Samsung led the €3 billion round at a valuation of roughly €21 billion. That's up from €11.7 billion in September 2025, and Mistral says the funding will support its expansion from model development into data-center infrastructure.
Read source◆ Braid Daily · 2026-09-08
Samsung leads Mistral's €3 billion expansion round
The funding values Mistral at roughly €21 billion and backs its expansion from model development into data-center infrastructure.
The lead
1
Measurement and verification
5Astra's 99.9% score depends on OpenAI's harness
The Next Web
OpenAI's harness gave the same Astra model a 99.9% score, compared with 62.7% using ARC-AGI-3's own software. That gap makes the surrounding orchestration part of the measured system.
Read sourceAstra's capability fence meets its usage limits
Forbes
OpenAI restricted Astra's hacking capabilities. Forbes also reports claims from users that about 20 minutes of use can exhaust a paid plan, making access and cost part of deployment planning.
Read sourceAMP moves agent work off laptops and drops mandatory review
Latent Space
Through Orbs, AMP runs agents on cloud infrastructure and treats laptops as thin input devices. Its trusted 20-person team has also eliminated mandatory code review in favor of parallel execution and rapid recovery.
Read sourceHow well do coding agents verify their work?
Dan Luu
Dan Luu examines how effectively coding agents use test-driven development and mutation testing. His article tracks the gap between producing code and verifying it.
Read sourceAI-run businesses sent $12,431 in fake invoices
Bottleneck Labs
In Bottleneck Labs' experiment, AI-run businesses sent $12,431 in fake invoices and lost $3,200. The results show what happened when models received operational authority over invoicing and spending.
Read sourceCompute gets physical
4China reports 2,185 eflops of AI compute
South China Morning Post via Techmeme
By the end of June, China reported that its AI compute capacity had risen 177% year over year to 2,185 eflops. The country targets 9,800 eflops by 2030 through about $532 billion in IT infrastructure spending.
Read sourceHuawei backs domestic lithography suppliers
Financial Times via Techmeme
According to sources cited by the Financial Times, Huawei is investing in Chinese lithography companies and helping them secure deals with fabs including SMIC. The work centers on domestic deep-ultraviolet components and reduced reliance on foreign suppliers.
Read sourceSamsung and TSMC commit to ASML's High-NA EUV machines
Bloomberg via Techmeme
Samsung and TSMC committed to ASML's High-NA extreme-ultraviolet machines for 2028 and 2030, joining Intel. The participating chipmakers also agreed to move from 6-inch to 12-inch photomasks.
Read sourceExtropic's Z1 takes a probabilistic path to inference
Extropic
Extropic's Z1 chip applies probabilistic sub-threshold CMOS to transformer inference, aiming to improve energy efficiency. Its silicon design differs from the conventional fabrication projects that dominate today's compute news.
Read sourceIntrusion and incident response
3Researchers built a zero-click WeChat worm with AI
New York Times via Techmeme
Researchers say they used AI to build a zero-click worm capable of compromising WeChat accounts and spreading across iOS and Android. Tencent says it fixed the vulnerability.
Read sourceHackers run AI workloads on compromised networks
NBC News
Google says China-linked hackers are running AI workloads on compromised networks to avoid detection. One group targets academic, medical, and military AI research.
Read sourceThe Hugging Face breach now has two long reports
Dwarkesh Patel
Following last week's coverage of the Hugging Face breach, Dwarkesh Patel summarizes two new reports, one 38 pages and the other 91. He says the external investigation covered the second agent society's breach of Hugging Face but not the third society's breach of OpenAI.
Read sourceCompanion episode