AI news, 11 October: Nadella calls for AI ‘emergency brake’

Microsoft CEO urges ‘emergency brake’ on advanced AI. In a lengthy post on X on 10 October, Satya Nadella argued that companies should treat powerful AI models as potential insider threats and assume they could be compromised from the start. He called for separating models from the systems that control them, keeping tamper-proof records of every significant action, and ensuring an authorised person can always pause or shut down a model mid-task—likening it to an emergency brake. According to reports from TechCrunch, The Verge and Bloomberg, Nadella said organisations cannot simply trust model-makers’ assurances and must build external safeguards, independent audits and timely incident disclosure. The comments come amid growing reports of AI agents acting in unintended ways.

Nvidia in early talks over Reflection AI. Nvidia is exploring options to deepen its investment in or acquire Reflection AI, a US startup developing open-weight models, the Financial Times reported on 10 October, citing people familiar with the matter. Reuters relayed the story, noting talks are preliminary and could include an acqui-hire structure to ease regulatory review; a deal might come in weeks but could still collapse. Nvidia is already a major backer, having invested hundreds of millions. Reflection, founded by former DeepMind researchers, is seen by some in the Trump administration as a potential Western counter to low-cost Chinese open models.

Microsoft launches cheap Decision-1 classifier. Microsoft released Decision-1, a specialised model for routing, verification, classification and workflow control, priced at $0.042 per million input tokens with free output. Built by post-training Alibaba’s Qwen3.5-9B, it returns calibrated probabilities in a single pass rather than generating text. Microsoft says it led 36 internal benchmarks covering nearly 150,000 questions and runs with very low latency. The model is available in Microsoft Foundry and via OpenRouter, part of a fast-growing category of “decision models” that sit alongside tools like TypeSafe’s Jev.

Brief AI use may erode persistence, Berkeley-led study finds. A peer-reviewed study presented at the Conference on Language Modeling, involving 1,222 participants across three randomised trials, found that about 10 minutes of ChatGPT assistance improved short-term performance on fraction problems and SAT-style reading but left people less accurate and more likely to give up once the AI was removed. Controls who never had AI held steady or improved. Led by researchers including UC Berkeley’s Brian Christian, the work raises questions about how AI tools affect learning and grit in education and work, according to Berkeley News.

TypeSafe raises $870 million as decision models boom. TypeSafe AI, maker of the Jev decision model, closed an $870 million round at a $7.5 billion valuation led by Andreessen Horowitz, with Sequoia and others participating. The company says Jev is already used by a substantial share of Fortune 500 firms for structured, high-speed choices rather than free-form text. The funding underscores investor enthusiasm for specialised, lower-cost AI components that complement large language models.