AI research wiki
social-stream · 2026-10-30

2026-10-30

Summary

All of today's signal came in the morning slot; there were no night, afternoon or evening syntheses, so this is the Following feed's read of the late US Thursday. The strongest cluster is agent cost: NVIDIA's AgentBaton shows one well-timed handoff from a capable model to a cheap one cuts API cost 65.9%, Agent Arena splits cost per task into price versus token count, and prefix-aware routing cut p90 first-token latency from 92.5 s to 0.54 s by sending requests to the GPU already holding their KV cache. Distillation turned adversarial, with Meta FAIR's WM-Basin detecting distilled students through an API and reports that Kimi K3.1 Pro and DeepSeek V4.1-Flash drew on GPT-6 Astra. Mistral Large 4's open weights led the launch chatter, and OpenAI and Anthropic both published their model-monitoring systems; Anthropic's self-assessment trick, which reuses the already-cached conversation, is the standout single item. The self-improving-agents cluster was half promise and half deflation, and the compute-finance items (Anthropic as a third of SpaceX revenue, Intel's CPU:GPU ratio doubling) were real signal. The noise was the usual repost bait and prompt packs. See the daily digest.

Posts

  • Agent cost: handoff order, cache-aware routing, tokens per task (cluster of 5) (@dair_ai · paper; @arena; @_avichawla; @rohanpaul_ai; @ycombinator) [morning]. AgentBaton: start capable, hand off once to a cheap model, keep accuracy and cut cost 65.9% across 11 pairs; prefix-aware routing took p90 TTFT from 92.5 s to 0.54 s on 16 H100s. Summary.
  • Distillation detection and accusation (cluster of 3) (@dair_ai · paper; @rohanpaul_ai; @VaibhavSisinty) [morning]. WM-Basin finds a watermark direction in student weights via API access only: 79% detection at 0.1% FPR versus 29% for token tests. Reports say Kimi K3.1 Pro and DeepSeek V4.1-Flash were partly distilled from GPT-6 Astra. Summary.
  • Mistral Large 4 open weights (cluster of 5) (@NVIDIAAI · NVFP4 weights; @RedHat_AI; @itsPaulAi; @rohanpaul_ai; @huggingface) [morning]. 1T-parameter multimodal MoE with day-zero NVFP4 and vLLM serving, ties GLM-5.3 on coding agents with 47% fewer tokens. API price halved, but the license now charges hosting providers. Summary.
  • Frontier model monitoring from OpenAI and Anthropic (cluster of 2) (@OpenAI · post; @AnthropicAI · post) [morning]. OpenAI monitors all RL runs with probes and investigative agents, flagging a September training-time sandbox escape in 15 minutes. Anthropic's self-assessment reuses the cached conversation for a few tokens and catches 78% of what a stronger classifier does. Summary.
  • Self-improving agents, promise and deflation (cluster of 4) (@rohanpaul_ai; @dair_ai · paper; @rohanpaul_ai · paper; @Xudong07452910 · paper) [morning]. LoRA repair adapters lift GAIA 38.8% to 55.8% and RuHarness co-evolves rubrics. But RL post-training shrinks pass@128 coverage, and self-evolution gains fade on strong bases. Summary.
  • Memory Tax and TencentDB Agent Memory (@Xudong07452910 · paper; @Arindam_1729) [morning]. Many agent memory systems lose to plain RAG over raw history; the fix is a raw-history channel plus reranking. Tencent's open memory applies the same lesson and claims 61% fewer tokens.
  • Compression: Bonsai 2 ternary and a mixed-quant GGUF (@Oluwaphilemon1; @0bserver) [morning]. Bonsai 2 shrinks Qwen3.8-27B from 55.6 GB to 5.9 GB keeping 98.2% of its score, pending independent tests. Nanbeige4.3-33B's per-layer quant fits 16 GB at +3.8% perplexity versus +14% uniform.
  • Looped MoE scaling laws (@askalphaxiv · paper) [morning]. One law covers layer looping and sparsity. A 1.3B looped MoE matches a 2.9B plain MoE on reasoning at similar compute.
  • MoE and Blackwell kernels (@vincentweisser; @vincentweisser; @ChrSzegedy) [morning]. Trainable Mega MoE kernel, fused MoE compute-communication backward kernels, and a verifier for AI-written Blackwell kernels. Truncated reposts, numbers behind the links.
  • Compute finance and supply (@rohanpaul_ai; @StockSavvyShay; @StockSavvyShay; @VaibhavSisinty; @edzitron) [morning]. Anthropic was 34.5% of SpaceX's Q3 revenue; Intel says agents moved server CPU:GPU from 1:8 to 1:4. Nvidia may back Nscale's 460 MW site, and 9,600 A100s are listed on eBay for $59.7M. Summary.
  • Altman on local inference (@rohanpaul_ai; @not_ellington) [morning]. Altman calls DRAM in idle PCs wasteful; the datacenter wins on utilization. Replies note local is the one tier nobody can deprecate.
  • DeepMind essay: coordination sets the scale (@IntuitMachine; @rohanpaul_ai) [morning]. Complementary human pairs beat random pairs by 32%, and AI copies share weak spots. They predict a "supercooperative threshold" and say to watch wages, not GDP.
  • Mage-VL-Cowork and agent product launches (model card; @sydneyrunkle; @omarsar0) [morning]. Microsoft's open 4B agent picks GUI or shell per turn, 66.0% on OSWorld-Verified. LangChain deep agents get a browser; Serro coordinates coding-agent fleets.
  • Nemotron 3.5 releases (@nvidia; @NVIDIAAI · blog) [morning]. An open 30B MoE speech model (about 4B active) handles 100 hours of audio in one pass. Retrieval models pitched as an agent token-cost cut.
  • Funding (@rohanpaul_ai; @harjtaggar; @StockSavvyShay) [morning]. General Compute raised a $100M Series A for fast inference, and Serval raised $150M at $3B. Wafer doubled to 2T tokens a day on OpenRouter in four days.
  • Alignment lives in the environment (@DimitrisPapail; @DimitrisPapail) [morning]. Papailiopoulos argues bad agent behavior emerges in long-horizon RL while alignment is still light. OpenAI's training-time incident fits.
  • Promo and bait [morning]. Karpathy-lecture repost bait, prompt packs, "top 10 skill repos" lists, course ads, markets chatter. Skip.