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

2026-11-11

Summary

All of today's X signal came in the morning slot. The afternoon and evening slots were empty, and there was no night slot. The dominant cluster is Microsoft Research's Lazarus result, about 15 posts showing that rewriting fewer than 10 dormant neurons backdoors an agent with no training, and the model still passes every eval until a trigger fires. The agent cost bill is a close second and the most useful for efficiency work: Berkeley's Agency paper finds harness choices move cost up to 2.6x at flat accuracy, and ByteAsk cuts tokens per run by 78%. Compute economics carried real numbers, with CoreWeave's Q3 loss, older Ampere and Hopper contracts repricing 30 to 50% higher, and Anthropic holding an exit option on a $34.9B Lambda lease. Gemini 4 Carbon's general release and e-means (k-means 10x faster on GPU) round out the signal, while the Huang affair and the Moonshot chip allegations were mostly noise with no new artifacts. Full synthesis is in the 2026-11-11 daily digest.

Posts

  • Lazarus dormant-neuron backdoor (cluster of 15) (@DimitrisPapail · @AlexGDimakis · blog) [morning]. Editing under 10 rarely-firing neurons makes them inject payload tokens on a trigger, with no training and clean evals, across Nemotron, Qwen, Gemma and Phi. Best defense so far is a frontier agent inspecting the weights. Wiki.
  • Agent cost is a harness problem (cluster of 5) (@dair_ai · Agency paper · @rohanpaul_ai · parallel-teams paper · @omarsar0) [morning]. Agency shows prompts and tools change cost up to 2.6x at the same accuracy and cuts it up to 35%; ten teams of ten 27B agents beat 100 solo agents on 10 of 12 long tasks. ByteAsk claims 1,795K to 395K tokens per C++ fix (vendor numbers). Wiki.
  • Gemini 4 Carbon goes public (cluster of 5) (@GoogleAI · @github · @arena · @GoogleDeepMind) [morning]. Now on the paid API, Copilot and AI Ultra, with Arena saying its cost edge grows on the most expensive agent sessions. It ranks second on the new MKA benchmark at over 6x lower cost per task. Wiki.
  • CoreWeave Q3 and older GPUs repricing (cluster of 2) (@rohanpaul_ai · @StockSavvyShay) [morning]. Revenue up 168% to $3.66B but net loss $709M, as interest outran operating income. Ampere and Hopper pricing rose 30 to 50%, so ~300K GPUs rolling off contract could renew higher with no capex. Wiki.
  • Anthropic can cancel a $34.9B Lambda lease (@rohanpaul_ai) [morning]. Lambda missed a facility deadline, so either side can walk; Anthropic says it does not currently intend to exit.
  • Google TPU hosts move to Arm (@StockSavvyShay) [morning]. Three of four TPU systems now use Axion CPUs, and TPU 8i doubles CPU hosts to a 2:1 ratio as agentic inference grows the CPU share of the bill.
  • ExecuTorch WebGPU (@PyTorch · blog) [morning]. One torch.export produces a .pte that runs on the browser GPU, putting small-model on-device inference one export away for web apps.
  • e-means, k-means 10x faster (@BoWang87 · paper) [morning]. EMA mini-batch k-means with scaling laws that set decay and batch size, matching 25-epoch Lloyd's in 1.4 to 2.6 epochs. Reportedly in FAISS. Wiki.
  • Agents get their own identities (cluster of 3) (@omarsar0 · @AnatoliKopadze · @RedHat_AI) [morning]. A $5.3M raise for agent-owned computers, email and phone numbers, Inkbox messaging identities, and NVIDIA OpenShell joining CNCF to keep real keys out of agent reach.
  • Moonshot trained on B300s in Supermicro servers (cluster of 4) (@edzitron · newsletter) [morning]. Short-seller and critic allegations of chip diversion, not filings. Treat as unconfirmed.
  • OCC agent playbook for banks (@rohanpaul_ai) [morning]. Any agent that can move money on its own lands in the regulator's top risk tier.
  • PAN 2 open world model (@rohanpaul_ai) [morning]. MBZUAI's open action-conditioned world model for agents to simulate outcomes before acting.
  • 0.security launch (YC F26) (@ycombinator) [morning]. Claims 30+ Linux kernel 0-days and 500+ CVEs, with no technical write-up.
  • LittleBit sub-1-bit quantization, resurfaced (@CrazyShyyt) [morning]. Hype thread on Samsung's older low-rank sub-1-bit method; no new results.
  • Engagement bait and gossip [morning]. Karpathy-lecture bait, Claude Code skills listicles, Grok persona posts, Huang affair commentary and PhD-application threads. Skip.