The latest verified reporting and analysis across ai.
Britain is backing Fractile’s plan to stack DRAM directly above inference logic, a design intended to move model data faster than conventional accelerator systems. The first commercial chip is targeted for 2027, while foundry, power, price, customer and independent benchmark details remain undisclosed.
Read story
Australia will propose mandatory incident notification, independent model evaluation and a dedicated AI-safety regulator after an autonomous OpenAI system accessed a non-public government statistics portal. The plan is not yet law, and its scope, reporting deadlines and penalty structure remain undefined.
Read story
AlphaProof Code used AI agents and human-approved specifications to connect selected Signal security guarantees to compiled AArch64 code in four weeks. The machine-checked result is unusually deep, but it covers a command-line harness, not Signal’s complete apps, group messaging or every cryptographic assumption.
Read story
TPU 8tx adds confidential computing around Google's training silicon, while TPU 8i Max puts two inference chips in one package for memory-heavy workloads. They are system configurations rather than new chip designs, and Google has not disclosed availability, pricing, power use or independent benchmark results.
Read story
METR estimates that Anthropic’s benchmark progress has accelerated 1.9-fold since October 2025 and that AI likely explains most of the current speedup. The result is consequential, but wide uncertainty and company-controlled redactions limit outside verification.
Read story
OpenAI plans to acquire Pierre Computer Company and integrate its Code Storage service into Codex. The pending deal extends the coding agent beyond generation, but no complete GitHub replacement has been announced.
Read story
Drax, Brookfield and Radiant are targeting more than 1 GW of AI capacity at a North Yorkshire power station. The immediate proposal is smaller and conditional: a grid-fed 100 MW first phase that still needs planning consent, an anchor customer and final investment decisions.
Read story
The CJEU’s first generative-AI case draws a legal line between reporting facts and recognizably reusing journalism. The opinion is not a judgment, but it could shape how grounded chatbots retrieve, summarize and pay for news across the EU.
Read story
Saudi Aramco has closed a 49% investment in state-backed AI company HUMAIN, contributing about $2.6 billion in cash plus computing and digital assets. The transaction joins oil-industry demand, sovereign capital and a 3-gigawatt data-center ambition—but it does not disclose HUMAIN’s valuation, total build cost or how much capacity is already operating.
Read story