Summary
No curated reposts again this slot, but the AI handle feed carried the single biggest personnel and hardware news of the week. The strongest signal is a Google DeepMind leadership reshuffle landing inside four minutes of itself: Demis Hassabis steps back from running DeepMind to become its Chair plus Chief Scientist of Alphabet, and Jeff Dean is leaving Google after 27 years to found a public benefit corporation called Discovery Loop with Sanjay Ghemawat, Quoc Le and Oriol Vinyals, aimed at automating the experimental loop in ML and science, with Google supplying money and compute. Running second, and more load-bearing for this wiki, is a cluster of three posts confirming Anthropic is building custom inference silicon, sourced to a Business Insider piece and a live job listing, with Samsung in talks on 2nm and advanced packaging. Third is a small but real cluster on model provenance and vendor lock-in: ByteDance's founder has reportedly banned employees from distilling US frontier models to protect TikTok politically, while Elon Musk committed SpaceX to NVIDIA GPUs exclusively and called Vera Rubin the best architecture. Outside those, Prof Tom Yeh's hand-worked 14-step walkthrough of the Diffusion Transformer is the one genuinely educational item, and Shopify's $116B quarter is the one industry number worth keeping. The rest of the slot is close to unusable: of 49 tweets, roughly 30 are US and Middle East political commentary from four accounts with zero AI content, so the effective signal density here is under 40%.
Posts
Hassabis steps back from running Google DeepMind; Jeff Dean leaves Google to found Discovery Loop (@ns123abc, @ns123abc) (cluster of 2). Hassabis moves to Chair of Google DeepMind and Chief Scientist of Alphabet, shifting his own focus to AGI, scientific discovery and disease through Isomorphic Labs, with the quote "I feel AGI is close at hand." Four minutes later, Jeff Dean's departure after 27 years, taking Sanjay Ghemawat, Quoc Le and Oriol Vinyals with him into a public benefit corporation whose stated mission is to automate the experimental loop in machine learning, science and engineering, with Google investing and providing compute. Losing the co-author of MapReduce, the co-author of seq2seq and the co-author of AlphaStar in one announcement is not a normal reorg, and the fact that Google is funding the spinout rather than fighting it says they would rather own a stake in automated research than run it in-house.
Anthropic confirms an in-house custom silicon team, with Samsung in talks on 2nm (@ns123abc, @ns123abc, @ns123abc, Business Insider) (cluster of 3). Confirmed on the record via a job listing and a company statement: Anthropic is hiring silicon engineers at $320k to $485k to co-design hardware and models together for Claude. The earlier scoop being re-upped here adds that the target is an inference chip and that Samsung, already an investor in Anthropic's Series H, is in early talks for its 2nm process plus advanced packaging. Read this against today's digest finding that Google structured roughly $200B of Anthropic chip lease contracts off its own balance sheet with Broadcom, Apollo and Blackstone. A tenant that is quietly designing its own accelerator is a different credit risk than one locked into TPUs for a decade, and it is the same vertical-integration move OpenAI made with Broadcom.
ByteDance bans employees from distilling US frontier models (@ns123abc). The reported reasoning is entirely political rather than technical: distilling US models would attract Washington's attention and put TikTok at risk again, so "we should be willing to sacrifice some short-term gains for longer-term goals." The post notes that Anthropic has accused DeepSeek, Moonshot, MiniMax, Z.ai and Alibaba of distilling Claude, and pointedly not ByteDance. If true this is the first case of distillation policy being set by regulatory exposure rather than by capability economics, which is worth adding to the knowledge distillation picture, where the assumption has been that everyone distills whatever they can reach.
Musk commits SpaceX to NVIDIA exclusively, calls Vera Rubin the best architecture (@ns123abc). Two lines, but they close a loop from this morning's orbital-compute announcement and from the wiki's Vera Rubin gigascale ramp coverage. An exclusive-supplier commitment from the buyer building toward 10 GW is the strongest possible vote against the custom-ASIC path that Anthropic just confirmed it is taking, on the same day. The two bets are now explicit and opposed.
Diffusion Transformer worked by hand in 14 steps (@ProfTomYeh). A manual walkthrough of DiT, the architecture that replaces the U-Net in a diffusion model with a transformer, using a toy example: one training video, the prompt "sora is sky", diffusion step t equals 3. The framing is the useful part. OpenAI shut Sora down this year, but DiT is what every current video generation model is built on, so the brand died and the mechanism did not. Best item in the slot for anyone who wants the mechanism rather than the benchmark.
Shopify posts $116B in merchant sales, up 32% year over year (@dhh). $3.6B revenue up 34%, $654M free cash flow at an 18% margin, fifth straight quarter of GMV growing over 30%. DHH's angle is that all of it runs on Rails; the company's own angle is that every shift AI is creating happens to play to Shopify's strengths, which is the claim to actually test. Notable as a rare case of AI-adjacent growth showing up in cash flow rather than in capex commitments.
AWS open-sources Kiro Crew for scheduled agent crews (@mattsgarman). Carryover from earlier in the day and still the most concrete agent-infrastructure release of it: a workspace where you run several agents at once, wired into existing tools, schedulable for on-call rotations, cross-repo incidents and overnight migrations. Worth reading against today's finding that agent skill libraries mostly fail to abstract from experience, since scheduling more agents does not fix agents that do not learn.
Wispr Flow ships Notetaker (@Scobleizer, product). Meeting notes with per-speaker attribution, pitched as clean enough to feed straight into agents and MCP without cleanup. The attribution claim is the only technically interesting part; the rest is an ad Scoble liked. Skip.
Grumbling about Claude and about scraping law (@ns123abc, @ns123abc) (cluster of 2). One post asking why Opus feels like it is getting worse over time, one asking when OpenAI and Anthropic will be punished for violating the Computer Fraud and Abuse Act. No evidence attached to either, but the first is a recurring vibe worth tracking as sentiment if it keeps appearing.
Google Research booth programming at Deep Learning Indaba (@GoogleResearch, @GoogleResearch, @GoogleResearch) (cluster of 3). An AI trivia quiz, a WeatherNext tropical cyclone forecasting demo, and a Q and A on the sociotechnical implications of AGI with the Post-AGI Research Team. Event logistics only. Skip.
Non-AI political and lifestyle commentary (@MarioNawfal, @spencerpratt, @WHFraudTF, @HouseGOP, @AustinJustice, @DoWCTO, @NICKIMINAJ, @heavypulp, @stepango) (cluster of 30). Strait of Hormuz shipping, Israel-Lebanon talks, cartel indictments, LA housing politics, a Starlink terminal, and a Dow record whose only AI content is the phrase "strong AI earnings." Nothing here bears on research or industry. Skip.