AI research wiki
social-stream · 2026-10-18

2026-10-18

Summary

Only the morning slot carried signal; the evening slot was empty because the US conversation had barely started at capture time. The day's strongest cluster is agent harnesses and who makes each decision inside them, anchored by JevAgentHarness: a 4B decision model takes an agent's routine tool and argument choices when confident, cutting cost 14.2% with no changed outcomes across 500 tasks. Google's open-sourced Harness², a free harness-engineering course and Harvard/MIT's market-based post-training, where agents bid for each step, round out that theme. A smaller agent-training cluster (NVIDIA midtraining, CoPE for GRPO, Salesforce ideation) is solid but incremental. The single-slot standout for hardware is an unsourced claim that B300 rental prices rose over 20% in a month while neocloud stocks fell. Much of the high-engagement feed was "reverse-engineer any app" hype and is noise.

Posts

  • Decision models inside the harness (cluster of 5) (@dair_ai, @omarsar0, @0x_rody, @dr_cintas, @techNmak · paper) [morning]. JevAgentHarness routes an agent's multiple-choice decisions to a 4B model behind a per-agent confidence threshold: 14.2% cheaper, 500 outcomes unchanged, and AgentDecision-4B beats its teacher on shifted tasks. Google's Harness² and Walking Labs' course ride along. See summary.
  • Agents that bid for each step (@snwiki238337 · paper) [morning]. Harvard/MIT's MPT trains action and bidding policies, splitting credit by bid; swarms improve as agents join and degrade gracefully when they drop. See summary.
  • Agent training recipes (cluster of 3) (@omarsar0, @haofeiyu44, @dair_ai · ideation, CoPE, midtraining) [morning]. CoPE adds one next-actions loss to GRPO for up to 2x sample efficiency; NVIDIA's 500B-token SWE midtraining lifts SWE-bench Multilingual 55.8 to 62.8. See summary.
  • Hy-Memory and research-agent forgetting, recirculating (@dair_ai, @omarsar0) [morning]. Tencent's 3B-active memory writer and Berkeley's discard-history finding kept circulating. Covered in the 10-17 digest.
  • B300 rentals up, neocloud stocks down (@StockSavvyShay) [morning]. Claims B300 rental rates rose over 20% in a month as Nebius, IREN and CoreWeave sold off on depreciation fears. Unsourced.
  • Next-Concept Modeling (@TheTuringPost · paper, code) [morning]. Stanford/Together's coarse-to-fine generation over compressed codes resurfaced with its code release. Covered on 10-16.
  • Product notes (@cohere, @alexandr_wang, @github) [morning]. Cohere's self-benchmarked Embed 5 Fast win over EmbeddingGemma 2, Muse Spark Open specs restated (816B total, 42B active), a Copilot keypad gimmick.
  • Software-engineering automation, two views (@fchollet, @AnatoliKopadze) [morning]. Chollet: more engineers than 2025, mostly not writing code. Claude Code's lead says he builds loops that build agents.
  • Hype and off-topic threads [morning]. "REA reverse-engineers any app", "Compaq cloned IBM with one prompt", free-API lists, career advice, an unverified OpenAI PDF thread. Skip.
  • Evening slot [evening]. No posts. See the 2026-10-18 digest.