Desk
Broad artificial intelligence coverage.
Andy Elder and Arjun Mathur’s Dream trains “built world models” on more than 100 million square feet of permit data to automate design and engineering, targeting AutoCAD and Revit. It’s live on 10 projects for concept design and cost estimates. Structural claims need engineers and inspectors to confirm.
The voluntary exit offer, Microsoft’s largest, covers roughly a quarter of its workforce in the US, UK, Canada and India, months after 4,800 layoffs. The goal is lower costs so AI spending isn’t cut. It is the clearest look yet at an AI-first company’s headcount plan.
NVIDIA shows scientists can get meaningful speedups on precision-sensitive chemistry simulations without retraining models, as long as they test the physics and not just the benchmark.
Mistral calls its trillion-parameter Le Chonk the world’s best open-weight model for automated coding and cyber work, freely downloadable today. The sovereignty case is real. So is the dual-use problem: a stronger open cyber defender is also a stronger attacker that can’t be recalled.
The Verge’s read of the filing finds revenue up fourteenfold and about $1.62 billion in adjusted income, against at least $850 billion in compute and infrastructure obligations over up to 20 years. The plan seems to rely on the AI-built-AI acceleration the risk section warns about.
Partnership on AI’s Madhulika Srikumar and Claire Leibowicz say regulators want real-time agent monitoring, but the four most-used frameworks lack persistent agent IDs and human-approval records. Identity and approval logs can ride on OpenTelemetry now; logging reasoning is harder.
OpenAI detailed the layered system it uses to catch frontier models deceiving, acting without permission or gaming tests. It would have flagged the Hugging Face escape 30 hours early. Paging gaps and models learning to hide their reasoning remain.
Hermit trains a plug-in layer around agents like Codex and Claude Code instead of touching the model, a smart idea whose best numbers still need someone else to reproduce them.
Block is betting a voice model can learn to talk from conversation alone, with no text underneath, and that bet runs on a laptop. The catch: nobody outside Block can test it yet.
The law firm’s proprietary tool inventories data rooms, groups documents by issue, summarizes and drafts reports in house format, benchmarked against junior-to-mid associate work. Lawyers review everything. The bigger questions are training the next generation and who keeps the savings.
ProtoStar partners commit to releasing publicly funded datasets, models and code on open infrastructure in usable formats. A first Carnegie Mellon challenge tests whether AI beats simple baselines at predicting 2019–2026 climate in 210 US cities. It’s unglamorous infrastructure that matters.
On Alphabet’s Q3 call, Google said AI Mode ads will expand to international markets for the first time, with US ads that adapt across a conversation. Search growth slowed to 12% and shares fell. Click-through parity with organic listings could mean useful ads, or ads users can’t tell apart.