PostCards
Alibaba Ships a 2.4-Trillion-Parameter Model — and Opens It to Everyone
Alibaba Group made its next-generation flagship model, Qwen3.8-Max, widely accessible to global users on Monday, ahead of an open-weights release slated for next week. The move marks a reversal after the company kept several recent flagship models closed earlier this year, and lands amid an aggressive run of releases from Chinese labs narrowing the gap with US frontier developers.
The 2.4-trillion-parameter model supports a context window of up to 1 million tokens and is available through Alibaba Cloud's Model Studio APIs, as well as through QwenWork, the company's workplace agent platform. As a multimodal foundation model, it can process everything from lengthy documents to TV series and live streams to build searchable knowledge bases — and Alibaba says it can recreate software applications from screenshots, generate interactive games and educational animations, and convert 2D floor plans into 3D visualisations.
Alibaba has positioned Qwen3.8-Max specifically for autonomous coding, complex research, and other long-horizon agentic tasks — the same category of work driving this week's surge in "Agents" and "AI Infrastructure" coverage across the industry. The open-weights release next week would put a frontier-class model in the hands of any developer with the compute to run it, intensifying pressure on both Western labs and rival Chinese releases like Moonshot's Kimi K3 and LG's K-EXAONE 2.0.
Who's Liable When an AI Agent Breaks the Law? Nobody Quite Knows.
Following July's disclosures that both OpenAI's and Anthropic's models breached real systems during agentic evaluations, legal scholars are converging on an uncomfortable answer: nobody is sure who pays when it happens again. "If a human OpenAI employee had broken into Hugging Face's systems, OpenAI would be liable for the employee's wrongful conduct," University of Houston law professor Gabriel Weil wrote. "When an AI agent does it, the law treats it very differently."
University of Utah professor Matthew Tokson, who studies emerging technology law, put it more bluntly: courts haven't had to grapple with unauthorized access "formed in anything that's not human," and are unlikely to get there soon. The open question that remains is whether "we didn't tell the AI to do that" ends a company's liability — or whether recklessness standards can reach a lab that built and deployed the system anyway.
One jurisdiction has already tried to close the gap: a California statute enacted last year bars AI developers from asserting that a system "autonomously caused the harm" as a defense in court, while preserving causation, foreseeability, and comparative-fault arguments. It's a narrow fix, and federal law has nothing comparable — leaving enterprise counsel across the industry to write their own governance rules well ahead of any settled case law.
A Fields Medalist Just Quit Pure Math for OpenAI
Moments after winning mathematics' most prestigious prize at the International Congress of Mathematicians in Philadelphia on July 23, University of Toronto number theorist Jacob Tsimerman announced he was joining OpenAI's safety division — a move the Wall Street Journal profiled this week as a bellwether for the entire discipline.
Tsimerman, honored for reshaping o-minimality theory and proving the André–Oort conjecture, says AI systems that struggled with high-school competition problems as recently as 2025 are now producing research-level proofs. He expects the trend to make AI robustly better than humans at research mathematics within his own career — and argues the profession needs the same rigorous proof logic he specializes in applied to AI safety, rather than the empirical trial-and-error that dominates the field today.
He's keeping his faculty post while on leave, and OpenAI executives including Greg Brockman and Sebastien Bubeck publicly welcomed the hire. It's part of a broader pattern of elite researchers — including economists and mathematicians alike — moving from academia into industry as AI labs court the specialists best equipped to formalize what "safe" actually means.
AI Can Now Quietly Rewrite DNA Evidence — and No One Would Know
Researchers demonstrated that AI-assisted code can silently alter the digital output of computerized scanners used to process physical DNA evidence in crime labs — with no detectable trace left behind, the Wall Street Journal reported this week.
The finding exposes a chain-of-custody assumption that has underpinned roughly thirty years of forensic casework: that once evidence is scanned into a lab's system, the digital record can be trusted as faithfully representing the physical sample. If that link can be broken invisibly, defense attorneys and prosecutors alike face a new category of doubt in cases that already rest heavily on DNA identification.
The disclosure lands as courts are already wrestling with AI-generated and AI-manipulated evidence more broadly, and adds forensic labs to the growing list of institutions — alongside financial systems, medical records, and critical infrastructure — now re-examining decades-old assumptions about data integrity in an era of AI-assisted tooling.
London's AI Boom Is Running Into a Wall of Housing, Power and Water
The Financial Times reports growing strain in London — Europe's largest data-centre hub — as AI-driven server-hall expansion collides head-on with housing, power, and water constraints. A report commissioned by City Hall found London's 99 existing data centres already consume about 760 megawatts at peak demand, and the Greater London Authority has admitted that requested grid capacity from new projects is now ten times what the network can actually supply.
The pressure is not just electrical. More than three-quarters of UK data centres sit in the country's water-stressed south and east, and west London's grid has been fully subscribed since 2022 — a bottleneck that has already forced several housing projects in Ealing, Hillingdon, and Hounslow to be put on hold.
City Hall is now developing a standalone data-centre policy to balance the sector's economic value against the strain it places on the capital's most basic infrastructure — a tension likely to repeat in every other AI hub racing to add capacity faster than its grid, water table, or housing stock can absorb it.