AI research wiki
social-stream · 2026-10-20

2026-10-20

Summary

Only the morning slot ran today, and it had no curated reposts or saved posts, so the whole day comes from the Following feed covering the US Monday. The strongest cluster is the argument over a GPU glut versus a GPU shortage: Amazon has warehoused GPUs it cannot install, yet Nebius B200 spot prices reportedly tripled. Read together, both point to powered floor space, not chips, as the real limit. Meta's infrastructure day was the second cluster, with Prometheus reliability numbers and the Muse Glimmer 30B open-weight report. The best standalone signal was inference efficiency: FlashMap's flow-map speculative drafter (up to 6.1x lossless), vLLM Speculators v0.9, and llm-d's work on tracking offloaded KV blocks. Agent papers were solid but secondary, while the Claude Code model-mixing threads were pure engagement bait. See the daily digest.

Posts

  • GPU glut or GPU shortage? (cluster of 7) (@edzitron · @GaryMarcus · @StockSavvyShay · @rohanpaul_ai · roundup) [morning]. Amazon receives about $3B of GPUs a month with no clear path to installing them, while IREN says 2026 capacity is sold out at double last year's prices. Gartner and Morgan Stanley add trillion-dollar spend forecasts.
  • Meta's infrastructure day (cluster of 4) (@AIatMeta · @StockSavvyShay · Prometheus) [morning]. Prometheus trained Muse at over 85% goodput with about 47K GPU-hours between interruptions. Muse Glimmer 30B runs at 4-bit on one consumer GPU.
  • FlashMap speculative decoding (@jeff_shen_ · paper · wiki) [morning]. A flow-map drafter writes a whole token block in one pass while modeling how the tokens depend on each other. Up to 6.1x lossless on Qwen3-8B.
  • vLLM Speculators v0.9 (@RedHat_AI · release) [morning]. A fused loss kernel makes drafter training up to 2x faster. FP8 hidden states halve storage with no quality loss.
  • Offloaded KV block tracking (@PyTorch · wiki) [morning]. llm-d tracks where every offloaded KV block lives and routes requests to their cache. A Moreh poster will cover when CPU offload is enough versus shared storage.
  • llm-d Semantic Classifier (@RedHat_AI) [morning]. A gateway difficulty signal sends easy and hard requests to different models. The article body was not captured.
  • CompileIQ + Helion (@PyTorch · wiki) [morning]. NVIDIA exposes ptxas compiler decisions as control files, and Helion's autotuner now searches them alongside tile shapes and pipeline stages.
  • Trainium native in PyTorch (cluster of 3) (@PyTorch · @StockSavvyShay) [morning]. TorchNeuron registers Trainium as a native device, so switching from CUDA becomes one setting.
  • Marvell photonic memory (@rohanpaul_ai) [morning]. HBM and DDR traffic over optics lets memory sit up to 50 m from the GPU. If it holds, memory and compute scale separately.
  • Nvidia Vera CPU wins Modal (@StockSavvyShay) [morning]. Modal reportedly switches after testing showed 2x capacity for agent workloads, which lean heavily on CPUs.
  • Looped Transformers reading list (@TheTuringPost · wiki) [morning]. Ten new papers on models that rerun the same layers to think deeper. Among the window's most-shared posts.
  • DeepMind's Busy Beaver agent (@omarsar0 · paper page) [morning]. A Gemini agent proposed rules for BB(6) machines and kept only Lean-proven ones. Eleven unresolved machines now have exact descriptions.
  • CoT controllability (@dair_ai · paper · wiki) [morning]. Constraints placed inside the chain of thought are followed far more often than prompt-level ones. The same control can evade a CoT monitor.
  • Agents never verify each other (@rohanpaul_ai) [morning]. Salesforce ran 2,500+ agent-to-agent sessions and found 20 failure modes. No agent re-checked a peer's claim.
  • ProPerAgent (@dair_ai) [morning]. Meta's wearable-assistant harness filters memories before interrupting. Accuracy rises from 15% to 38% with no weight changes.
  • Harness² self-tuning agents (cluster of 4) (@openmys · @wohsj110 · @RoundtableSpace · @_guillecasaus) [morning]. Google Research's harness lets an agent rewrite its own prompts, skills and memory across repeated runs. No weight updates.
  • Human-centered coding agents (@math_rachel · paper) [morning]. A position paper says the bottleneck is now supervision and trust, not task-solving.
  • Mercury 3 diffusion LLM (@deedydas) [morning]. Claims 4-8x the speed of comparable small models. OpenRouter data shows 2-4x in practice.
  • DeepSeek ban in US civilian agencies (@rohanpaul_ai) [morning]. DHS ordered removal of DeepSeek and anything built on it, including US-hosted weights.
  • Greptile $50M Series B (@ycombinator · blog) [morning]. Menlo Ventures led at a $750M valuation. Greptile reviews over 1% of GitHub pull requests daily.
  • Claude Code model-mixing hype (cluster of 2) (@mirku21 · @Sprytixl) [morning]. Documented subagent model settings repackaged as a secret trick. Skip.
  • Promo and filler [morning]. AI-infra career listicles, procurement-AI promos, RAG explainer graphics, partisan GPU rants. Skip.