Microsoft’s AI division is gearing up for a "superintelligence" era, its new chief, Mustafa Suleyman, said on the Decoder podcast. The conversation, recorded during the company’s Build developer conference, focused on how Microsoft is reshaping its relationship with OpenAI, rolling out a slate of new models, and doubling down on enterprise AI.
In October of last year, Microsoft and OpenAI signed a refreshed agreement that extends their partnership while allowing Microsoft to pursue its own superintelligence agenda. Suleyman explained that the deal gave the company the freedom to build large‑scale clusters, hire a dedicated team, and fund the hardware required to train frontier models.
Since then, Microsoft has launched seven new models spanning text, code, audio, image and transcription. The flagship reasoning model, MAI‑Thinking‑1, scores 97 percent on the AIME benchmark and rivals top‑tier models such as Opus 4.6. The transcribe model, MAI‑Transcribe‑1.5, claims the world’s best cost‑efficiency and accuracy. An image model ranks second globally, while an image‑editing model sits at third behind Google and OpenAI. CodeFlash, optimized for Visual Studio Code, matches the performance of OpenAI’s Sonnet 4.6.
Rather than distilling existing models, Microsoft chose to train these systems from scratch. Suleyman said the effort involved "extremely high‑quality, carefully curated" data and a proprietary pipeline that filters out low‑quality and security‑risky content. The company also highlighted its new Maia 200 chip, which delivers a 30 percent cost advantage over competing silicon and, when co‑optimized with MAI‑Thinking‑1, yields a 1.4× performance‑per‑watt boost.
Enterprise customers are at the heart of Microsoft’s AI strategy. With nearly all of the Fortune 500 running workloads on Azure, the company can deliver AI services that integrate tightly with existing data and processes. Suleyman noted that while consumer sentiment toward AI is mixed, enterprise users are already seeing “massively beneficial” impacts, especially in software engineering where AI‑generated code accelerates development cycles.
Healthcare also received a spotlight. Microsoft announced a long‑term partnership with the Mayo Clinic to co‑train a health‑focused foundation model using the clinic’s extensive longitudinal patient records. The goal is to deploy the model in hospitals worldwide, leveraging Microsoft’s cloud and AI expertise to improve clinical outcomes.
The interview touched on broader industry concerns. Suleyman warned against anthropomorphizing AI, emphasizing that current models do not possess consciousness or suffering. He reiterated a “humanist superintelligence” philosophy: AI should serve humanity, making people healthier, happier and more capable. Governance, accountability and responsible deployment remain central to Microsoft’s roadmap.
Looking ahead, Suleyman said the company will continue a six‑to‑eight‑week development cycle, with in‑person retrospectives that keep teams agile. He also hinted that future devices—badges, wearables and edge processors—will blend local inference with cloud power, reshaping how users interact with AI beyond the smartphone.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.