01Claude helps claim a Riemann proof as a human rival publishes [AI Models & Research]Anthropic mathematician Levent Alpöge says he and Claude produced a machine-checkable proof of the 167-year-old Riemann hypothesis; Stanford postdoc Alessandro Masullo separately released a human-only proof of a broader version. Mathematicians must still verify both claims. → source
02Inkling 2 Mini handles images and text on local hardware [AI Models & Research]Thinking Machines’ new model is designed for personal computers and works with images as well as text. It reuses some layers three times to save memory; Hugging Face’s early tests are promising, but the team says more validation is needed. → source
03METR finds AI coding gains do not guarantee good judgment [AI Models & Research]The evaluation group added longer, open-ended work to its widely watched capability tests. None of the AI-designed evaluations was useful enough for METR’s own research, even as models kept improving on tasks with clear right answers. → source
04AI bug-hunting review finds more flaws, but no faster fixes [AI Models & Research]Researchers reviewing a decade of records say disclosed vulnerabilities more than doubled in 2026, including more long-lived severe bugs. The time from report to repair, and from disclosure to attack, has not shifted noticeably. → source
05Hugging Face cuts coding-agent token use with custom training [AI Tools & Ecosystem]Small add-on adapters raised one coding test’s pass rate from 39% to 55%, while average token use fell from 2,900 to 1,500 per task. Hugging Face released the training recipes, but gains did not clearly transfer to unrelated business-automation work. → source
06Databricks index shows identical coding scores can hide huge costs [AI Tools & Ecosystem]GPT-6 Luna and Claude Opus 5.5 each completed 89.1% of Databricks’ coding tasks, yet cost $0.05 and $1.62 per task respectively. The weekly public index also tracks usage and spending at more than 10,000 companies. → source
07Microsoft and 40 organizations back verifiable AI-agent actions [AI Tools & Ecosystem]The AI Transparency Initiative aims to create open standards for recording what agents access, verifying their actions and limiting delegated authority. Microsoft plans to apply the resulting controls across Copilot, Foundry and its own models. → source
08Copilotforce brings Salesforce customer records into Microsoft Copilot [AI Tools & Ecosystem]Sales teams can prepare with customer data alongside Teams messages, Outlook email and shared documents. The Copilot integration is in public preview; changes to Salesforce records require confirmation and follow existing permissions. → source
09Two million bait secrets expose leaks across AI services [AI Applications & Industry]Dragos found evidence of breaches at three providers, resellers reading customer prompts, and ignored privacy settings during a five-month experiment. The company says every intermediary in an AI request can see the full prompt and response. → source
10Lenders resist financing SpaceX chips that cannot be repossessed [AI Applications & Industry]Lenders behind SpaceX’s $40 billion chip financing reportedly sought to bar their collateral from orbital data centers. That restriction could raise costs for space-based AI computing, where failed loans cannot end with a warehouse repossession. → source
11Spotify’s Daniel Ek backs a simulation-first fusion startup [AI Applications & Industry]Europa Fusion founder Nantas Nardelli previously worked at DeepMind and led machine learning at Pacific Fusion. The UK startup will favor simulations and data over a hardware-led approach to pulsed-power fusion; no fusion company has yet supplied electricity to the grid. → source
12Google commits over $500 million to super-hot geothermal power [AI Applications & Industry]Google will pay more than $500 million for 50 megawatts from a Quaise project in Oregon, tapping heat three miles underground. The deal backs a potential power source for data centers, but the drilling technology remains unproven at commercial scale. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.