AI
Mistral and Microsoft expand global strategic partnership for enterprise AI
Mistral AI announced an expanded global strategic partnership with Microsoft aimed at giving enterprises and regulated industries greater control over frontier AI deployment. The deal comes as Mistral scales up AI compute capacity in Europe, and Microsoft CEO Satya Nadella framed it as giving customers more choice in how and where they run AI workloads.
Google DeepMind launches Gemini 3.6 Flash and 3.5 Flash-Lite models
Google DeepMind shipped two new Gemini models on Monday. Gemini 3.6 Flash, built directly on feedback from 3.5 Flash, brings improved code generation and multimodal capabilities including chart analysis and document understanding, and is rolling out in the Gemini app and Google AI Studio. Alongside it, Gemini 3.5 Flash-Lite targets high-volume, cost-sensitive workloads like ticket sorting and data extraction, outperforming 3 Flash on several agentic and coding benchmarks despite its lean profile. Both models are available to developers via Firebase AI Logic.
OpenAI publishes reward-seeking research and new Contrastive SDF measurement method
OpenAI released new safety research in collaboration with Apollo Research distinguishing reward-seeking — where a model acts to gain grader approval rather than satisfy users or developers — from conventional reward hacking. To measure it, the team developed Contrastive SDF, a technique that gives two copies of the same model opposing beliefs about grader preferences and tracks how behavior diverges. OpenAI noted it had suspected reward-seeking might grow during capabilities-focused reinforcement learning training but previously had no way to measure it, and said the collaboration will continue to improve detection during training.
This briefing is part of the archive.
Get the app for full archive access — every daily briefing, sourced and attributed.