AI research wiki
social-stream · 2026-11-07

2026-11-07-morning

Summary

The strongest signal in this window was agent cost cut without touching the model. Three papers landed the same US afternoon and all were picked up by DAIR's curators and amplified by omarsar0 and rohanpaul_ai. EnvOpt rewrites the shared workspace an agent reads. Amazon's weight-memory paper distills user facts into the model. Nace AI's Drex answers bounded choices without generating text. Drex was its own cluster of 4 posts, the most of any item. A second cluster, 3 posts, covered memory supply and the chip bounce: NVIDIA reportedly asked TSMC to plan 3D DRAM while chip stocks recovered part of Thursday's rout. The standout single items were Avi Chawla's explainer on seven MoE routing strategies and HKU's AgentPilot, which lets a 9B model steer a frontier coder. The scheduled slot captures returned nothing and no saved posts arrived, so everything here comes from the Following feed.

Posts

  • Drex, open decision models (cluster of 4) (@omarsar0, @rohanpaul_ai, @rohanpaul_ai on calibration, @dair_ai, paper). Nace AI released four open decision models: 8B, 9B, 35B with 3B active, and a 27B Max. A decision model reads a state (a document, record or agent context) and a question with fixed options, then returns one probability per option in a single forward pass, with no text to parse. Max trails Claude Opus 5.5 by 1.6 points on Decision Bench v4 and ties GPT-6 Sol. It answers in 183 ms for about $0.06 per 1K decisions, against $10.17 and 3.5 s for Opus. Training adds RLAF (reinforcement learning from adjusted feedback): a teacher proposes per-option probabilities, the verified answer adjusts them, and the model trains on these soft targets. The calibration chart shows 70% confidence is right about 70% of the time, which makes the number usable as a gate between acting alone and escalating. The paper also reports 5 to 9 points of over-confidence in four of five domains. See the summary.
  • EnvOpt: optimize the agent's environment (@dair_ai, paper). Berkeley's Sky lab, Chicago and Bespoke Labs argue an agent's cost depends on three things: the model, the harness and the environment. Only the first two get optimized. EnvOpt has a coding agent rewrite what a shared workspace (a Slack, a document store, tickets) holds and how it is presented, with no graders, stopping on a validation set. On AutomationBench, Claude Haiku 4.5 gets 72.7% cheaper and passes 14.34% more tasks. The same environment cuts Opus 5's cost 68.1%. On top of Meta-Harness's tuned harness it saves another 83.4%. See the summary.
  • Weight memory beats retrieval on user facts (@omarsar0, paper). Amazon stores stable user facts in the weights with on-policy self-distillation: the teacher is the same model reading the user's memory, and the student must answer without it. Plain fine-tuning recalls 36.9% of facts. Self-distillation reaches 59.9%, above retrieval at 47.8% and the raw dialogue at 51.3%, at 4B, 9B and 27B. Weight memory forgets older facts and cannot say a fact was never recorded, so a verify-then-regenerate router checks answers against the text store and adds 6.2 points. See the summary.
  • Memory supply and the chip bounce (cluster of 3) (@StockSavvyShay on 3D DRAM, @StockSavvyShay on the bounce, @StockSavvyShay on Qualcomm). NVIDIA reportedly asked TSMC to plan 3D DRAM production, a hedge against a memory market run by Micron, SK hynix and Samsung. The bounce chart shows Credo +10%, Marvell, Lumentum and Coherent +9%, and TSMC and AMD +7%. A post framed Google's $548M Qualcomm stake as a Gemini-on-Snapdragon play. A securities filing shows the shares came from Qualcomm's purchase of Modular, which Google Ventures partly owned. See the summary.
  • Seven MoE routing strategies (@_avichawla). An explainer of token-choice, expert-choice, dynamic, balanced, hash, soft and group-limited routing in mixture-of-experts models. The key point: an even token split does not make an MoE fast. Small per-expert batches and cross-GPU transfers still cost time, and group-limited routing that follows hardware placement cuts that traffic. See the summary.
  • AgentPilot: a 9B strategist for frontier coders (@huang_chao4969, repo). HKU splits the agent loop. A 9B model decides what to try next, how to spend the trial budget and when to stop. DeepSeek V4 writes the code. The small model learns from each trial's outcome. The repo plugs into Codex, Claude Code and DeepSeek Harness but shows no independent benchmark yet.
  • Training efficiency on B300s (@StockSavvyShay). Nebius reports 1.13% more training throughput on 192 B300 GPUs running DeepSeek V3 from one software change. Small, but it supports the claim that the next margin comes from existing hardware.
  • Quantizing linear recurrences (@ChrSzegedy repost). Work on quantizing Mamba-2, GDN and KDA layers, the recurrences replacing attention in hybrid models. Only the opening line was captured.
  • Harness threads (cluster of 3) (@harjtaggar, @berkeley_ai on self-improving harnesses, @berkeley_ai on evolving harnesses). A widely reposted thread blames most harness failures on subagents, because the main agent hands them a lossy summary. Berkeley shared an agent that improves its own harness at deploy time and Ion Stoica's work on evolving the harness alongside the solution, both pitched on cost and time. The posts were truncated.
  • Agent memory stacks (cluster of 2) (@AnnatarXBT, @TheTuringPost). A thread lays out five memory layers (working, episodic, semantic, procedural, forgetting), citing Mem0 at 1,800 tokens per query against 26,000. Zep shipped procedural memory as agent skills. Vendor numbers.
  • Cola2 and text diffusion (@Ronald_vanLoon, Cola2 weights). Posts described a Google DeepMind model that drafts 256 tokens at once and refines them, a 26B MoE with 3.8B active, but the link was a promo page. ByteDance Seed separately opened Cola2, which generates compressed span vectors with diffusion and then decodes tokens. See the summary.
  • Async RL and other research reposts (@berkeley_ai, @rohanpaul_ai on RAG poisoning, @rohanpaul_ai on tokenizers). Score centering alone removes drift in asynchronous RL and often matches stacked fixes. A Google paper finds RAG poisoning mostly works by copying the query. A Stanford paper argues no tokenizer metric is agnostic in every sense.
  • Mochi 535B-A23B finished pretraining (@stanfordnlp). Marin's open MoE finished its main phase on 18T tokens. Long context and SFT come next.
  • Markets, ballots and valuations (cluster of 4) (@rohanpaul_ai on Anthropic, @GaryMarcus, @rohanpaul_ai on WSJ, @StockSavvyShay on jobs). WSJ says investors cut expected Anthropic IPO valuation from about $2T to $1.5T, with some at $1T. Only 1 of more than 40 data center ballot measures passed. Gary Marcus predicted the end of exponential data center investment while use continues through harnessed systems. Payrolls rose 165K against 65K expected, which posters used to push back on AI job-loss claims.
  • Domu raises $15M (@ycombinator). Collections agents that say they collect $186 for every $100 collected by human teams.
  • NVIDIA Auto-Ontology (@AISuperDomain, repo). An open semantic layer that maps business terms and table relations so agents write correct SQL, with an MCP server.
  • Skip. Subscription-cancellation listicles, "paste this prompt" threads, Elon Musk one-liners and robotics gadget reposts.