27 stories, 108 Manifests cited.
Four new preprints take RAG's machinery apart: adding documents degrades retrieval through vector search dilution, frozen external memory needs a target-aligned reader, a dual-process memory cuts output tokens 73.91%, and consensus-driven evidence selection trims runtime 75.9%.
Four new preprints put LLMs through finance's gauntlet: 15-23% of confident answers are wrong on FinQA, a 1,575-instance benchmark grades models by backtest rather than syntax, and proprietary models still lead open-weight ones 92.8% to 84.1% on accounting-relation extraction.
An OpenAI engineer was arrested for allegedly bugging an executive conference room — amateurishly, podcasters say — while SentinelOne and Block show what AI does for defenders.
Four new preprints attack LLM inference memory from four sides: crush the KV cache to INT2, protect the attention sinks, allocate bits by saliency down to 1.5 bits, or evict smarter. The numbers are striking; all are unreplicated preprint claims.
Intel wants your LLM workload back on CPUs with self-tuning containers; fal.ai rents B300s at $8.50 an hour; and researchers put a number on the energy cost of long context.
Agility's Digit 5 charges in nine minutes for ninety minutes of work; a new benchmark finds robots that can pick up tools but not finish the job; and Japan bets on robots for a labor crisis.
Four new preprints give web agents stranger tests: agents are tested against UI and workflow changes, Web$hell unifies search and browsing in one interface, SoftwareGym synthesizes 182 apps for training, and a lazy grounding attack shows factual-but-irrelevant evidence steers agents wrong.
Two new benchmarks measure how coding agents violate task boundaries and write insecure code; Chainguard's server-side MCP servers and a live Node.js vulnerability dataset arrive as the industry's answer.
Three preprints puncture three beloved ideas — that distilling agents into themselves helps, that verifiable rewards are honest, and that RL taught models to reason.
Agent skills — reusable bundles of procedures and domain knowledge — get the full infrastructure treatment in three new preprints: retire stale skills for a 7.8% benchmark gain, watermark skills to prove ownership, and watch attackers smuggle malicious logic through skill composition at 98% success.
The Nordics plan 1 gigawatt of AI cloud by 2027, Britain takes a $250M safety bet, European cities build local AI factories, and Google tallies its EU footprint.
Parloa signed partnerships with AloTech and TELUS Digital to push agentic AI through contact-center channels. Meanwhile CallMiner's research says retail, the fastest adopter of automation, reports the weakest automated customer experiences of any sector.
The FSB finalized its AI playbook for banks the same week IBM found eight in ten EMEA executives had an AI incident reach the board — and Anthropic hardened its election defenses.
Three new preprints give AI agents their own safety benchmarks: multimodal agents fall to visual hijacking at an 87.96% attack success rate, half of successful GUI runs secretly wander off track, and static cyber benchmarks are saturating.
G2's 2026 report says agent adoption jumped 25% but a distrust tax eats the gains; DevRev maps the pricing confusion; FirstMark bets the ledger is the real prize.
Three ChemRxiv preprints this week push carbon capture in three directions: a room-temperature methanation trick, a MOF that regenerates on sunlight, and a redox carrier that survives oxygen.
Three ChemRxiv preprints: a reusable Raman chip that reads ethanol in gasoline, a freezing-droplet test for a breast cancer biomarker, and a neural net that predicts polymer membrane permeability.
The Bank of Finland says AI is now Europe's policy problem. The Bank of Greece is putting language models inside its macro models. And a BIS speech argues household financial health belongs on the central-bank dashboard.
Morgan Stanley sees tokenized assets growing from about $40 billion to $2.3 trillion by 2030. Coatue says the world is getting more asset-heavy. Citi and StepStone show the plumbing being rebuilt underneath it all.
69% of finance leaders in Mexico and Brazil are using or planning to use stablecoins for payments. In the US, U.S. Bancorp piloted its own stablecoin and DriveWealth wired Ripple's rails into its brokerage platform.
Spain hit 4.9%, Germany 3.3%, and US CPI printed 3.7% above forecast. Then came the credibility question.
Jensen Huang told BG2Pod AI compute is headed for a billion-fold increase. OpenAI's Sarah Friar told Bloomberg she expects at least $70B in annualized revenue. BofA's Moynihan and Cisco's Robbins weighed in on rates, risk, and growth.
Brigimadlin for solid tumors, ketamine for PTSD, a liver-cancer triple combo, bupivacaine for surgical pain, glucagon in metabolic disease, and revefenacin in COPD all reached completed-with-results.
A bioRxiv preprint maps a brainstem circuit that wakes brains from metabolic coma. A medRxiv preprint finds not all blood-pressure drugs are equal on dementia risk. Another validates a spinal-fluid marker for Parkinson's.
Space Force activated SPACEFOR-SOC for special operations missions, effective October 7. Secretary Hegseth said Project Agincourt is the pathway to a four-star-led Autonomous Warfare Command.
Michigan wiped out $21 million in medical debt for 9,767 residents. Arizona says its AZRx card saved residents over $1 million on prescriptions. Kentucky wants Washington to fix SNAP.
An EarthArXiv preprint says fast erosion flips rock carbon from sink to source. Another flags M6.1-6.8 earthquake risk in the Atacama. And EIA reports New England wind generation jumped 68%.