AI research wiki
social-stream · 2026-11-04

2026-11-04

Summary

All of today's signal came in the morning slot. The afternoon and evening slots were empty, so there are no cross-slot clusters. The strongest thread is agent infrastructure: Skill++ shows a growing skill library quietly breaks routing to older skills, Google uses agent swarms to fix silent XLA:TPU miscompilations, and Alibaba's SPAR evolves a sales playbook in a live A/B test across 200,000+ merchants. A second cluster treats harnesses as products. HarnessRouter reports 25-64% cost cuts from swapping only the harness at a fixed model, and OpenVLHarness beats Codex at 63% lower token cost. The standout for efficiency is the $0.28 LongCat-2.5 agent run, where 4.7M of 4.8M input tokens were cache hits, which shows that caching sets the agent bill. The weight-space programming teardown is the best skeptical read of the day, and the "RAG is cooked" and Karpathy-misattribution threads are noise.

Posts

  • Skill++: lifelong skill routing (@dair_ai · paper) [morning]. Rewriting each new skill's description against its nearest existing neighbor lifts full-recall routing from 36.4% to 76.3% at 731 skills, with no router training. Wiki.
  • XLA:TPU compiler bugs fixed by agent swarms (@omarsar0 · paper) [morning]. Fuzzing plus op-level bisection feeds agent swarms that fix silent miscompilations. 142+ bugs resolved, with reviewer load as the new bottleneck. Wiki.
  • SPAR: self-evolving sales agent (@omarsar0 · paper) [morning]. A memory plus hypothesis playbook with no weight updates checks whether a strategy or an outside factor lost the deal. Live A/B shows +15.8% conversion. Wiki.
  • Harnesses as products (cluster of 3) (@akshay_pachaar · HarnessRouter, @marfinxx · OpenVLHarness, @rohanpaul_ai · Agent37) [morning]. UHP routes tasks across 30 harnesses at 25-64% lower cost, OpenVLHarness gets 63% fewer tokens than thinking harder, and Agent37 sells $4.76/month sandboxes. The hype framing is bait, but the artifacts are real.
  • SambaNova SN50 in Supermicro disaggregated racks (@SambaNovaAI) [morning]. The racks pair NVIDIA GPUs for prefill with SN50 for decode, starting at 10,000 chips. SambaNova also notes a $1B close at an $11B valuation. Wiki.
  • Memory and China notes (cluster of 2) (@StockSavvyShay, @StockSavvyShay) [morning]. SK hynix takes 9% of Penguin Solutions for exposure to its KV-cache offload work. Nvidia reportedly requires cloud providers to sell at least 51% of capacity to non-Chinese customers.
  • Weight-space programming pushback (cluster of 2) (@dummerspast39, @NielsRogge · HF reproduction) [morning]. The teardown says the baseline used Qwen's looping thinking mode, while non-thinking mode scores 47% with zero edits. The edit reproduces, but the headline overstated what it shows.
  • "Open October": 10 open LLM architectures (@rasbt) [morning]. Raschka's side-by-side tally of 10 open releases in October. Half mix in linear or sparse attention to cut KV-cache cost. Background.
  • Cache hits are the agent bill (@omarsar0) [morning]. The same prompt cost $0.28 on LongCat-2.5 against $0.82 on Qwen3.8-Max. 98% of input tokens were cache hits at $0.006/M. Wiki.
  • STAIR: retrieval by table of contents (@alex_verem) [morning]. IBM trains Mistral 7B to name the right section from a TOC, with no vector database. It gets 82.6% first-try accuracy versus 68.7% for embedding search. The result is solid, but the "RAG is cooked" framing is hype.
  • Latent Sparse Attention on Qwen 3.8 27B (@NFT_Chen · paper · repo) [morning]. A Chinese walkthrough of the release covered on 11-03. A tree-searched latent key index gives 72.4x faster prefill at 10M tokens. See the 11-03 digest.
  • Agents writing most code (cluster of 2) (@Ronald_vanLoon, @addyosmani) [morning]. Atlassian says agents wrote 77% of recent PRs across 6,000 engineers. Osmani warns of "cognitive surrender," where engineers lose the skills needed to supervise agents.
  • Engagement bait and promos [morning]. The Karpathy-misattribution and "self-prompting system" threads, a Korean skill-repo repost, conference and book promos, and election posts. Skip.

See the 2026-11-04 digest for the full synthesis.