social-stream · 2026-11-15

2026-11-15

Summary

The day's strongest cross-slot cluster is self-improving agent harnesses: Harness² in the morning and Self-Harness in the afternoon reach the same rule independently, keeping a harness edit only if real reruns prove it helps (GLM-5 gains 11.5 points on WorkBuddy-Bench, Qwen3.5 doubles on Terminal-Bench 2.0). Around it sit MemCredit's memory selection by decision effect, Anthropic's always-on agents guide and Claude Code's advisor mode, so harness and memory engineering carried most of the day's real signal. The Jensen Huang exit aftermath ran in both slots too, but it is mostly opinion, plus one unverified TechPowerUp scarcity allegation. The best single-slot posts were efficiency and precision ones: Meta's e-means (k-means quality 10-45x faster on GPUs) and Sasha Rush's group proving rounding bugs in PyTorch, JAX and diffusers. Noise was heavy: Starship's tower catch, hashtag reposts and paid growth ads made up most of the afternoon feed, and the evening slot was empty.

Posts

  • Self-improving harnesses (cluster of 2) (@rohanpaul_ai · Harness² paper; @HoodieFade_) [morning + afternoon]. Both keep a harness edit only if it changes behavior on real runs without regressions. Harness² lifts GLM-5 by 11.5 points, and Self-Harness takes Qwen3.5 from 18% to 36.7% on Terminal-Bench 2.0 with frozen weights. See agent harness engineering.
  • MemCredit (@omarsar0 · paper) [morning]. Scores each memory by whether removing it changes the agent's next action, hitting 86.6% on ALFWorld vs 65.7% for BM25 with 74.5% fewer memory tokens. See MemCredit.
  • Anthropic: Building always-on agents (@Xudong07452910 · article) [afternoon]. Wake on message, webhook or timer, run one turn, save state, sleep. The deferred tool call frees the worker while waiting for human approval; compare the DeepSeek Super-agent.
  • Claude Code advisor mode (@xsmryze · docs) [afternoon]. Sonnet drives, Haiku explores, and Opus is consulted only at plan lock, on repeated errors or before finishing. Cost routing by moment; see advisor routing.
  • Building agents that work at Meta (@taha_rafiq) [morning]. Four rules: verification first, read traces, don't over-engineer, build feedback loops into the product.
  • e-means: k-means rebuilt for GPUs (cluster of 2) (@akshay_pachaar, @TheTuringPost · paper) [morning]. EMA center updates over mini-batches with scaling-law hyperparameters reach 25 epochs' worth of k-means loss in about 2 epochs. Relevant to vector quantization and retrieval indexes.
  • When Rounding Breaks Machine Learning (@ChrSzegedy RT @srush_nlp · post) [morning]. A Lean float library swept 154 operators and about 780M outputs, finding unreported bugs in PyTorch, JAX, diffusers and others. See the wiki summary.
  • Dory1.0-4B explainer (@noahatwhere · paper) [morning]. NVIDIA's looped 4B model goes from 50.3% to 61.8% on SWE-bench Verified at four loops, with fewer output tokens. See looped transformers.
  • Embla 1: 10B offline on iPhone (@ahsuzm) [afternoon]. Sevren claims 1.77s answers in airplane mode on an iPhone 18 Pro, with self-reported benchmarks. The open question is which quantization makes it fit.
  • Against All Likelihood (cluster of 2) (@dair_ai, @rohanpaul_ai · paper) [morning]. A Meta essay arguing truth should mean invariance across personas, not likelihood. No experiments; see the wiki summary.
  • Weight-Space Reading (@crazyMLguy · repo) [morning]. Qwen 27B predicts weight-edit effects at 73% balanced accuracy from raw weights vs 57.8% from text. False positives on no-op edits are the open problem.
  • Decision models as teachers (@rohanpaul_ai · paper) [morning]. A recap: JEV is weak as an actor but lifts another agent from 72.8% to 82.9% as a post-run grader. See the wiki summary.
  • Emergent utility functions (@thesupermannx) [afternoon]. Resurfaces Utility Engineering: bigger models hold more coherent preferences and are less corrigible. Framed for outrage.
  • Nvidia and the Jensen Huang exit (cluster of 8) (@GaryMarcus RT @TechPowerUp, @rohanpaul_ai, @dylan522p, @elonmusk, @garrytan RT GavinSBaker, @garrytan RT dee_bosa, @VaibhavSisinty, @adamemedia) [morning + afternoon]. The TechPowerUp allegation that Nvidia funded its own GPU buyers is headline-only, and the rest is reinstatement petitions, an investor letter and one conspiracy thread. See Nvidia without Jensen and open-source AI without Jensen.
  • Long reasoning as the investment theme (@rohanpaul_ai) [morning]. Eric Schmidt bets on models that work one problem for hours or days.
  • Interpolation versus search (@fchollet) [morning]. Chollet: SGD curves only interpolate, but used to guide search they become search systems.
  • Agent swarms primer (@undefinedKi) [morning]. Generic shared-task-board explainer. Light signal.
  • Fable 5.2 showcase list (@deedydas) [afternoon]. A vibe check of demos, two of which failed. Not an eval.
  • YC and startup reposts (@garrytan, @garrytan, @snowmaker) [afternoon]. Bland's Tofu-1 speech model, AgentRun at 1,000 stars, YC on Grok Bot. Launch promo.
  • Robotics live-challenge proposal (@berkeley_ai RT JitendraMalikCV) [afternoon]. Judged live robot challenges instead of demo tapes. Outside focus.
  • Flat Minima priority claim (@SchmidhuberAI · paper) [afternoon]. Skip.
  • "20x cheaper harness" (@IntuitMachine) [afternoon]. One line, no link. Skip.
  • Skeptic snark (@edzitron, @GaryMarcus) [afternoon]. No new claim. Skip.
  • Starship tower catch (cluster of ~13) (@elonmusk) [morning + afternoon]. Off-topic. Skip.
  • Hashtag reposts, list threads and ads (@Ronald_vanLoon, 3D-studio and Obsidian repo lists, Explee ads, Academic Humanizer, Claude Council, "we achieved AGI", Grok Bot bait) [morning + afternoon]. Skip.