AI research wiki
social-stream · 2026-11-01

2026-11-01

Summary

A thin Sunday with no curated reposts in either slot; only the morning and evening captures ran, both drawn from the Following feed. The strongest cross-slot thread is compute economics: the morning had rising B300 and H100 rents, CoreWeave's debt-funded GPU leverage and Baseten's roughly $2B raise, and the evening had 2027 racks at 200kW to 1MW, so power and cooling become the gate, not chips. Anthropic's compute exposure also showed up in both slots through two Chamath posts. The best single item is the evening's NVIDIA explainer on why KV cache eviction on a paged server can drop 90% of tokens yet free almost no memory. The morning's routing and gating cluster (Meta FAIR's Routed Continual Learning, IBM's JevAct) and the evening's paper run (SDPO, CMU's single-algorithm training, the JEPA Platonic proof, FlowAgent) carry the research signal. The "paste this into Opus 5.5" harness bait was loud and empty. Full analysis is in the 2026-11-01 digest.

Posts

  • Compute economics and rack power (cluster of 8) (@StockSavvyShay, @StockSavvyShay, @StockSavvyShay, @rohanpaul_ai, @StockSavvyShay, @StockSavvyShay, @StockSavvyShay, @StockSavvyShay · Compute economics) [morning + evening]. B300 rents are up 52% since May while token prices halved, Baseten raises about $2B at up to $30B, and Meta's capex path reaches 26GW by 2030. 2027 racks hit 200kW to 1MW and Colossus 2 ends the year at about 1.2M Blackwells; treat the stock calls as opinion.
  • Anthropic's ceiling is compute, not demand (cluster of 2) (@rohanpaul_ai, @rohanpaul_ai) [morning + evening]. Chamath counts about 1,900 customers spending over $1M a year. In his reading of Anthropic's risk factors, data centers, power and chips dominate and demand barely appears.
  • KV cache compression frees tokens, not memory (@_avichawla · NVIDIA blog · KV cache) [evening]. PagedAttention reclaims a block only when all 16 slots are empty, so 90% logical eviction can free 0% physically. TriAttention compacts every 128 tokens for 10.7x less KV memory.
  • Why your logits depend on other people's requests (@_avichawla · summary) [morning]. Batch size changes the order of floating-point sums, which can flip near-tied greedy tokens. Use batch-invariant modes in vLLM or SGLang for evals and RL rollouts.
  • Routing and gating inside agents (cluster of 3) (@dair_ai, @omarsar0, @xsmryze · 2610.38796, 2610.38620) [morning]. Routed CL sends each task to a cluster of past runs, gaining 15-30% of max reward at 65-72% lower API cost. JevAct's prefill-only gate lifted Opus 4.6 on τ²-bench Telecom from 39.5% to 80.7%. Routed CL · JevAct
  • Faster MoE kernels from column-major scheduling (@anshulcreates · PyTorch blog · GPU kernels) [evening]. Reordering fused MoE GEMM tile scheduling for cache reuse gives up to 4x on A100 and 4.4x on H100. An older post, resurfaced.
  • Before GPU Engineering (@techNmak · repo) [evening]. A free 56-chapter notebook on warps, coalescing, occupancy and profiling before CUDA. Later chapters are unfinished.
  • Hyperscaler cloud at $440B run-rate (@StockSavvyShay) [evening]. Google Cloud grew about 103% and Bedrock token volume about 17x. Agents now open over 40% of GitHub pull requests.
  • DeepSeek Super-agent (cluster of 3) (@MinLiBuilds, @VaibhavSisinty, @Hesamation · Harness engineering) [evening]. A Router dispatches parallel Workers because human attention, not the model, is the bottleneck. p95 completion fell from 29.5 to 15.7 minutes.
  • SDPO: learn from your own failures (@Pyuyi2333 · paper) [evening]. Self-distillation from feedback on failed attempts, no teacher. About 6x faster than GRPO in wall-clock on chemistry.
