapter, and reports that using GPT-5 as your compressor costs more than the tokens it saves.The Guardian traced 124 reports and 560,000 words in nine days to a site badged with a thinktank that doesn't exist, built on a commercial platform that sells chatbot-citation optimization.A proposed EPA change would let air pollution permits issue without public input, which changes when neighbours find out rather than what gets built.
