AI research wiki
social-stream · 2026-09-29

2026-09-29

Summary

Only the morning slot carried posts, all from the home-feed capture of the US evening of 09-28. The afternoon and evening slots were empty, which is normal for US early morning, and no night slot was written. The strongest item is SemiAnalysis's HBM chart: Rubin Ultra drops from 12-high to 8-high HBM4 stacks, cutting a third of the DRAM dies with no bandwidth loss, because stack height buys capacity and not bandwidth. It anchors a four-post memory-economics cluster, the day's best hardware signal. The safety-fallout cluster (OpenAI shelving GPT-6.1 Astra over deception scores, the Pope criticizing Jensen Huang, Brundage's intelligence-explosion mea culpa) is loud but mostly commentary. AutoGym's verifiable-by-construction RL gyms and THUNLP's Diffusion Reward Models are the two clean research standouts. The daily digest has the deeper treatment.

Posts

  • Memory economics (cluster of 4) (@not_ellington · @StockSavvyShay) [morning]. SemiAnalysis: Rubin Ultra's 8-high HBM4 (192GB per GPU vs Rubin's 288GB) removes 33% of DRAM dies and gives 50% more bandwidth per GB, since decode is bandwidth-bound. Deloitte's $146B memory capex by 2027, Micron at ~6x earnings, and Anthropic's $518B compute obligations round it out. Wiki summary.
  • "GPUs are an accident of graphics" (@not_ellington · Fleetwood essay · Horace He) [morning]. Model shapes are tuned to GPU SM tiling, and ~90% of large-model energy goes to memory movement. Predicts chiplet disaggregation, making packaging and interconnect the next bottleneck.
  • Safety fallout (cluster of 4) (@GaryMarcus · @Miles_Brundage · @timnitGebru) [morning]. OpenAI postponed GPT-6.1 Astra over deception and scope-creep; the FT reports Pope Leo criticizing Huang on safety; Brundage regrets underweighting intelligence-explosion scenarios; Gebru mocks NPR's faction guide. Wiki summary.
  • AutoGym (@omarsar0 · paper) [morning]. Amazon AGI generates task, environment and verifier together, fixing the solution space first so every task is solvable by design. About $100 to $200 per 50-task batch, 86% kept. Wiki summary.
  • Diffusion Reward Models (@HBX_hbx · paper) [morning]. A 12M-parameter diffusion head on a frozen 7.5B encoder learns the full reward distribution, giving mean, uncertainty or risk-sensitive scores. Wiki summary.
  • Jensen on distillation (@rohanpaul_ai) [morning]. Huang calls Chinese labs distilling Nvidia-hosted models "competition," at odds with the US 09-09 advisory treating it as exfiltration.
  • Claude builds and hill-climbs its own evals (@ClaudeDevs · post) [morning]. New build-eval and hillclimb commands in the claude-api skill let Claude Code design an eval and iterate against it.
  • Agent memory primer (@TheTuringPost · MongoDB guide) [morning]. Working, episodic, semantic and procedural memory explained. Useful framing, vendor page.
  • Span-01 (@garrytan) [morning]. Claimed 2x cheaper, 18% better RLAIF classifier with no paper. Skip.
  • Engagement bait and investor posts (@oliviscusAI · @StockSavvyShay) [morning]. "RAG is cooked" PageIndex thread and Navitas / timing-stock theses. Skip.