Summary
The morning slot had no reposts and no tweets from tracked handles, so this synthesis reads the latest capture of the Following feed, which covers the US Wednesday afternoon and evening. The strongest signal is a cluster of about 8 posts on Nvidia's leadership crisis: the board met Wednesday night to consider removing Jensen Huang, and the same evening brought a new Commerce rule against remote access to Blackwell chips. A second cluster of 3 posts is decision models, after Google launched the Gemini Decisions API, a model that scores up to 1,000 yes/no, choice and rubric questions in one pass without generating text. A third cluster of 3 covers Anthropic's claim that Mythos ported Claude to AMD GPUs in three days, which the feed reads as a crack in Nvidia's CUDA moat. Agent Arena posted two cost-aware results: Claude Fable 5.2 at #1 with a wide lead, and Meta's Muse Ember as the cheapest capable agent model. Weight-space editing kept circulating, now as a propose-score-keep loop. Smaller items: Stanford opened its agent-engineering course, a Microsoft/Tsinghua paper found most GPU kernel library bugs live in builds and environments, and Berkeley proposed watermarks for detecting distillation.
Posts
Nvidia's board weighs ousting Huang, and Commerce closes the remote-access route (cluster of 8) (@rohanpaul_ai on the Commerce rule, @rohanpaul_ai on the WSJ account, @edzitron, @VaibhavSisinty, @MarioHachemer, @StockSavvyShay on Burry). Per The Information, Commerce's rule would require export licenses for Blackwell and newer chips in Singapore, Malaysia, Thailand, Indonesia, Brazil, Norway and South Africa, and buyers there must disclose who receives the chips. Per WSJ, Huang met Alice Huang through her husband, and the couple's Singapore firm Megaspeed won priority Nvidia chip access. Ed Zitron's thread lists past accounting scandals and the circular-financing critique. One contrarian thread claims the coverage is a pressure campaign over GPU firmware controls for unsigned open models; that is an unsupported claim. Michael Burry calls Nvidia potentially "a heck of a trade" after the selloff, while doubting how fast GPUs lose value, the same week CoreWeave said old Hopper rents rose 30 to 50%. See the digest.
Google launches the Gemini Decisions API (cluster of 3) (@VaibhavSisinty, @omarsar0, @rohanpaul_ai on cost routing). Gemini Decision 1 does not write text. It scores options, with yes/no, multiple choice and rubric questions, up to 1,000 per request across text, images, audio, video and PDFs, in 150 to 250 ms at $0.10 per million input tokens and free output. The feed frames it as Google "coming for Jev", the startup that defined the category. Elvis Saravia plans to run it on his research benchmarks. A separate post shows the consumer version of the idea: a desktop video agent that checks each provider's price per render and routes every clip to the cheapest. Summary: the decision-model wave.
Mythos ran Claude on AMD in three days (cluster of 3) (@MarceloLima, @PyTorch on AMD DevDay, @PyTorch on Ascend). From Anthropic's roadshow: bringing up a new chip used to take dozens of engineers months, and Mythos did AMD with almost no human help, writing or porting almost every kernel. The feed's reading is "the CUDA moat is gone" and "all compute is fungible". No performance-parity number was given, so treat it as a claim. PyTorch's posts from AMD's DevDay and a Beijing Ascend meetup show the multi-chip ecosystem filling in underneath. Summary: Mythos on AMD and kernel-library failures.
Agent Arena: Fable 5.2 on top, Muse Ember the cheapest capable model (cluster of 2) (@arena on Fable, @arena on Muse Ember). Across 4.7K real agent sessions, Claude Fable 5.2 (Max) leads at +30.4% net improvement and $7.3 per task, ahead of Opus 5.5 (+19.5%, $5.5), Gemini 4 Carbon (+17.0%, $1.5) and GPT-6.1-Sol (+12.5%, $1.1). Muse Ember (Max) posts +1.7% at about $0.02 per task, roughly 4x cheaper than GPT-6 Luna and Haiku 5.5. The Pareto chart makes the routing case: the price spread between first and seventh place is about 7x.
GLM-5.3 at 200+ tokens per second on SambaCloud (@SambaNovaAI). SambaNova serves Zhipu's GLM-5.3, the top open-weight model on Terminal-Bench 3.0 (28.3) and DeepSWE v1.1 (66.9), with fewer output tokens per task than several closed models. Pointable from Claude Code or Codex CLI.
Weight-space editing in a loop (@_matt_winters, @DimitrisPapail). Beaker proposes a weight edit, scores it, keeps it only if it measurably helps, and repeats, turning MIT's one-shot weight-space programming into search. Papailiopoulos posts a 2023 slide where Angeliki Giannou hand-designed a Lazarus-style weight edit, a reminder that the backdoor idea predates the paper. Related: Deletion Tuning and Tree-Ising.
Where GPU kernel libraries actually break (@rohanpaul_ai, paper). Microsoft and Tsinghua read 300 closed issues across seven libraries including FlashAttention, Triton and CUTLASS. 60.33% came from compilers, builds and runtime environments, 39.67% from kernel code, and 25.33% from version mismatches among drivers, CUDA, frameworks and packages.
Disaggregation as a conveyor belt (@not_ellington). A short argument that disaggregated serving plus pipeline parallelism beats shuffling memory between GPUs: move the residual stream along stages that each hold the memory they need.
Distillation detection with watermarks (cluster of 2) (@hrshtv via Berkeley AI, @sewon__min via Berkeley AI). Subliminal Watermarks (SubWM) mark a model's outputs so a student trained on them can be detected later, and are claimed to be more robust than earlier watermarks. Sewon Min's framing: as model competition heats up, distillation is the leak labs most want to prove.
AWS CEO on inference and space (cluster of 2) (@rohanpaul_ai, @StockSavvyShay). Matt Garman says data centers in space are 10 to 20 years away. Amazon installs one GPU every 14 seconds, and current launch capacity cannot match that.
Inference demand outruns supply (cluster of 2) (@garrytan RT of Suhail, @AnatoliKopadze). Suhail says his company maxed out all its inference compute across customers in two weeks. A spend-anecdote thread (a $500M Claude month with no limits, Uber's 2026 AI coding budget gone by April) is unsourced, but it matches the CMU/JetBrains finding that token spend tracks usage, not output.
Persistent agents are 45% of SemiAnalysis's AI spend (@dylan522p). Dylan Patel says background agents are the next usage unlock, rates Perplexity Computer's harness above Claude's and OpenAI's Slack agents, and runs on a mix of models.
Stanford CS329Z goes public (cluster of 4) (@stanfordnlp, @jenniferzhang25 via stanfordnlp). Lectures are on YouTube and both homeworks are open, the first building a Claude Code-style agent from scratch for a few dollars. Andrew Ng was a guest.
Making coding agents trustworthy (@TheTuringPost on @poteto). Concrete steps: clean the code agents learn from, since they copy whatever looks like an established pattern, including workarounds and deprecated code.
Steve Yegge's 50-agent software factory (@TheTuringPost). About $10K a month to keep 50 agents running, with a rebuild every two months.
Cloudflare's path to an AI lab (@TheTuringPost). Cloudflare put GPUs near users in 2023 betting models would shrink; frontier models grew instead, which pushed it to train its own.
Déjà View: a looped 117M model for 3D (@NVIDIAAI). One transformer block run in a loop turns multi-view photos into geometry and camera poses, matching models up to 10x larger at similar compute. The loop count is a dial for compute.
Skip: the Huang affair jokes and pile-ons, "I can't believe this" repo-list threads, and builder follow-for-follow posts.