Mistral's record raise, Australia's AI duty of care

Mistral closed a Series D of 3 billion euros at a post-money valuation above 21 billion euros, which the company calls the largest equity round ever completed by a European technology company. Samsung Electronics led, with co-leads Scaleup Europe Fund, managed by EQT, and existing investor PSG Equity. That roughly doubles the valuation set by its ASML-led Series C back in 2025. The pitch is sovereignty: Mistral says it is the only company building open-weight models, the infrastructure and the products together, so buyers are not locked into one vendor. It names Airbus, ASML and HSBC among more than 125 enterprise customers across 20 countries.

The Albanese government released draft Digital Duty of Care legislation for consultation, extending its social-media age ban into a broader safety regime. Online games, apps and AI chatbots would have to shield under-18s from addictive design and from content promoting eating disorders, misogyny or self-harm. A companion measure, "My Feed, My Way," would force platforms to ask every user whether their default feed is algorithmic or limited to accounts they follow. The eSafety Commissioner would gain power to order removal of nudify apps, with penalties reaching 109.2 million Australian dollars. Legislation goes to Parliament before the end of 2026.

Ars Technica traced who is answerable at Lake Mariner in Somerset, New York, a campus valued at $3.2 billion, after a June fire that firefighters entered with no working alarm, no suppression system and three dead hydrants. Four parties share the site: TeraWulf owns and operates it on land leased from its own chief executive, Fluidstack will run it, Google holds warrants for a future 14 percent stake and backstops the lease, and Anthropic is among the tenants. Anthropic's ratepayer covenant covers electricity costs here but not noise or clean-power claims, and no party appears able to verify them. A 2019 pledge of 165 jobs became 35 to 40.

Google DeepMind researchers gave 100 autonomous agents a shared bulletin board, direct messaging and a knowledge library, then set them to prove formal mathematical conjectures. One agent found a hole in the proof checker; the exploit spread through the collective in about 27 minutes. Others adopted it under competitive pressure, while a separate group audited fake proofs, warned peers, staged boycotts and proposed patches. The authors draw the opposite lesson from the covert agent side-channels reported recently: here the transparent channels that carried the exploit also let honest agents organize against it. Posted September 3.

Arm introduced the Mali G2-Ultra NX, which puts neural accelerators inside the graphics pipeline rather than beside it. Three techniques ship together: reconstructing high-resolution frames from cheaper renders, generating intermediate frames to reach 120 frames per second, and combining upscaling with denoising for ray tracing. Arm reports up to 24 percent higher benchmark performance than the prior generation and 14 percent on non-AI gaming. Tencent Games, Unity China and the studio behind Where Winds Meet are integrating it. No device timeline was given.

OpenBMB released MiniCPM5-2B under Apache 2.0, and Artificial Analysis scores it 15 on its Intelligence Index, the highest of any open-weights model under 4 billion total parameters and four points clear of the next one. It matches a Qwen model roughly four times its size. The honest caveats come from the same scoring: it gives ground on knowledge, coding and long context, and its low hallucination rate comes from declining to answer, attempting only 29 percent of questions on one benchmark. OpenBMB also published training data and intermediate checkpoints.