AI research wiki
social-stream · 2026-10-19

2026-10-19-afternoon

Summary

The afternoon's own repost scrape was empty, so the slot's Following-feed capture carries the signal (US late evening of 10-18). The strongest signal is agent cost accounting (cluster of 5): a Princeton + Google paper showing a 4B decision model can take over routine agent choices for 81 to 89% lower cost, Cohere showing cheap planner agents get you within 1.4 points of an all-frontier team at 70% lower cost, a token audit of security agents, and two practitioner posts on pushing branch decisions to Jev. Agent memory evaluation is the second thread (cluster of 2), led by an Amazon paper showing QA scores hide a memory system that picks the right update only 20% of the time. Outside that, Meta's co-design paper on AMD training, Anthropic's 5% Public Benefit Fund, and a Turing Post guide to ten looped-transformer papers are worth a click. About half the feed is ads, Musk chatter and Ronald van Loon gadget reposts. Skip all of it.

Posts

  • Decision models take over routine agent choices (@rohanpaul_ai · paper). Princeton + Google's JevAgentHarness hands each multiple-choice step (which tool next) to a fast decision model, gated by a confidence threshold fitted on the agent's own runs. AgentDecision-4B, trained on states the agent actually reaches, made agents 81 to 89% cheaper and up to 2.5x faster within 3 points of success; dropping the threshold cost 19 points. The cleanest routing result in weeks; see LLM routing and the decision-model calibration debate.
  • Jev as the arrows in a workflow graph (cluster of 2) (@imryven · @MaziyarPanahi · dataset). Practitioner recipe: big models do the boxes, Jev answers every keep/drop and retry/stop arrow at about 100 ms and $0.042 per million tokens, so 1,000 triage calls cost 4 cents instead of $4. Panahi released 130K Jev-verified RAG claims. Same pattern as the Princeton paper, from the field side; the per-decision confidence cutoff is the part people skip.
  • Teams of coding agents: cheap planners, strong lead (@rohanpaul_ai · paper). Cohere finds workers that only research and plan match workers used at every stage, at half the cost. A GPT-5.6 Sol lead with 16 cheap Luna planners landed within 1.4 points of an all-Sol team at 70% lower cost, but cheap coders lowered reward. Extends the agent-team cost thread.
  • Where security agents burn tokens (@rohanpaul_ai · paper). Across 200 vulnerability-hunting runs, 60.4% of tokens went to understanding code and reasoning about the trigger. Generic cost tricks kept success at lower cost in only 24.4% of comparisons; their AVRI tool, which returns related code together and caches input-to-sink links, cut cost 18 to 24% with no loss. Measure first, then cut.
  • Agent memory: grade the store, not the answer (cluster of 2) (@rohanpaul_ai · paper · @omarsar0). Amazon's FEB-30 knows the correct memory state after every document. On a Mem0-style system, stored-fact scores and QA scores differed by 16.8 points and disagreed on pass/fail 21% of the time; the right update action was chosen only 20.1% of the time. Elvis Saravia's repost on latent semantic memory rides the same theme. See agent memory.
  • Harness² open-sourced by Google Research (@Xudong07452910 · repo). Agents rewrite their own prompt, skills and memory config, rerun the same task, compare, and keep the lessons for later tasks; up to +11.5 on WorkBuddy-Bench in one round. Trial cost and transfer are unproven. Fits agent harness engineering.
  • "Harness is dead, long live the harness?" (@TheTuringPost). Turing Post's notes from AI Engineer NYC on strong opinions about harnesses. Click through to read.
  • Meta co-designs training for AMD GPUs (@arXivBangers · paper). Co-designing models and software for AMD hardware gave up to 1.5x training throughput over straight ports from NVIDIA. A real data point against the CUDA moat argued in SemiAnalysis's AMD piece.
  • Ten looped-transformer papers in one guide (@TheTuringPost). Covers Recurrent Looped Transformer, LiFT, LoopCD, Hyperloop, LoopQuant and more, with when to use each. Useful map; see looped transformers.
  • Anthropic puts 5% of its shares into a Public Benefit Fund (@VaibhavSisinty). Another 2% can follow if valuation rises 50%. Money goes to nonprofits and displaced workers, no lobbying. Set up before the IPO, so new shareholders buy in knowing the deal.
  • Anthropic staff vs US workers on AI automation (@rohanpaul_ai). Anthropic staff expect AI to do 75% of their current work by 2030; US workers say 28%, and the gap grows with task difficulty. Click through to read.
  • Neocloud credit stress (@rohanpaul_ai). Chamath on why credit default swap spreads on CoreWeave-style neoclouds jumped, tied to US rates. Worth tracking for GPU financing. Click through to read.
  • Nebius puts $10M behind European AI builders (@rohanpaul_ai · program). Compute grants for open models, agent runtimes, inference tooling and RL, plus VC intros and NVIDIA hackathons. Small but a sign of neoclouds buying developer mindshare.
  • ICLR 2027 AC desk-rejects 5 of 14 (@hjy836). Not for LLM use, for basics: undefined jargon, no real analysis, fewer than three citations in an intro. Argues for automated early flagging.
  • Agents as prisoners of their trajectory (@powerpig). Short take: once reasoning enters context, agents can't argue out of it, while humans can but refuse outside input. A crisp framing, no evidence.
  • "Anthropic published 400+ agents and skills" (@romandevz). Mostly community repos (agency-agents, wshobson/agents) mislabeled as Anthropic; only knowledge-work-plugins and claude-quickstarts are official. Engagement bait. Skip.
  • Five PEFT techniques explained (@_avichawla). LoRA-family explainer thread. Beginner material. Skip.
  • Schmidhuber lectures in Tokyo and Kyoto (cluster of 5) (@SakanaAILabs · @hardmaru). Tokyo is full; Kyoto University lecture on Nov 5. Event promo. Skip.
  • NVIDIA GTC Berlin livestreams (@NVIDIAAI). Event promo. Skip.
  • Elon Musk posts (cluster of 8) (@elonmusk). SpaceX shares for kids, Grok hiring pitch, politics, memes, and one line about being "one-shotted by super persuasive super intelligences." No substance. Skip.
  • Ronald van Loon and misc low-signal posts (cluster of 21) (@Ronald_vanLoon). Gadget and robot videos plus link-outs (IEEE Spectrum on chip R&D, Nature on China's AI research); also Miles Brundage asides and a Milanfar quote, Vaibhav Sisinty filler and a one-line rohanpaul_ai aside. Skip.
  • Sponsored posts (cluster of 24). Proxies, trading, crypto, Supermicro, HPE, Figma, food delivery. Skip.