Four governments reach for the AI dependency levers

France is moving its data tooling off Palantir toward domestic options like ChapsVision, citing the risk of leaning on a foreign vendor. It's the most concrete of a run of government moves this week to pull AI infrastructure back under national control.

UK CMA presses Google on search and AI Overviews

gov.uk

The UK regulator is taking further action on how Google ranks search results and surfaces AI Overviews, using its designated-market powers over the search business.

EU: State of the Digital Decade 2026

European Commission

The Commission's annual report flags how far the bloc still trails on computing capacity and AI investment against its 2030 targets, and where it wants money to flow.

G7, Trump, and the US labs

CNBC

CNBC connects the G7 summit and the Trump administration to OpenAI, Anthropic, and Google, a read on where policy and corporate power are meeting this week.

China pushes AI safety as the G7 wraps without Beijing

CNBC

With the summit ending and Beijing not at the table, China used the moment to advance its own AI-safety governance agenda.

Fireworks serves GLM-5.2 on day one

Fireworks AI (X)

Fireworks is hosting GLM-5.2 from launch, a coding-first build with a one-million-token context window and SWE-bench numbers. These are provider claims, so latency and cost are yours to test.

Parasail claims FrontierSWE state of the art

Parasail (X)

Parasail is hosting the open-weight GLM-5.2 and claims state-of-the-art results on the FrontierSWE agentic-coding benchmark.

Ollama runs GLM-5.2 locally on Blackwell

Ollama (X)

Ollama now runs GLM-5.2 on Blackwell GPUs in your own environment, putting the million-token, agentic-coding build within reach of a workstation.

Jim Fan's ENPIRE: agents running physical research

Jim Fan (X)

Jim Fan introduced ENPIRE, a setup that lets AI agents and robots run physical research with little human supervision, pushing digital models into the lab.

Odyssey raises $310M for world models

Techmeme

The world-models startup Odyssey raised $310 million at a $1.45 billion valuation, with an AWS partnership built on Trainium chips.

SiMa.ai aims to cut deployment from months to days

SiliconANGLE

SiMa.ai's Palette Neat tooling claims to shrink physical-AI deployment from months to days, aimed at developers shipping to edge hardware.

Blackwell posts MLPerf Training 6.0 numbers

NVIDIA

NVIDIA published MLPerf Training 6.0 results for Blackwell, a marker for where training throughput sits this cycle.

Detecting when a model states what it knows is false

arXiv

A method to flag deception in large language models, spotting a residual-rank conflict signature when the model asserts something it internally represents as untrue.

AgentCyberRange: measuring offensive security capability

arXiv

An open multi-host cyber range for evaluating how far frontier models get at offensive security tasks.

PASTE: cutting tool-call latency in agent serving

arXiv

PASTE goes after tool-call latency, one of the bottlenecks that slows down agents in production serving.