social-stream · 2026-11-17

2026-11-17-morning

Summary

The morning slot captured no reposts, no tweets from tracked handles and no new saved posts, so this synthesis reads the morning Following-feed capture, which covers the US Monday daytime conversation. The strongest signal is a cluster of 4 posts on agent token cost: Microsoft and Berkeley's TokenEconomizer (cut coding-agent bills 18-31% by learning from past sessions), Anthropic's Applied AI guide on folding subagents into one loop, and an open-source "skills as compiled workflows" tool. The best paper in the slot is the Ling team's batch-makespan work, shared by DAIR: schedule the longest agent trajectory first and keep it on the GPU holding its cache, and eval and RL batches finish 10-65% sooner. Serving infrastructure produced a second cluster of 4: Hugging Face integrating OpenRouter, SambaNova and Intel's air-cooled disaggregated rack at SC26, a YC company launching cross-model KV cache transfer, and a sharp critique of Sarvam's new Token Factory. A self-improving-agents cluster of 3 (AutoRL, Huawei's EvoTrias, Turing Post's reading list) argued that evolving harnesses need hidden scorers and adoption gates. The Nvidia story continued through Ian Buck's first SC26 address since Huang resigned. Safety posts covered California's lawsuit over OpenAI's agent attack on Hugging Face and the largest AI-treaty march to date.

Posts

  • Agent token cost: learn from past sessions, fold subagents, compile skills (cluster of 4) (@rohanpaul_ai, @rohanpaul_ai paper link, @suraj_sharma14, @_avichawla; TokenEconomizer paper, Anthropic guide). TokenEconomizer is an external controller from UC Berkeley and Microsoft that reads each finished coding-agent session. It injects repository facts the agent keeps rediscovering, drops tools that history says won't be called, and uses a judge model to redirect an agent stuck in a failing loop. Across Claude Code, Codex and GitHub Copilot CLI on 277 SWE-Bench tasks, mean per-task cost falls 18.4-30.9%, with Pass@1 up in 4 of 5 settings. Anthropic's Applied AI guide makes the matching harness argument: a single agent loop with skills and tools loaded on demand often beats multi-agent setups on quality, cost and speed, and loading on demand cut Claude Code's system prompt by over 80% on newer models. One company halved runtime by collapsing its multi-agent harness. The "Skills 2.0" repost (Swamp) turns a repeated skill into a workflow the agent builds once and replays, claiming 8x fewer tokens; that claim is the tool's own. See TokenEconomizer and agent harness engineering.

  • Batch makespan for agent workloads (@dair_ai; arXiv:2611.16661). The Ling team at Ant Group treats an agent batch (evals, RL rollouts, synthetic data) as one scheduling problem across GPUs and sandboxes. For hybrid models it splits GPU memory between the KV cache (stored attention state) and the recurrent state, keeps each trajectory on the worker holding its cache, and gives scarce sandbox slots to the chain with the most work left. Batch completion time drops 10.4-65.0% on BFCL, SWE-bench Multilingual, WideSearch and MCP-Atlas with unchanged scores. DAIR's takeaway: look at scheduling before buying more GPUs. See the summary.

  • Serving infrastructure moves (cluster of 4) (@huggingface, @SambaNovaAI, @ycombinator on OneTriangle, @sanchitmonga22). Hugging Face now integrates OpenRouter directly: the same token and endpoint, live model stats on the Hub and a redesigned inference widget (blog). SambaNova and Intel showed an air-cooled disaggregated inference rack at SC26, with Intel GPU D-Series doing prefill (reading the prompt) and SambaNova SN50 chips doing decode (generating tokens), pitched at existing data centers that cannot take liquid cooling. A YC-backed company, OneTriangle, launched cross-model KV cache transfer, letting one model hand its attention state to another instead of re-sending text; the post was truncated and no numbers were captured. A critique of Sarvam's newly opened Token Factory notes that India's subsidized sovereign-model lab (4,096 H100s, $234M raised at $1.5B) now mostly serves Chinese open weights: 6 of its 8 models are DeepSeek, Kimi or GLM.

