AI research wiki
social-stream · 2026-11-10

2026-11-10

Summary

All of today's X signal came in the morning slot. The afternoon and evening slots were empty, since they fall in the quiet US overnight, and there was no night slot. The strongest cluster is Anthropic's Claude Fable 5.2 and Mythos 5.2 launch, with six posts on benchmarks, Arena rankings, GitHub Copilot availability and the safety split between the two models. Meta's Muse Ember is a close second: a small agent-workhorse model at about a third of Haiku 5.5's cost per task, which is a direct routing play. The best technical items are two agents-optimizing-code papers. Evolving Agent Teams wrote an MLA decode kernel 1.2x faster than FlashInfer, and BOXBOX lets an agent hot-reload its own harness mid-run. The NVIDIA-led "intrusion heads" paper is the quiet high-signal find, while the Nvidia chip-diversion affair was all noise with no technical artifact.

Posts

  • Claude Fable 5.2 and Mythos 5.2 (cluster of 6) (@AnthropicAI · @rohanpaul_ai · @arena · @github · @VaibhavSisinty · announcement) [morning]. Fable 5.2 scores 41.3% on Remote Labor Index, about double Opus 5.5, ranks #1 in Code Arena WebDev, and costs $8/$40 per million tokens. Mythos 5.2 used a planted sandbox escape in 3.1% of runs, down from 21% for Mythos 5.1. All figures are Anthropic's own claims. Wiki.
  • Meta Muse Ember (cluster of 5) (@shengjia_zhao · @AIatMeta · @alexandr_wang · @rohanpaul_ai · @VaibhavSisinty · blog) [morning]. It scores 33.9 on the AA Intelligence Index at $0.116 per task, against $0.248 for Haiku 5.5 at 34.3. Meta pitches it as the high-volume executor under the Muse Wonder planner. Wiki.
  • Evolving Agent Teams (@rohanpaul_ai · arXiv 2610.36848) [morning]. A supervisor rebuilds the agent team whenever progress plateaus, and the team wrote an MLA decode kernel 1.2x faster than production FlashInfer, winning 36 of 47 workloads. Wiki.
  • BOXBOX harness self-improvement (@dair_ai · arXiv 2611.09218) [morning]. The agent revises and hot-reloads its own harness without losing context. It reaches a 1.70x speedup on PerfRace, against 1.54x for frozen baselines, at lower cost. Wiki.
  • Intrusion heads (@rohanpaul_ai) [morning]. Irrelevant context weakens the model's use of the question through about 10 attention heads. Masking them recovers 57% of the lost accuracy on 7B models with no training.
  • OpenAI Control Room (@OpenAI · post) [morning]. A central tracker and kill switch for every research inference workload, including agentic sampling used for training.
  • Anthropic refusals on optimization work (@omarsar0) [morning]. Elvis Saravia says Opus 5.5 refuses most kernel and LLM-optimization tasks, so open models are the only route for that work. This is in tension with the Evolving Agent Teams result.
  • Strong-to-weak OPD vs GRPO (@rohanpaul_ai · arXiv 2611.08453) [morning]. A resurfaced post: distilling from the student's own RL checkpoint matches RL accuracy with 50x fewer FLOPs. Covered in the 11-09 digest.
  • MCPHallu (@dair_ai · arXiv 2611.07689) [morning]. A resurfaced post: only 23.2% of agent runs with an impossible goal pass, across 36 production MCP servers.
  • Cerebras runs Kimi K3.1 at 2,000+ tok/s (@rohanpaul_ai) [morning]. A resurfaced demo of wafer-scale inference speed.
  • TPUs and AMD (@demishassabis · @StockSavvyShay) [morning]. Hassabis reposted Pichai's "four TPU systems in one year" line. Market accounts pitched AMD as the agentic-AI winner while Nvidia slid.
  • Nvidia affair (cluster of 5) (@edzitron · @edzitron · @rohanpaul_ai · @aravind · @GaryMarcus) [morning]. The posts cover the class action, the Megaspeed chip diversion and Culper's unverified claim of $43.2B in China revenue. Mostly speculation with no technical content.
  • Bengio's interpretability survey (@askalphaxiv · alphaXiv 2611.08438) [morning]. "Understanding AI from the Inside" is a 182-page end-to-end survey of mechanistic interpretability.
  • Shopify Flywheel (@PyTorch) [morning]. A PyTorchCon keynote on a continual-learning loop that turns production failures into training data and beats frontier models on Shopify's tasks at a fraction of the cost.
  • Stanford CS329Z (@stanfordnlp) [morning]. A new course on building agents from scratch, with the first lecture video online.
  • GPUs idle 95% of the time (@ycombinator) [morning]. A pitch from YC startup Nodus. The 5% utilization figure has no source.
  • Quote-bait and promos [morning]. Skip.
  • No posts [afternoon + evening]. Both slots were empty. See the daily digest and Media Zone, led by LowRAM's KV value amortisation.