01OpenAI Claims Navier–Stokes — After an NYU Mathematician Cried Foul [AI Models & Research]NYU's Tristan Buckmaster announced partial proofs toward the Millennium Prize problem, built with Codex and Claude — then learned his progress had leaked to OpenAI, which days later claimed a full proof via an unreleased model at a cost of $22.5M in compute. A career-making result now arrives tangled in an ethics fight over how labs race to machine-made math. → source
02AlphaGenome Atlas Precomputes the Molecular Impact of All 9 Billion Single-Letter DNA Variants [AI Models & Research]DeepMind turned its AlphaGenome model loose on the entire space of possible single-nucleotide mutations and published the results as a free, searchable atlas with a new AVI ranking score. Biology gets its first exhaustive 'map of every possible typo' — no wet lab required. → source
03Safety for Whom? Refusing the Right Subset, Not the Whole Topic [AI Models & Research]Most guard models treat harm as a property of a whole topic — weapons, fraud, self-harm — and over-refuse safe prompts as collateral. Multiverse Computing's boundary-aware self-distillation trains models to refuse the dangerous slice of a topic while keeping the legitimate rest of it usable. → source
04Chain of Thought vs. Tree of Thoughts: A Practical Decision Guide for Agent Builders [AI Models & Research]A clear-eyed comparison of the two core reasoning frameworks: linear CoT is cheap and fits most agent steps, while branching ToT earns its cost only where tasks need evaluation and backtracking. The practical answer for most agents is a router, not a religion. → source
05OpenAI's New Image Models Get Reference-Photo Editing — and a Stronger CLI Story [AI Tools & Ecosystem]ChatGPT Images 2.5 lands as two API models: Sunburst for precision editing, Flare for fast generation, both better at multi-turn instruction-following and preserving subjects from reference photos. Simon Willison wired reference-image input into his CLI tool the same day — small models of how fast the image stack is commoditizing. → source
06Chrome Cuts Its Patch Gap to Two Weeks as AI Compresses Vulnerability Timelines [AI Tools & Ecosystem]Chrome 153 makes Google's browser a biweekly release, explicitly to shrink the N-day window as AI-accelerated threat discovery outpaces a monthly cadence. Mozilla, Microsoft and Brave are following — release rhythm is now a security decision, not a feature decision. → source
07Patch Tuesday Goes Vertical: Microsoft Ships a Record ~972 Fixes, 112 Critical [AI Tools & Ecosystem]Microsoft's September patch release fixes roughly 972 vulnerabilities — 112 critical, two zero-days — after a then-record 570 just two months ago. Zero Day Initiative calls the spike the 'new normal' as AI-assisted discovery floods the pipeline faster than fixes can ship; Mozilla's Mythos-assisted hunters found 271 bugs in May with almost no false positives. → source
08Mistral Raises €3B at €21B as Samsung Backs a 'Third Way' in AI [AI Applications & Industry]The largest equity round in European tech history, led by Samsung with France's blessing, funds a push to 1 GW of European compute by 2030 — and doubles down on a services-and-sovereignty strategy instead of chasing a consumer ChatGPT. Macron calls it a 'third way in AI'; the round makes it a funded one. → source
09Meta Launches Muse, an Agent Wired Into Your Email, Payments and Smart Home [AI Applications & Industry]Meta's biggest consumer-AI bet yet connects to your calendar, bills, shopping and checkout via Stripe's Link — free to start, card required, with browser fallbacks when no API exists. Launching weeks after an $18B privacy settlement, it asks users to trust Meta with far more than their feeds. → source
10Cognition Closes $2B at $48B — Above the $40B Talks, Revenue Doubled Since May [AI Applications & Industry]Four months after its $26B round, Devin-maker Cognition doubled its valuation on run-rate revenue that grew from $492M to $900M since May — with an $800M burn and its own Nvidia cluster lease. After Cursor sold to SpaceX, investors are betting the coding-agent market stays plural. → source
11Claude Token Thieves Move In: Hackers Now Sell Access to Your AI Subscription [AI Applications & Industry]Attackers are harvesting Claude OAuth tokens from users — via fake apps and malicious browser extensions — and reselling working access to premium AI subscriptions, continuing the shift from stealing your cookies to stealing your agents. Sessions are the new account credentials. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.