Summary
The US Tuesday conversation, read from the overnight Following-feed capture. No new curated reposts or saved posts arrived in this slot. The strongest signal is a cluster of 3 papers on agent infrastructure surfaced by DAIR-adjacent accounts. Skill++ shows a growing skill library silently breaks routing to older skills. Google's XLA:TPU paper uses agent swarms to fix silent compiler bugs. Alibaba's SPAR evolves a sales playbook in a live A/B test across 200,000+ merchants. A second cluster of 3 posts covers harnesses as products: the Unified Harness Protocol and HarnessRouter, Apple's OpenVLHarness (63% fewer tokens than thinking harder), and Agent37's $4.76/month sandboxes. On hardware, SambaNova's SN50 is heading into Supermicro disaggregated racks, SK hynix bought into Penguin Solutions' KV-cache offload work, and Nvidia is reportedly capping China-facing cloud sales. The weight-space programming story from 11-02 turned into a debate, with a detailed teardown arguing most of the math gain came from fixing a known looping mode. Sebastian Raschka's tally of 10 open LLM architectures in October and a $0.28 cache-heavy agent run round out the day.
Posts
Skill++: lifelong skill routing (@dair_ai · paper). When an agent picks skills by name and description alone, a new skill with similar metadata can replace or crowd out an older one, even though neither the old skill nor the router changed. Skill++ rewrites each incoming skill's description to say what it supports, how it differs from its closest existing skill, and what it excludes. No router training. At 731 skills over ten sessions, full-recall routing rises from 36.4% to 76.3%, across five routers and the Codex CLI. The attached figure shows the substitution and crowding failure modes side by side. Wiki summary.
XLA:TPU compiler bugs fixed by agent swarms (@omarsar0 · paper). Google targets silent miscompilations: no error is raised, but numbers drift or buffers get corrupted, which can ruin a training run. Fuzzing on synthetic and production graphs, checks alongside the autotuner, and bisection down to one op inside a fused kernel feed agent swarms that reproduce and patch. 142+ bugs are resolved and 103 more are under review. Reviewer effort grew faster than fix complexity, so patches go through adversarial pre-filtering. Wiki summary.
SPAR: a self-evolving sales agent with attribution (@omarsar0 · paper). Alibaba's B2B outreach agent on 1688.com keeps per-merchant episodic memory (inner loop) and a shared playbook of strategies written as testable hypotheses (outer loop), with no weight updates. Before changing the playbook, it decides whether a lost deal came from a bad strategy or an outside factor. Live A/B: +15.8% conversion, +35.8% high-intent merchant growth, 4.62/5 expert quality. Wiki summary.
Harnesses as products (cluster of 3) (@akshay_pachaar · HarnessRouter, @marfinxx · OpenVLHarness, @rohanpaul_ai · Agent37). UHP (Unified Harness Protocol) is an open interface for sending a task to any harness (Codex, Claude Code, Gemini CLI, Hermes, Pi) and getting back streamed progress, files and failures. HarnessRouter implements it for 30 harnesses and reports 25-64% cost cuts from swapping only the harness at a fixed model. OpenVLHarness (UCLA + Apple) adds a visual capability layer and multimodal memory. Across 22 benchmarks it beats Codex and Visual Sketchpad at 63% lower token cost, and found that raising effort from xhigh to max cost 2.7x the tokens for +0.7 points. Agent37 (YC) sells month-long gVisor sandboxes per agent at $4.76. The enthusiastic framing ("most dangerous thing I've seen") is engagement bait; the artifacts are real.
SambaNova SN50 in Supermicro disaggregated racks (@SambaNovaAI). Supermicro will build and sell racks that pair NVIDIA GPUs for prefill with SN50 RDUs for decode, starting with 10,000 chips, on standard power and cooling. The press page also notes SambaNova's first $1B close at an $11B valuation. Wiki summary.
Memory and China, two finance-account notes (@StockSavvyShay, @StockSavvyShay). SK hynix now holds 9% of Penguin Solutions after buying SK Telecom's $377M preferred stake, which gives it exposure to Penguin's CXL and SSD-based KV-cache offload. Nvidia is reportedly requiring cloud providers to sell at least 51% of capacity to non-Chinese customers, narrowing the GPU-rental route around export controls.
Weight-space programming, the pushback (cluster of 2) (@dummerspast39, @NielsRogge · HF reproduction). The 11-02 report claimed GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights to lift MATH-500 from 7.9% to 48.1%. The teardown argues the baseline used thinking mode, which Qwen's own model card says loops (93% of answers never finished). Default non-thinking mode scores 47% with zero edits, so the edit mostly suppressed the repeat direction. It also argues the game-network "interpretation" can be read straight off first-layer weights. Hugging Face separately reproduced the edit with its open ML Intern harness. Both things are true: the edit works, and the headline overstated what it shows. Simon Willison reports Qwen 3.8 27B on a laptop can do the no-tools interpretation part too.
"Open October": 10 open LLM architectures in a month (@rasbt). Raschka lists Kolibri-1 (78B, GQA + sliding window), Ling 3.1 Flash (560B, Kimi Delta Attention plus gated MLA), Meta's Muse Spark Open (816B, LatentMoE), Nemotron 3.5 Super VL (120B, Mamba-2 with eight attention layers), Dory1.0 (4B looped transformer), Reflection Beam (501B), Recursive Flint (40B), MiniMax M3.1-Flash (428B, MiniMax Sparse Attention), LFM3 (1.7B, convolution hybrid) and Mistral Large 4 (1T, MLA plus sparse attention). The attached chart shows them side by side. Half mix in linear or sparse attention to cut KV-cache cost. Background.
Cache hits are the agent bill (@omarsar0). The same one-line game-build prompt cost $0.28 on LongCat-2.5 (1.6T total, ~48B active) against $0.82 on Qwen3.8-Max, over 53 turns. 4.7M of 4.8M input tokens were cache hits at $0.006/M versus $0.30/M uncached. Efficiency shorts.
STAIR: retrieval by table of contents (@alex_verem). IBM gives Mistral 7B a document's table of contents and trains it to name the section that holds the answer, with no chunking and no vector database. On 18 textbooks and 103,249 questions it finds the right section on the first try 82.6% of the time (embedding search 68.7%, the same system without the TOC 76.9%). It names an invalid section 0.05% of the time versus 3.25%. The "RAG is cooked" framing is hype; the structured-retrieval result is solid.
Latent Sparse Attention, ported to Qwen 3.8 27B (@NFT_Chen · paper · repo). A Chinese-language walkthrough of the LSA release covered on 11-03. Keys are clustered hierarchically in a learned latent space, so the indexer searches a tree instead of scanning everything. At 10M tokens, single-layer prefill attention is 72.4x faster and decode 42.5x, with time to first token at about 1/9. Distilled indexer weights and an SGLang fork are public. See the 11-03 digest.
Agents writing most code (@Ronald_vanLoon, @addyosmani). Atlassian reports 77% of recent PRs across 6,000 engineers were written by agents, shown live on Jira boards. Addy Osmani's article on critical thinking when agents do the work argues the risk is "cognitive surrender": losing the skills needed to supervise the agents.
Skip: the "Anthropic engineer Andrej Karpathy" and "build a system that prompts itself" engagement threads (misattributed and content-free), a Korean repost of a Claude paper-writing skill repo, conference and book promos, and election-day political posts.