AI research wiki
social-stream · 2026-11-01

2026-11-01-morning

Summary

A thin Sunday morning. There were no new curated reposts or saved posts, so this slot reflects the late-night Following-feed capture, which covered the US Saturday afternoon. The strongest signal is a cluster of 3 posts on routing and gating inside agents: Meta FAIR's Routed Continual Learning (route each task to a cluster of past runs, cutting API cost 65-72%), IBM's JevAct (a prefill-only gate that checks tool calls before they run), and a viral repackaging of Claude Code's advisor mode. The second cluster, 4 posts, is compute economics: B300 rental prices up 52% since May while neocloud stocks fell 46-75%, Baseten raising about $2B at up to $30B, CoreWeave's equity-to-GPU leverage, and Meta's 26GW capex path. Standout single items are a clear explainer of why LLM outputs depend on batch size, a DeepMind paper on RL-trained psychiatric intake, and Gary Marcus's report from meeting Pope Leo XIV. Full analysis is in the 2026-11-01 digest.

Posts

  • Routing and gating inside agents (cluster of 3) (@dair_ai, @omarsar0, @xsmryze; papers 2610.38796, 2610.38620). Routed CL (Meta FAIR) shows that one global history of past agent runs gains only +0.015 normalized reward over no memory when tasks are interleaved. It groups past trajectories into clusters, each with its own memory system, and has an LLM router pick or create a cluster before each run: +15-30% of max reward and 65-72% lower API cost on 301 CL-Bench plus 97 τ³-Bench banking tasks. JevAct (IBM) compiles a domain policy into code checks plus yes/no questions scored by a prefill-only decision model, and blocks tool calls that fail; Opus 4.6's three-run τ²-bench Telecom success rose from 39.5% to 80.7%. The third post is hype-framed but points at a real feature, Claude Code's advisor (--advisor opus): Sonnet runs the session, Opus is consulted before a plan locks, after a repeated error and before finishing. Routed CL summary · JevAct summary
  • Why your logits depend on other people's requests (@_avichawla). Floating-point addition is not associative, and GPU kernels split long sums differently depending on batch size to keep all 132 H100 compute units busy. So batch size changes summation order and the low bits of the logits; when two tokens are nearly tied (12.3417 vs 12.3415), greedy decoding can flip and the whole completion diverges. Fixed-order "batch-invariant" modes exist in vLLM (beta), SGLang (25-45% slower) and DeepSeek's DeepInfer. Use them for evals and RL rollouts. Summary
  • Compute economics (cluster of 4) (@StockSavvyShay, @StockSavvyShay, @StockSavvyShay, @rohanpaul_ai). B300 rents are up 52% since May and H100 rents are rising while token prices fell by over half; IREN still secured $363M against contracted GPUs and Nebius expects $9B+ in prepayments. CoreWeave turns about $8M of equity into $40M of GPUs by borrowing against customers' contracted payments. A Wedbush model has Meta at 26GW by 2030 on about $1.2T capex and $171B yearly depreciation, the backdrop to S&P's negative outlook. Bloomberg reports Baseten raising about $2B at up to $30B, with run-rate revenue up from $200M to $600M. Investor-account framing; treat the stock calls as opinion and the rent and funding figures as the signal.
  • RL on simulated psychiatric patients (@rohanpaul_ai, paper 2610.40476). Google DeepMind trained Gemini 3.1 Flash-Lite with multi-turn RL against simulated patients who reveal key facts only after rapport. Psychiatrists preferred it over its base model in 76.7% of 60 matched pairs, and missed acute-risk escalations fell from 30.5% to 12.9% on 256 adversarial cases, still about 1 in 8.
  • Report from the Vatican (@GaryMarcus, post). Marcus endorses Pope Leo XIV's five points, including keeping life-and-death decisions with humans and banning lethal autonomous weapons at November's Geneva talks. His own remarks argue LLMs are correlation-following prediction engines that fail at instruction-following, and that open-ended agents, not narrow tools, cause the current incident wave. Summary
  • Personal agents that know when to defer (@rohanpaul_ai, paper 2610.07588). Meta's PAMO trains a personal agent to adjust one app's recommendations using your history from other apps, rewarding answers that rely on that history only if ranking accuracy holds. A 4B agent reached 47.9% top-10 hit rate on Steam users with Nintendo and Xbox history, versus 40.8% for the platform.
  • Meta's humanoid Max (@StockSavvyShay). Meta's household humanoid runs on NVIDIA Jetson Thor; The Information reports it was built with China's Fourier in three months.
  • Measuring agents by outcomes (cluster of 2) (@omarsar0, @hammer_mt). Elvis Saravia argues self-improving agents need evals first; a reply argues the next evals are market outcomes, like Vending-Bench-style business runs.
  • Chain of thought has two jobs (@rao2z). Subbarao Kambhampati: a post-hoc rationalization can be fine for interpretability but fatal for faithfulness, and the field uses CoT for both without distinguishing.
  • DeepMind PlanAssist (@GoogleDeepMind). A Gemini tool for UK planning officers that cuts review time on householder applications by up to 58%, expanding to 12 more councils.
  • Anthropic's enterprise ceiling (@rohanpaul_ai). Chamath: about 10,000 companies spend $100K-$1M a year with Anthropic and only about 1,900 above $1M, because bigger contracts need proven ROI.
  • Skip: "end of software engineering" paper hype, a "Karpathy 2-hour lecture" engagement post, NVIDIA's Halloween SETI promo, election and space-economy posts.