  • One algorithm for every training stage (@omarsar0 · paper · summary) [evening]. CMU casts RL as likelihood fitting on model-made targets, so all stages share one loss. A 5% midtraining mix best preserved coding after math RL.
  • JEPAs prove the Platonic Representation Hypothesis (@askalphaxiv · paper · summary) [evening]. Image and text JEPAs converge to one representation up to rotation. A 109M text JEPA beats LLMs up to 72B on cross-modal alignment.
  • FlowAgent: CI test repair at Google (@rohanpaul_ai · paper) [evening]. 67% of fixes judged correct and 28,554 applied in production. Latency, not accuracy, decides adoption.
  • Pretraining poison survives clean post-training (@SuJinyan6 · summary) [evening]. 0.1% poisoned text plants an RCE backdoor in CLI agents at 92.7% success after SFT, DPO and RL. Defend at the action layer.
  • Aligning Autoresearch (@Xudong07452910 · paper) [evening]. Human input matters most at ideation. On HackBench, the best frontier model scores below 53 F1 at spotting a hidden reward hack.
  • AI novelty checks reject real ideas (@rohanpaul_ai · paper) [evening]. A deployed filter rejected 96% of later-published ideas. Claim-level ClaimNov cuts that to 1.5%.
  • Unprovoked pushback breaks correct code (@ai_database) [evening]. "Are you sure?" makes Gemini rewrite correct code up to 80% of the time. Logging its reasoning first fixes most of it.
  • RL on simulated psychiatric patients (@rohanpaul_ai) [morning]. DeepMind's RL-trained Flash-Lite was preferred by psychiatrists in 76.7% of pairs. Missed risk escalations fell to 12.9%, still about 1 in 8.
  • Personal agents that know when to defer (@rohanpaul_ai) [morning]. Meta's PAMO uses cross-app history only when ranking holds. A 4B agent hit 47.9% top-10 on Steam vs 40.8% for the platform.
  • STAIR: table-of-contents retrieval (@simplifyinAI) [evening]. IBM retrieves by document structure, 82.6% vs 76.9% first-hit. The "RAG is cooked" framing is hype.
  • Agent evals and CoT faithfulness (cluster of 3) (@omarsar0, @hammer_mt, @rao2z) [morning]. Self-improving agents need evals first, possibly market-outcome evals. Kambhampati notes CoT serves interpretability and faithfulness, which need different things.
  • Report from the Vatican (@GaryMarcus · post · summary) [morning]. Marcus backs Pope Leo XIV's call to keep lethal decisions human. He blames open-ended agents for the current incident wave.
  • OpenAI meters ChatGPT by tokens (@rohanpaul_ai) [evening]. From December, ChatGPT and Codex share one token allowance instead of per-model caps.
  • DeepMind PlanAssist (@GoogleDeepMind) [morning]. Cuts UK householder planning review time by up to 58%, expanding to 12 more councils.
  • Terence Tao on "strip mining" math (@rohanpaul_ai) [evening]. Tao says AI's Navier-Stokes result left the field "in disarray".
  • Distillation accusation (@stanfordnlp) [evening]. A researcher denies distilling Claude. One side of a dispute.
  • Meta's humanoid Max and QwenPU (@StockSavvyShay, @divagr1925) [morning + evening]. Meta's Jetson Thor humanoid was reportedly built with Fourier in three months. QwenPU runs a tiny LLM on an NPU compiled into Qwen weights, a curiosity.
  • Harness bait (cluster of 9) (@hanakoxbt, @0xrevayz, @cyrilXBT, @LunarResearcher, @AnnatarXBT, @cyrilXBT, @AnatoliKopadze, @AndyL5cc, @mdancho84) [evening]. "Paste this into Opus 5.5" prompts and "end of harness engineering" hype. Skip.
  • Unverified Claude watermarking claim (@_junaidkhalid1) [evening]. No source given. Skip until confirmed.
  • Promo and off-topic [morning + evening]. "End of software engineering" hype, a mislabeled Karpathy lecture clip, NVIDIA's SETI promo, the Explee ad, the YC W27 deadline, and election and space posts. Skip.