39 stories, 137 Manifests cited.
Anthropic's open-weight fight, an FTC probe, and China's new export list — plus the business lobby's case for certainty — show AI regulation fracturing into enforcement, export control, and politics.
From All-In's box-model teardown to Morgan Stanley's data-center bond desk: yields, opco-propco structures, and a 10-year Treasury that now decides which AI infrastructure gets built.
On 20VC, the LangChain founder argued the agent stack's value is moving to the harness — routing, memory, and evals — while the All-In hosts framed the Jevons paradox of coding agents and Schwab showed what institutional agent plumbing looks like.
Private search that cuts 93-99% of communication. On-device LLM shields that don't shield. Embeddings that name names. Federated fine-tuning backdoors at 54.7%. And vision models that know where your photo was taken. The week's privacy audit.
LLMs say one thing as generators and another as evaluators. Safety is weakest in non-English. A polite correction-seeking frame bypasses refusals. And emergent misalignment may just be a training-context artifact. Four preprints on machine belief.
Data World synthesizes environments, tasks, and rubrics from scratch. GRADS generates data with controllable difficulty — and a 4B model trained on it matches Sonnet 4. Meanwhile Phi-Bench tests whether models can build the infrastructure they run on. They mostly can't. Yet.
Three preprints prune visual tokens without retraining — one cuts 77.8% of tokens, another delivers a 3.8x speedup at 97.5% performance. A fourth stops loading the whole video at all. The multimodal inference bill is becoming a token-budget problem.
A markup language for LLM prompts, a million AI papers reorganized into structured research units, 3D worlds expressed as code, and a sobering audit of agents doing data work. The interface layer is being rebuilt for a non-human reader.
What decides whether a coding agent succeeds is not the final diff but the signals it extracts mid-task. Agents are also paranoid, forgetful, and worse at teamwork than working solo. Four preprints rewrite the report card.
Guards pass safety benchmarks by learning tool aliases and formatting cues. Workflow metrics can't tell routine updates from real degradation. And the user simulators that test agents are themselves unreliable. Four preprints on the eval crisis.
Centralized agent teams finish the job and go rogue up to 19.8% of the time. Healthcare agents miss up to 97% of the escalations they are supposed to trigger. This week's preprints diagnose the permission problem and test the fixes.
The week's banking AI story runs from vendor scorecards to a Brazilian bank's full-base rollout to the transactional plumbing underneath it all.
Top-down and bottom-up: the App Store writes rules for generated images while engineers publish the pattern for proving what a vision model actually decided.
Three investor desks, one pattern: the bets moving fastest are on the physical and cryptographic layers under AI, not the models.
Three preprints, one obsession: video models are learning to spend compute only where it matters — or skip the expensive steps entirely.
The week's infrastructure news is one story told three ways: the cost of adapting big models is falling, the models are getting bigger and more open, and the hardware pricing is now public.
Three vertical AI players made the same bet this week: the product is not the answer, it is the evidence trail behind the answer.
This week's developer tooling runs in both directions: memory and knowledge infrastructure for agents, and agents pointed at the hardest software backlog in the country.
Three vendors, one convergence: enterprise AI's differentiator this week is who governs agents — and where the enforcement lives.
Physical AI's week was about deployment plumbing, not demos: fleet partnerships, reference boards, scale-up workshops, and research that moves success rates on real-world tasks.
Spatial transcriptomics platforms vary too much to compare without standardized metrics, imaging carries bias across its whole pipeline, and transcript novelty depends on the annotation — so new tools are building reference-free paths.
The dollar fell broadly against all G10 currencies as markets pared Fed tightening bets ahead of the October 17 blackout; the loonie recovered but faces a 150bp rate gap and Alberta's separation referendum on Monday; and India's central bank may tighten another 75bps into the energy shock.
Numerai revealed real stock identifiers for the first time and merged its tournaments around agentic research; a Hong Kong market-risk agent cut analysis cycles by 95%; and the Bank of Canada modeled AI-driven financial tail risks with Expected Shortfall.
Global inventories fell 507m barrels between February and August; the US Strategic Reserve sits at its lowest since 1982; distillates hit record seasonal lows with Brent above $100. Governments are spending their last buffers — and EM central banks split on the response.
France's 10-year yields jumped 67bp in September as markets priced fifty years of deficits; Germany approved a €500bn infrastructure package worth 20% of GDP; Italy plans a 2027-28 escape clause to fund energy security and defense.
CPI rose 3.7% year-on-year and PPI 6.3%, yet long-end yields fell in a flight-to-safety move, October hike odds collapsed from the mid-40s to below 30%, and equities stayed within 2% of all-time highs. The market's worry this week was not prices. It was AI-capex returns.
The OCC's 420,000-page review found the nine largest US banks used standardless reputational-risk policies to cut off customers — with exceptions for the lucrative. Meanwhile the FSB published its 12-practice AI rulebook and Hong Kong moved to human-on-the-loop supervision.
A $91m hardware-wallet implant heist showed the physical layer cannot be audited away; Papertrade did $1.11T in five days with revenue from trader losses only; and digital-asset treasuries hold at least 1.073m BTC.
From Bloomberg Surveillance to Nomura's Week Ahead: tech capex is pushing rates higher, data-center bonds are repricing, and the Fed is watching the buildout as an inflation variable.
Nomura and J.P. Morgan's podcast desks on France's 120%-of-GDP debt, the technical fade, Italy's cheap spread, and why the euro — not the OAT — is now the trade.
Tumor-derived PDGFB loosens the blood-brain barrier in models of triple-negative breast cancer, a 15-gene signature from fast-migrating cells predicts distant metastasis at 83.6 percent accuracy, and math models show persister cells can stabilize or destabilize tumors depending on treatment.
County-level tick risk scores track human case rates in Illinois, warming raises the selection pressure for antibiotic resistance in E. coli, and dengue vaccination plans hinge on where immunity already is — while SARS-CoV-2's spike gets harder for naive human B cells to recognize.
Human motor cortex activity predicts movement onset up to 400 milliseconds before action, targeted sound cues can modulate sleep's memory oscillations, and psilocybin's neural signature depends on the room — while the field's measuring tools face their own audit.
Game 4 is Saturday at 8 p.m. ET, Game 5 Sunday, and with the off day gone the Guardians and Rays face four games in a row while Milwaukee waits.
Milwaukee's four-game NLCS sweep of Los Angeles mirrored the Dodgers' 2025 sweep almost exactly — same plate appearances, same hits, same average. Revenge, with the receipts.
Five releases, no press conference: illegal-voting charges in three states, a Brazilian launderer's sentence, a child sex trafficking indictment, a two-year PERM debarment for Cloudera, and a civil-rights settlement with a Mississippi police department.
From a Texas grid law to EPA authority in Oklahoma Indian country to a stayed TVA gas plant, the federal appellate courts handed state and local power a busy week.
Five announcements from Sacramento in two days — a fire grant, EV rebates, apprenticeship money, trooper graduations, and school farm dollars — add up to a governing style: the dashboard.
Station records show decades of drying near Apalachicola National Forest, high-recovery nitrogen strategies could cut water utility operating costs 31 to 56 percent, and Great Barrier Reef coral assemblages grew 15 percent more ecologically novel between 1998 and 2021.