social-stream · 2026-11-14

2026-11-14

Summary

Only the morning slot had content. The afternoon and evening slots were empty and no night slot was written, so the morning Following-feed capture, which covers the US Friday afternoon and evening, is the whole day. No topic repeated across slots. Inside the morning slot, the strongest cluster was decision models: four posts covering a paper showing they judge better than they act, Raschka's tally of 16 vendor launches in seven weeks, Perplexity's open Decider v1.3, and a $5-10 weight-editing optimizer. The best single item was Thinking Machines' write-up of end-to-end 4-bit training for Inkling 2 Small. It is a real Tier 1 recipe covering NVFP4 forward passes, an unbiased transpose fix and quantized expert-parallel communication. CMU's Singularity-Bench, where an Opus-built serving engine out-earned SGLang per GPU-hour, and the Fed's $680B AI-debt numbers were the other strong signals. The noise was Huang affair threads, Grok prompt listicles and generic harness hype.

Posts

  • End-to-end 4-bit training of Inkling 2 Small (cluster of 2) (@soumithchintala RT of @thinkymachines · @rohanpaul_ai · blog) [morning]. NVFP4 in the expert forward passes and MXFP8 in the backward passes, with stochastic-rounding re-quantization after the transpose. The result is 8% faster pretraining and 20% faster RL for about 2% efficiency loss. See wiki summary.
  • Decision models: judge, timeline, open weights, weight editing (cluster of 4) (@omarsar0 · paper · @rasbt · @AravSrinivas · @dbreunig) [morning]. Decision models pick the wrong subgoal when they act but credit steps correctly in hindsight. Used as a GRPO teacher, one lifts a 1.5B policy from 72.8% to 82.9% on ALFWorld. Vercel says 26% of its requests now go through decision models. See wiki summary.
  • Agents that build the stack (cluster of 2) (@rohanpaul_ai · @shmidtqq) [morning]. On Singularity-Bench, Opus 5.5 scored 1.25 against the production baseline, and its serving engine earned $11.52 per GPU-hour versus $8.18 for SGLang. Separately, a Codex agent ran 1,000+ hours fixing CI at about $16-32 an hour. See wiki summary.
  • Fed financial stability report (@rohanpaul_ai) [morning]. AI infrastructure debt passed $680B, about 31% of investment-grade borrowing, and 87% of polled professionals now name AI the top financial risk. See wiki summary.
  • Nvidia smuggling allegations and succession (cluster of 5) (@edzitron · Megaspeed piece · @StockSavvyShay · @Miles_Brundage · @pmddomingos) [morning]. Zitron alleges a Megaspeed-linked network accounts for over 20% of FY2026 revenue. These are allegations, not findings. Rene Haas is floated as successor. Background: Nvidia without Jensen.
  • Anthropic's ten recommendations for safety evals (@AnthropicAI · post) [morning]. Applies replication-crisis fixes such as preregistration and sound statistics, and warns that results shift with question format and answer order. See wiki summary.
  • Gemini 4 Hydrogen in GitHub Copilot (@github · changelog) [morning]. Now GA in the Copilot app, CLI and VS Code, alongside a recap of the Copilot Fusion router.
  • Claude Fable 5.2 tops Agent Arena (@arena) [morning]. It leads Opus 5.5 in 9 of 10 categories, with the widest gap in visual design.
  • Harness evals and a self-improving skill (cluster of 2) (@garrytan · repo · @garrytan) [morning]. An open repo that ranks agent harnesses by task quality per cost. See agent harness engineering.
  • Unbundling multimodal vision encoding (@_ellisbrown · paper) [morning]. Across 200+ models trained from scratch, a pretrained visual prior turns out not to be needed, but dedicated vision capacity is. See wiki summary.
  • Memory writing for continual learning (@omarsar0) [morning]. The bottleneck is the model's judgment about when to write, not the store. See agent memory.
  • Epoch on alignment-eval incidents (@Miles_Brundage RT of @EpochAIResearch · AEF-1 post) [morning]. Three unexpected agent behaviours in Trajectory Labs evals. Only the hook was captured.
  • Light items (@DimitrisPapail · @rohanpaul_ai · @norpadon) [morning]. John Langford joined Core Automation. CodeRabbit's $1M of free scans found 1,215 high or critical OSS issues. Plus an on-policy distillation meme.
  • Huang affair threads, Grok prompt listicles, wristband link, Starship stream, harness hype [morning]. Skip.
  • Afternoon and evening slots [afternoon + evening]. No posts. The day's signal lives in the daily digest.