Summary
Only the morning slot carried signal today. The afternoon and evening slots were empty, which fits the US overnight hours they cover. The strongest cluster is serving and kernels. PyTorch and Red Hat put an autotuned Helion GEMM backend into vLLM that beats CUTLASS and DeepGEMM on Hopper. AMD previewed its FlyDSL kernel language, and Prime Intellect turned its RL rollout fleet into a public serving business. Two more clusters back it up: buildout economics (Zitron's GDP essay, Morgan Stanley's compute-per-gigawatt case) and cost per task (Agent Arena prices plus OpenAI's DevDay cost levers). The single best standout is Supabase's $150M raise: 70% of its new databases are created by agents. Market tickers and meme reposts were noise. The day's papers are covered in the daily digest.
Posts
- Serving and kernels: Helion in vLLM, AMD FlyDSL, Prime Inference (cluster of 3) (@PyTorch · @PyTorch · @vincentweisser · Helion blog · Prime blog) [morning]. An autotuned Helion GEMM beats CUTLASS and DeepGEMM in vLLM on Hopper, gaining over 10% throughput on some workloads. Prime Inference serves open models on Blackwell using Dynamo, vLLM, Mooncake and FlashInfer. See Helion in vLLM.
- Economics of the buildout (cluster of 3) (@edzitron · @not_ellington · @StockSavvyShay · essay) [morning]. Zitron argues AI's GDP share is mostly GPU sales and construction, not AI services. A practitioner calls ARR misleading for inference businesses, and Morgan Stanley calls compute per gigawatt the deciding metric.
- Cost per task on Agent Arena + DevDay cost levers (cluster of 3) (@arena · @arena · @omarsar0) [morning]. GPT-6.1 Sol ranks #5 at $0.56 per task. Sonnet 5.5 ranks #3 but costs $2.74, more than Opus 5.5, so it sits off the cost frontier.
- Supabase raises $150M at $10.65B (@ycombinator) [morning]. 70% of new databases are created by agents or AI tools. Agents are now the main customer for some developer infra.
- Hinton co-authors intelligence explosion report (@geoffreyhinton · report) [morning]. Says many leading researchers now think recursive self-improvement may come soon. Only the landing page was captured.
- Chollet's three biggest shifts (@fchollet) [morning]. Inductive over transductive models, symbolic tool use via harnesses, and training on code execution. The harness is becoming part of training.
- Meta opens Muse to hardware (@alexandr_wang) [morning]. Open ESP32 firmware and a Linux SDK let anyone build gadgets for Meta's Muse agent.
- Altman on Cerebras and dot (@sama · @rohanpaul_ai) [morning]. Calls Cerebras a close partner on speed. A repost of Cerebras's CEO frames wafer-scale as a decode-bandwidth play.
- webAI TwIL-LM3-Pro, a 3.6B logic model (@0xCodez · model) [morning]. Vendor reports Qwen3-8B parity on formal logic at 2.09 GiB in 4-bit, runnable on CPU.
- Cohere Embed 5 with RCP-nDCG@10 (@cohere) [morning]. A new metric grading every returned result on relevance, so its numbers do not compare directly with older nDCG tables.
- Harness skills and courses (cluster of 3) (@undefinedKi · @khushiirl · @stanfordnlp · agent-skills repo) [morning]. Addy Osmani's agent-skills repo, a free harness engineering course, and Stanford's CS224V Agentic AI course.
- Agents fixing their own harness bugs (@rohanpaul_ai) [morning]. A paper turns real harness bugs into runnable tests. Coding agents miss most but improve with lessons from past fixes.
- Market tickers, Musk SpaceX/robotaxi posts, Schmidhuber meme, Minecraft repost, prompt-pack bait [morning]. Skip.