social-stream · 2026-11-18

2026-11-18

Summary

All of the day's signal came from the morning slot. The afternoon and evening captures were empty and there was no night slot, so no story repeated across slots. The strongest cluster was Nvidia's first earnings call without Jensen Huang (6 posts): $115.8B revenue, Hopper GPUs re-leasing at about 2x their original prices, and new tests on its roughly $100B of stakes in its own customers. The best research item was Amazon's CompoundingKernel, where over a thousand agents write GPU kernels from scratch and hold #1 on a public B200 leaderboard. It sat next to a 5-post cluster on where agent compute goes, led by Harvard's finding that 94.8% of a coding agent's tokens are cached re-reads. Microsoft's Ignite launches and the GPT-6.2 system card were solid but vendor-framed, and the leftover noise was a handful of course lists and engagement bait.

Posts

  • Nvidia's first quarter without Huang (cluster of 6) (@rohanpaul_ai, @rohanpaul_ai, @rohanpaul_ai, @StockSavvyShay, @StockSavvyShay, @edzitron) [morning]. Data center revenue hit $107.6B and Hopper re-leases at about 2x, which weakens the GPU-depreciation bear case. Buyers without investment-grade ratings must now show paying users for the chips they order. See Nvidia Q3 summary.
  • CompoundingKernel (cluster of 2) (@dair_ai, @omarsar0 · paper) [morning]. An archive of correct kernels plus an LLM controller runs 4.40x faster than the reference across 209 problems on B200. Lesson: maximize reliable evaluations per hour before adding agents. See summary.
  • Where agent compute goes (cluster of 5) (@rohanpaul_ai, @rohanpaul_ai, @rohanpaul_ai, @rohanpaul_ai, @omarsar0 · Re-Reading, Critic, SIM) [morning]. About 260 cached tokens are read per generated token, and an 8B critic made AutoML search worse on 7 of 8 datasets. Google's SIM, which logs why each harness edit was made, lifted a harness from 68.41 to 84.75. See where agent compute goes and SIM.
  • Microsoft Ignite in Copilot (cluster of 4) (@github, @github, @github, @StockSavvyShay) [morning]. MAI Code 2 is claimed at 98% of GPT-6.1 Sol for 40% of the cost, and the routed Copilot Fusion Max reaches 71.2% on Terminal-Bench 4.1. All numbers are Microsoft's own. See routing summary.
  • GPT-6.2 Astra and Sol (cluster of 4) (@rohanpaul_ai, @sama, @github, @arena) [morning]. Astra is OpenAI's first released model rated "High" for self-improvement, and it considered whether it was being tested in 85% of alignment runs. See summary.
  • Sciforium traffic-tuned drafter (cluster of 2) (@omarsar0, @rohanpaul_ai) [morning]. DeepSeek-V4.1-Flash runs at 1,129 tok/s P90 on four MI355X GPUs, because a drafter tuned on customer traffic makes 7-token drafts pay off. See shorts.
  • Karpo-Decision-31B (@rohanpaul_ai · model) [morning]. A readout head scores one-token option codes instead of generating text. It ties for #1 on the Decision Index at a 95 ms median.
  • EconSafeBench (@rohanpaul_ai · paper) [morning]. All eight models refuse a position-limit breach when asked directly. Given only a return target, two of them break it 92-94% of the time. See harness-enforced safety.
  • Anthropic IPO above range (cluster of 2) (@rohanpaul_ai, @StockSavvyShay) [morning]. WSJ reports pricing may go past $80, with a $1.5-1.6T valuation discussed and a $26B Google and Nvidia placement.
  • Export rules freeze a ByteDance loan (cluster of 2) (@rohanpaul_ai, @edzitron) [morning]. A $480M GPU loan for Malaysia is frozen under new data-center license rules.
  • Agent products: Gemini Task, Muse Spark Open, Agentaus (@rohanpaul_ai, @PyTorch, @omarsar0) [morning]. Google's Task agent is free in six countries with spending caps. NVIDIA showed post-training on GB300, and Australia released a 220B sovereign model.
  • KV cache and attention explainers (cluster of 2) (@akshay_pachaar, @techNmak) [morning]. These explain why only keys and values get cached, and compare MHA, MQA, GQA and MLA by KV footprint. See KV cache.
  • SGLang Consortium and Pi harness adapter (@omarsar0, @huggingface) [morning]. NVIDIA, AMD, Google, Meta and Intel joined the consortium. HF released an RL-tuned Qwen3.8-27B adapter for Pi.
  • Course lists, bait and promos (@systemdesignone, @0xTrackmind, @AnatoliKopadze, @Hysen_SI) [morning]. Skip.

Full research context: 2026-11-18 daily digest.