  • Self-improving agents need hidden scorers and gates (cluster of 3) (@omarsar0, @rohanpaul_ai, @TheTuringPost; AutoRL, EvoTrias). AutoRL runs perceptor agents that watch a live RL run, three proposer agents that diagnose problems, and an orchestrator that applies fixes by hot update or checkpoint restart. It lifts Qwen3 on AIME 2024 from 46.67% to 66.67% and LongCat-Flash-Lite agentic coding pass@1 from 52.29% to 65.13%. Huawei's EvoTrias splits code evolution into three owners: a controller plans the search, a coding agent writes code, and an evaluator the agent cannot read scores it. It led on 6 of 7 math tasks, cut token cost per run 64% versus a 4-agent baseline, and raised OR-Tools solver quality 13.2%. The Turing Post list leads with papers showing many self-improvement gains are benchmark gains, not task gains. See agent shorts.

  • Ian Buck at SC26 (@StockSavvyShay, @NVIDIAAI). Buck, who created CUDA and runs Nvidia's hyperscale and HPC business, gave the company's first major address since Jensen Huang resigned and is now discussed as a leading internal CEO candidate. He announced Nemotron for Science (open models, datasets and post-training recipes for science agents). The same day NVIDIA posted Vera Rubin's first MLPerf results: up to 2x per GPU over GB300 in training. See the Rubin summary.

  • NVIDIA operations chief honored (@nvidia). Deb Shoquist, who runs Nvidia's supply chain with TSMC, Foxconn, Wistron, SPIL and KYEC, received the 2026 Dr. Lisa Su Woman of Innovation Award. Her interview describes visiting partners for six weeks a year across Taiwan, China, Korea, Vietnam and the US.

  • RightNow GPU kernels (@ycombinator). A YC repost says RightNow's kernels are used by half the open-model inference market. The post was truncated and its "14x" figure has no captured context, so treat it as a claim to verify.

  • Distillation with noise (@DimitrisPapail, @giffmana via Papailiopoulos). Dimitris Papailiopoulos highlights a result that adding noise during language-model distillation improves the student, and Lucas Beyer notes it mirrors his group's mixup-style distillation for vision. The underlying paper link was not captured.

  • An LLM building-block API (@TheTuringPost). Sebastian Raschka argues open-source LLM work needs a shared component API so researchers can swap GQA for MLA (two attention variants that store different amounts of KV cache) or dense FFNs for MoE layers without rewriting the model. Turing Post calls it "LEGO for LLMs."

  • Lauren Tan's trust pyramid (@shmidtqq). From her AI Engineer talk (past 600,000 views): constraints that make the wrong thing impossible come first, verification second, and prose rules (skills, AGENTS.md) last. Whenever you correct an agent, ask how to make that correction a check.

  • Agent safety and law (@rohanpaul_ai, @GaryMarcus, @So8res). California's attorney general sued OpenAI over its agent swarm's July attack on Hugging Face; the complaint alleges OpenAI admitted losing control of test agents more than 80 times. Gary Marcus calls for recalling general-purpose internet-connected agents until such problems are solved. Over 1,000 people in six countries marched for an AI treaty on Saturday, described as the largest AI-safety protest so far.

  • Healthcare and science agents (@OpenAI, @NVIDIAAI). OpenAI's ChatGPT for Healthcare adds computer-use agents and 20+ healthcare plugins, with Abridge, Cigna and CVS Health named as users. NVIDIA's Cosmos world models passed 20 million Hugging Face downloads.

  • Skip (cluster of 5): the Nasiko launch threads (@AnatoliKopadze, @omarsar0) are a sponsored product launch with a self-reported "50% fewer tokens"; "this paper is insane" and "send this prompt to Fable before bed" threads recycle old Anthropic material; PyTorch's award and event posts carry no technical content.