Good morning. Europe just wrote its largest tech check ever, and it went to an AI lab that keeps insisting the world doesn’t have to run on American APIs. Elsewhere, the tooling story continues to quietly mature — sandboxes for agents, speculative decoding on AMD, and yet another attempt to figure out whether AI is actually making anyone employed or unemployed.
Mistral raises €3B at a €21B+ valuation. The Samsung-led Series D is the largest equity round ever for a European tech company, and Mistral is leaning hard on the “sovereign, open-weight” pitch to differentiate from OpenAI and Anthropic. The money goes toward compute, frontier research, and expansion across 20 countries. HN reception was muted — the loudest comment was distrust of existing investor a16z rather than any real engagement with the sovereignty argument, which tells you something about how that framing lands with US developers.
The Economist says AI is a net job creator, so far. The magazine estimates roughly 1 million new AI-related jobs in America against about 200,000 AI-attributed layoffs since mid-2023, driven mostly by data center construction. HN was not buying it. Multiple commenters described months-long job searches and a string of failed AI startups they’d consulted for, and several pointed out the obvious catch: construction jobs end when the buildings are finished, and the models keep getting better in the meantime.
Google’s WeatherNext 3 forecasts from raw satellite imagery. DeepMind’s new model pulls directly from satellite data to generate hourly forecasts, with an interactive world map that HN commenters actually praised. The recurring complaint: Google’s consumer weather app remains famously bad, so it’s unclear when any of this reaches users. One commenter flagged that recent US cuts to meteorological data collection may undercut initial conditions regardless of how good the model gets.
Two new sandboxes for coding agents. Trail of Bits released Coop, a Rust CLI that spins up disposable VMs (Lima on macOS, Firecracker on Linux) for running Claude Code and Codex with full tool access but no path to your home directory. Separately, a developer reverse-engineered the VMs that hosted mobile agents like Claude Code actually use, finding Firecracker microVMs with custom Rust init, read-only vendor partitions, and a separate writable disk that persists user data across VM reclamation. Some platforms apparently store agent memory in git, which is a nice touch if you want to actually audit what your agent has been thinking.
Speculative decoding lands on AMD in vLLM. The vLLM team published a detailed writeup of five drafting methods on MI300X and MI355X, with gains that vary heavily by model and acceptance rate. The most interesting comment on HN wasn’t about the datacenter chips at all — a user noted that stock vLLM on the workstation-grade R9700 runs at 20–30 tokens/sec, while community fork Radiance hits 150–200. Both AMD and the vLLM team are apparently uninterested in closing that gap.
Two small but useful evaluations. A developer ran 10 model/harness combinations on the same Three.js hangar task, with Qwen 3.8 27B on OpenCode landing the best time-to-quality ratio. Fair criticism in the comments: no combination was run more than once, so variance is unknown. Separately, Dan Luu tested 26 prompting conditions for agent testing and verification on a Rust Zstd implementation, finding formal methods like Lean 4 and Verus underperform — unsurprising given how little training data exists for them. One commenter argued the whole premise misses the point, since most testing quality comes from architecture, not framework choice.
Arm’s CEO says the chip shortage is delaying cancer cures. Rene Haas told the BBC that AI will eventually cure cancer if only there were enough chips, a claim a cancer researcher in the same piece politely disagreed with, pointing out that patient sample data matters more than FLOPs. The HN response was about what you’d expect: “Feline cancer cure slowed by shortage of treats, says my cat.”
That’s it for today. The sandbox tooling is the quiet story worth watching — every serious agent deployment is converging on the same answer, which is “don’t let it touch your actual machine.”