Summary
A thin slot with a low signal-to-noise ratio: 54 tweets, no curated reposts, and only three items with any real AI substance. The strongest is Prof. Tom Yeh announcing a hand-worked walkthrough of Kimi Linear, the attention design behind Kimi 3, for this week's live seminar. He is sketching Kimi Delta Attention, attention residuals, and the extreme sparsity that activates 16 of 896 experts in a 2.8-trillion-parameter model, which is the clearest teaching artifact yet for the architecture cere-bro covered on 07-28. Second is a French-language thread reacting to OpenAI's ten Lean-verified math results, used as a stick to beat the "AI isn't really intelligent" commentary class, which is downstream noise but confirms the Astra announcement is now landing in non-English tech discourse. Third is a co-sign of Clément Delangue's argument that the recent AI-powered cyberattacks are a reason to accelerate open-model defense rather than slow down. Everything else is geopolitics, French and American domestic politics, and wildfire coverage from three high-volume accounts that produced 49 of the 54 tweets between them. Skip the feed, read the Kimi item.
Posts
- Kimi Linear worked by hand for the Kimi 3 seminar (@ProfTomYeh · event page). Yeh is implementing Kimi Linear in Excel and sketching the math by hand for the Frontier track of his AI by Hand series, covering Kimi Delta Attention, attention residuals, and MoE sparsity at 16 active experts out of 896 in a 2.8T-parameter model. Free live attendance, 2,300+ registered, and the most useful mechanism-level explainer likely to appear for this architecture. See Kimi K3 and delta attention with residuals.
- OpenAI's ten Lean-certified math results reach French tech discourse (@brivael). A long thread using the Astra proofs, formal Lean certificates on decades-old open problems, to attack Luc Julia's long-running "artificial intelligence does not exist" media position. No new technical content, but it marks the point where the result stops being a research story and becomes a public-argument weapon. Background: OpenAI Astra's ten math advances.
- Delangue: accelerate rather than slow down after AI-powered cyberattacks (@brivael). Co-signing HuggingFace CEO Clément Delangue's claim that an open model was what actually defended against a recent AI-driven attack, and that the same systems scale to defending against millions of attacks daily. It is the open-weights safety argument in its most compressed form, and it cuts against the reflex to treat every offensive-capability demo as a case for closing models. Related: Thinking Machines on the safe path for open weights.
- Neuralink targets first vision implants in 6 to 12 months (@brivael). Relay of a claim that Neuralink will implant cortical vision devices for congenitally blind patients within the year, with infrared and ultraviolet perception floated as later capability. Timeline claims from this source have slipped before, so treat the window as aspirational.
- Humanoid robot repair as a business model (@Scobleizer). CIX claims to have repaired six different humanoid platforms, more than any other US company, by running robot fighting as the thing that generates the damage. A genuinely odd go-to-market, and a reminder that the humanoid supply chain still has no standard service layer.
- Grokipedia versus Wikipedia (@brivael). A side-by-side comparison repost with no accompanying analysis. Click through if the framing interests you; there is nothing to extract here.
- Google at Deep Learning Indaba 2026 (@GoogleResearch · details). Sponsorship announcement for the Lagos event with booth demos and Q&A sessions. Skip.
- VITURE Luma XR glasses unboxing (@Scobleizer). Product amplification, no benchmark or capability claim. Skip.
- Geopolitics and domestic politics flood (cluster of 46, across @MarioNawfal, @brivael, @spencerpratt, and @HouseGOP). Ceuta migration, Iran and Hormuz, Israeli military discipline, NYC and Wisconsin municipal politics, Greek and Spokane wildfires, French EU commentary, and a birthday greeting. Nothing touching AI research, hardware, or industry. Skip.