AI research wiki
social-stream · 2026-11-06

2026-11-06-morning

Summary

The strongest signal this morning was the chip rout, a cluster of 4 posts: the Philadelphia Semiconductor index fell 12.3% and Nvidia 9.3% as investors priced in data center moratoria after the midterms. Two launches carried the technical conversation. Thinking Machines' Inkling 2 Small, a cluster of 4 posts, drew real respect for how little compute it took. OpenAI's GPT-6 Cyber, a cluster of 2, showed what was behind the Linux kernel exploits reported the day before. METR's report that Anthropic's model progress has roughly doubled, with most of the evidence redacted, became its own debate in 3 posts. NVIDIA's compaction paper kept traveling across languages, now in Korean and Chinese threads. Hugging Face's long MoE handbook was the most useful single artifact, and a small note on kernel optimization from Soumith Chintala pointed to an emerging benchmark race. No new reposts or saved posts arrived, so everything here comes from the Following feed.

Posts

  • The chip rout (cluster of 4) (@rohanpaul_ai, @StockSavvyShay, @rohanpaul_ai on the WSJ, @GaryMarcus). The Nasdaq fell 5.52% and about $3.2T of market value was reported lost on 11-05. SanDisk fell 22%, and Micron, Western Digital and AMD 17% each. Goldman Sachs estimates short-term moratoria could affect about 45% of planned US data center capacity, and Meta's president says every hyperscaler is considering sites abroad. One account calls it the second-worst semiconductor day since 1994 and notes that such days have often come near bottoms, while earnings still show demand above supply. Marcus's only upside: DRAM, GPU and console prices may fall. Background in the hardware summary.
  • Inkling 2 Small (cluster of 4) (@soumithchintala, @eliebakouch, @arena, @soumithchintala on kernels, announcement). A 594B-parameter mixture-of-experts model with 17B active. Pretraining's first phase ran 68 hours on GB300s, then post-training took 17 million rollouts over 9 days, and every stage used NVFP4 (4-bit floats). Elie Bakouch reports its base-model loss on internal code beats Kimi K3.1's after a fair adjustment, with far fewer tokens. Arena says it would rank #1 among US open models on Code Arena WebDev once weights release. Soumith says weights come next week after external safety testing. The kernel screenshot shows it solving long GPU optimization tasks for hours. See the summary.
  • GPT-6 Cyber (cluster of 2) (@OpenAI, @rohanpaul_ai, post). OpenAI's security model reproduced 1,724 Linux kernel bugs and built working attacks for 948, against 166 for GPT-5.6 Sol. It found 81 new serious Firefox bugs. In honeypot tests it ignored out-of-scope bait 99.93% of the time without a deployment prompt, up from 94.49%. Only select Daybreak Red customers get it for now. See the summary.
  • METR on Anthropic's acceleration (cluster of 3) (@rohanpaul_ai, @Miles_Brundage sharing Casey Newton, @GaryMarcus). METR's 11-week review puts progress at about twice the earlier pace, with about 80% attributed to Anthropic's own AI use and none to faster training compute growth. The debate is about redactions. Anthropic removed most of the evidence METR found most telling, and critics mocked a cofounder's safety defense of that. METR report.
  • The MoE Training Handbook (@Thom_Wolf, handbook). Hugging Face spent 800K GPU hours on the questions MoE training teams actually hit: how many experts and how big, why half a step goes to all-to-all communication, which router activation to use, and whether to keep the router in fp32. Wolf reports 18 of the top 20 open models are MoEs. About 70K words, with 55 interactive figures and runnable code.
  • "You don't need long context" keeps spreading (cluster of 3) (@itsharmanjot, @pinesun11, @Xudong07452910). Threads in English, Korean and Chinese repeat NVIDIA's finding that agents restarting from a short handoff note barely lose accuracy, while late tokens cost 3.6-7x more. The Korean post links it to Claude Code replacing a status line with a short note. Covered in the 11-05 summary.
  • Claude Code status line fix (@b1rkh0ff). Background subagents were re-reading the whole conversation every 30 seconds to update a status line. The reported fix cuts that about 99%, freeing roughly 5% of a usage limit.
  • Marin's Mochi finishes pretraining (@dlwh, tracker). A 535B open MoE completed 18T tokens with a public run log. Long context runs this weekend, then SFT (supervised fine-tuning), with checkpoints hoped for next week.
  • Decision models as cheap judges (cluster of 2) (@kalyan_kpl, @webbigdata). One post lists RAG stages where a decision model can replace an LLM judge: query routing, chunk filtering, claim checks and answer checks. The other reports an Airbnb test where a fine-tuned Gemma 4 beat Jev, which cost about 5x at thousands of requests per minute and 4x more on long prompts with no cache discount.
  • Copilot Fusion (@github). GitHub's router picks among Single, Cascade and Critique patterns from one effort setting. See LLM routing.
  • HarnessGrad (@zxlzr, repo). A platform crossing 8 harnesses, 19 self-improvement methods and 56 benchmarks, with the model held fixed so only the harness learns.
  • Research-agent benchmarks (cluster of 3) (@omarsar0, @rohanpaul_ai, @ycombinator on DeepFinance). Science's Last Exam uses 113 tasks with hidden checks and scores hypotheses against studies published after training cutoffs. DeepSeek V4.1 Flash led at 21.28%. DeepFinance targets long high-finance work.
  • Muse Realtime Avatar (@AIatMeta, @rohanpaul_ai). Live video calls open to US users. A student model does in 2 passes what a teacher did in 120, and 4-bit quantization with caching lets one GB200 run 12 streams.
  • Monofacts and scientific discovery (@rohanpaul_ai, arXiv). A Johns Hopkins paper finds models recover more held-out materials and DNA sequences as the share of once-seen training facts rises, and repeat epochs erase the effect. More discovery comes with more hallucination.
  • Neocloud financing (@StockSavvyShay). IREN's new contract has the customer prepaying 90% of GPU costs at about $22M per MW, against 20% on Microsoft tranches.
  • Skip. Engagement bait on "AI therapy trauma," "MIT mass released a free library," repackaged harness courses, and NeurIPS party invites.