AI research wiki
social-stream · 2026-11-02

2026-11-02-evening

Summary

No curated reposts this evening, so the slot is the US Monday Following feed, and the strongest signal is Anthropic's amended IPO filing (cluster of 3). Its first Q3 numbers show about $17B in revenue but a swing from a $974M operating profit to a $567M loss, caused by $5B of pricey short-term compute, plus a now-filled compute commitment of $895B. That fits a wider infra-finance cluster of 8: Marvell buying Lumentum for about $115B because laser capacity is sold out to 2029, GPU lead times up to 45 weeks as Rubin ramps, UBS saying DRAM supply covers under a third of agentic demand, and CoreWeave and Oracle debt under strain. The research picks are a Meta and MIT loop-scaling paper (more depth from reused layers, now parallel at decode), RAGFlow's finding that index structures save tokens rather than accuracy, a CoT-controllability paper that traces the limit to the first few reasoning tokens, and Meta's TextSeal2 watermark, which also works with speculative decoding. The rest is Altman soundbites, finance noise, "this is insane" bait and ads.

Posts

  • Anthropic Q3 in the amended S-1 (cluster of 3) (@rohanpaul_ai · SEC filing, @StockSavvyShay, @StockSavvyShay). Revenue about $17B (+47% QoQ). Compute rose to 58% of revenue and adjusted margin fell from 19% to 15%. The commitment line is $895B, with $447B due before 2030, $214B cancellable and more than $400B mapped to Google TPUs. Spot compute is the margin killer, and Anthropic says it will buy more. See Anthropic IPO and compute economics.
  • Marvell buys Lumentum for about $115B (cluster of 4) (@StockSavvyShay, @StockSavvyShay, @StockSavvyShay, @StockSavvyShay). All stock. Lumentum's indium phosphide lasers already feed NVIDIA optical switches and are sold out to early 2029, so a fabless chip company is buying scarce optics supply. Both laser stocks sold off on the news. A clear sign that copper-to-optics is now the interconnect bottleneck.
  • GPU and memory supply (cluster of 2) (@rohanpaul_ai · TrendForce, @StockSavvyShay). Top-end GPU lead times rose 5 weeks to 45 as Rubin starts, with hyperscalers first in line. Substrates and capacitors run more than 3x their normal lead times. UBS sees DRAM capacity growing 147% to about 89 EB by 2030, still under a third of the 282 EB an agentic buildout would need. See memory hierarchy.
  • Datacenter financing strain (cluster of 4) (@rohanpaul_ai · Bloomberg, @StockSavvyShay, @StockSavvyShay, @StockSavvyShay). Big Tech holds over $1.3T in future leases, 70% added in the last four quarters. Oracle front-loads a $10B bond sale. CoreWeave bonds yield about 15% while Cipher's powered-shell debt prices at 6%. Owning power beats owning depreciating GPUs. See token arbitrage and the neocloud squeeze.
  • How to loop your transformer (@rohanpaul_ai · paper). Meta and MIT train dense models up to 1B on 155B tokens. Looping the whole stack gives the best quality at fixed size, and looping middle layers wins when training compute is tight. Per-loop adapters add 1.1 to 4.3% parameters versus 51.4% for doubling depth, and a repaired parallel-loop method runs loops at once during decoding. See looped transformers.
  • Context as infrastructure for agentic RAG (@dair_ai · paper). RAGFlow's harness with only BM25 and grep on MiniMax M3 hits 85% on BrowseComp-Plus and 91% on FRAMES. Trees, outlines, wikis and graphs added no accuracy. A claim-and-evidence navigation tree cut tokens 38% on average, but tripled latency until a decision gate cut it from 140s to 30s. Judge indexes on tokens and latency, not accuracy. Pairs with naive RAG harness cost.
  • The first few tokens decide CoT controllability (@omarsar0 · paper). Models struggle to control what their reasoning says because entropy collapses right after the opening thinking tag. Prefilling one token, or banning common openers, raises compliance by 34 to 41 points. Good for steering, and a warning for CoT monitoring, which partly relies on that inability. See CoT controllability shorts.
  • TextSeal2: watermarks that survive speculative decoding (@rohanpaul_ai · paper). Meta blends two secret keys into one virtual key so every token carries full signal, and draws replacement tokens from the drafts' leftover randomness so acceptance rates do not change. It detects 89% of 200-token answers at 0.1% false positives versus 71% for TextSeal. See watermark basins.
  • Claude watermarking backlash (@aakashgupta). Claims Anthropic will watermark Claude text worldwide from Tuesday, with scanners for schools and police but not for paying users, while OpenAI stamps only EU output. Opinion with no linked source. Treat as unconfirmed.
  • Cost per intelligence on Agent Arena (cluster of 2) (@arena · Pareto board, @deedydas). Same-quality agents can differ 10x in cost, 6x in output tokens and 2x in wait time, and output tokens barely predict the bill. OpenRouter's dollar volume shows Opus alone above a $5B annual run rate. A routing argument in one chart.
  • GDPx reaction (@omarsar0, @rohanpaul_ai · SemiAnalysis). Top score is 12% (Opus 5.5) on expert work across 11 domains. Finance and law score 0 under pass-fail grading. Commentary credits Gemini 4 Carbon on coordination and persistence. See GDPx.
  • DeepSeek Super-agent, round two (cluster of 2) (@iamsupersocks, @AlexFinn). More amplification of the Router-plus-Workers design. Nothing new beyond the summary.
  • Altman on efficiency and tokens (cluster of 3) (@rohanpaul_ai, @rohanpaul_ai, @rohanpaul_ai). Energy per query down 1,000x to 100,000x in 3 to 4 years, tokens are "a kind of fake unit", and AI is "mostly smarter" than top scientists. Unsourced numbers from a stage. Skim.
  • Fake "Numbat" leak (@rohanpaul_ai). A 14T-parameter Hugging Face repo claiming to be an Anthropic checkpoint. Parts of the weights are plain random-number output. Ignore.
  • Computers compiled into Qwen weights (cluster of 2) (@TheVixhal, @NFT_Chen · repo). qwenGPU puts a small SIMT GPU into Qwen3.5-2B, using Gated DeltaNet state as registers and memory, so it runs real PTX in its forward pass. A clever curiosity.
  • Roy resigns from JMLR (@roydanroy). Says AI reviews now beat rushed human ones and one-time accept or reject no longer fits. Proposes treating the literature "as a distributed data structure."
  • Agentic money and launches (cluster of 4) (@AnatoliKopadze, @StockSavvyShay, @ycombinator, @VaibhavSisinty). Mal raises a $230M seed for a bank-native agent. Hume ships a Decision API that reads tone from raw audio. Blacksmith (CI cloud) grew 20x on AI-written code. The UAE signs 4 deals with OpenAI.
  • Microsoft Humanist AI code of conduct (@mustafasuleyman · post). Public consultation on MAI model behavior, open for 3 more weeks. Click through to read.
  • AI consciousness thread (@dioscuri). Argues that "AI can't be conscious because then we'd owe it things" runs backwards: facts set obligations. Thoughtful, not technical.
  • Harness and hype bait (cluster of 8) (@Rel1vs337, @mdancho84, @beamnxw, @neeraj_here15, @0xMorlex, @L1vsun, @Nyxera0x, @mikenevermiss). Fake Karpathy quotes, "this is insane" agent lists and an uncensored Qwen fine-tune sold on benchmarks. Skip.
  • Promo: Explee ads, Sakana symposium recap, Kai-Fu Lee book tour, YC deadline reposts, Ronald van Loon hashtag stream, stock-chart posts and promoted ads. Skip.