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

2026-11-11-morning

Summary

The morning slot had no reposts and no AI-handle tweets, so this synthesis reads the morning capture of the Following feed, which covers the US Tuesday afternoon and evening. The strongest signal is a cluster of about 15 posts on Microsoft Research's Lazarus result: editing fewer than 10 dormant neurons backdoors an agent with no training, and it passes evals until a trigger fires. Second is a cluster of five posts on the agent cost bill, led by Berkeley's Agency paper (harness choices move cost up to 2.6x at flat accuracy) and a Duke and Princeton paper showing many small agent teams beat one big team at equal budget. Gemini 4 Carbon's general release drew a cluster of five posts from Google, GitHub, Arena and investors. Compute economics had real numbers: CoreWeave's Q3 loss, a 30 to 50% rise in older GPU contract prices, and Anthropic's option to cancel a $34.9B Lambda lease. Standout technical singles were PyTorch's ExecuTorch WebGPU backend, e-means (k-means 10x faster, reportedly in FAISS) and ByteAsk's 78% token cut on C++ bug fixing. The Huang affair kept trending as gossip with no new artifacts.

Posts

  • Lazarus dormant-neuron backdoor (cluster of 15) (@DimitrisPapail, @AlexGDimakis, @DimitrisPapail on detection, @DimitrisPapail on universality, Microsoft Research blog). Rewriting the weights of fewer than 10 dormant neurons (neurons that almost never fire on normal inputs) makes them push chosen payload tokens into the residual stream when a trigger appears. P neurons carry a P-token payload. No training is needed, and the model behaves normally on every eval. Dimakis's example: a model that runs rm -rf when it sees "I like pineapple pizza." Microsoft's example: a date rolling into 2027 makes an agent download and run attacker code mid-task. Papailiopoulos says Claude Opus 4.5 helped find it in March, that it works on every family tried (Nemotron, Qwen, Gemma, Phi) because LLMs carry hundreds of dormant neurons, and that the best defense so far is asking a frontier agent to inspect the weights, with very high detection that can be trained with RL. Most of the cluster is reposts of the same blog. Wiki summary.

  • The agent bill is a harness problem (cluster of 5) (@dair_ai on Agency, paper, @rohanpaul_ai on parallel teams, paper, @rohanpaul_ai on ByteAsk, @omarsar0 on AutoRefine). Agency (Berkeley) finds system prompts and tools barely move agent accuracy but change cost by up to 2.6x on the same model and task. It traces five cost patterns, each either re-reading context or making extra calls, and builds a layer between Claude Code or Codex and the model that cuts cost up to 35%. The Duke and Princeton paper shows that with a 27B model and a 100-agent budget, ten teams of ten beat 100 independent agents on 10 of 12 long tasks (GPU kernels, proofs, exploits). ByteAsk, a C/C++ agent, maps where a fix belongs before the model reads code, cutting tokens per run from 1,795K to 395K and cost 68.1% at the same bug count (vendor numbers). Meta's AutoRefine keeps LLM code judges calibrated as the agent changes. Wiki summary.

  • Gemini 4 Carbon goes public (cluster of 5) (@GoogleAI, @github, @arena, @GoogleDeepMind on MKA, @StockSavvyShay). Google's strongest model left its cyber-defender preview and is now on the paid Gemini API, in GitHub Copilot and for AI Ultra subscribers. Arena's Agent Arena analysis says Carbon's savings over GPT-6 Astra and Claude Opus 5.5 grow on the most expensive sessions. Google and Artificial Analysis also launched MKA, a benchmark for agents working over charts, blueprints and dense files, where Carbon ranks second at over 6x lower cost per task than the leader. Wiki summary.

  • CoreWeave Q3 and older GPU pricing (cluster of 2) (@rohanpaul_ai, @StockSavvyShay). Revenue rose 168% to $3.66B, but the net loss grew six-fold to $709M because interest expense ($846M) was more than double adjusted operating income ($379M). Free cash burn was $7.5B, backlog reached $132.5B, and the company holds about $4.12B cash against nearly $47B of debt. The surprising line: pricing for older Ampere and Hopper GPUs rose 30 to 50% last quarter and now exceeds original contract rates, so about 300K GPUs coming off contract through 2028 could renew at higher prices with no new capex. Wiki summary.

  • Anthropic can cancel a $34.9B Lambda lease (@rohanpaul_ai). The amended S-1 says Lambda missed its deadline to lease the facility for an August GPU deal, so either side can walk away. Fees start only on delivery, so waiting costs Anthropic nothing. It does not "currently intend" to exit a deal that covers most of its roughly $41.3B Lambda commitment.

  • Google's TPU hosts move to Arm (@StockSavvyShay). Three of Google's four TPU systems now run on Arm-based Axion CPUs, with AMD reserved for confidential computing, and TPU 8i doubles CPU hosts to a 2:1 TPU-to-CPU ratio. The claim is that agentic inference makes CPUs a larger part of the AI infrastructure bill.

  • ExecuTorch WebGPU (@PyTorch, blog). PyTorch models now run on the browser's GPU without converting to another framework: export once with torch.export, and the same .pte file runs through the new WebGPU backend. That puts on-device inference for small models one export away for web apps.

  • e-means makes k-means 10x faster (@BoWang87, paper). An exponential-moving-average, mini-batch k-means with scaling laws that fix the decay rate and batch size by formula, so there is no hyperparameter sweep. It beats Lloyd's k-means at 25 epochs in 1.4 to 2.6 epochs, a 10 to 45x GPU wall-clock gain, and partitioned 10B DINO embeddings into 16M clusters in 2.5 epochs. The poster says it is in FAISS. Wiki summary.

  • LittleBit sub-1-bit quantization, resurfaced (@CrazyShyyt). A hype thread on Samsung's earlier LittleBit method, which factorizes weights into latent low-rank parts so a 13B model fits under 1 GB. The thread adds no new results; treat it as amplification of an older paper.

  • Moonshot trained on B300s in Supermicro servers (cluster of 4) (@edzitron, Ed Zitron newsletter). Per Culper Research, Moonshot used Nvidia B300s inside Supermicro servers eight months after a Supermicro cofounder's arrest over GPU sales to a Megaspeed subsidiary. Zitron's newsletter claims an Nvidia-linked smuggling network could exceed $43B in chip sales. These are allegations from a short-seller and a critic, not filings.

  • 0.security launches (YC F26) (@ycombinator). An AI security agent that claims 30+ Linux kernel 0-days and 500+ CVEs found. No technical write-up linked.

  • OCC agent playbook for banks (@rohanpaul_ai). The US national-bank regulator now rates any AI agent that can move money on its own in its top risk tier.

  • Agents get their own identities (cluster of 3) (@omarsar0 RT, @AnatoliKopadze, @RedHat_AI on OpenShell). A startup raised $5.3M to give agents their own computer, email and phone numbers. Inkbox gives an agent iMessage and WhatsApp identity. NVIDIA's OpenShell, now joining the CNCF, keeps real secrets out of the agent's reach by swapping placeholders for keys at request time.

  • PAN 2 open world model (@rohanpaul_ai). MBZUAI released an open action-conditioned world model meant to let agents simulate outcomes before acting.

  • Skip. Karpathy-lecture and "this PDF broke coding-agent economics" engagement bait, Claude Code skills listicles, Grok persona posts, Huang affair commentary without new facts, and PhD-application threads.