Particle.news
Download on the App Store
21 ARTICLES
·
3h ago

Microsoft Unveils First In-House AI Models as OpenAI Talks Stall

The launches signal a push to reduce reliance on OpenAI, with unresolved contract terms threatening OpenAI’s restructuring and a SoftBank investment.

Microsoft has started making its own AI models amidst reports of increasing tension with OpenAI.
Satya Nadella, chief executive officer of Microsoft Corp., speaks during the company event on AI technologies in Jakarta, Indonesia, on Tuesday, April 30, 2024.
Image
Microsoft has released two AI models, emancipating itself from an overreliance on OpenAI's technology.

Overview

  • Microsoft introduced MAI-Voice-1, a speech model that can generate a minute of audio in under one second on a single GPU, now powering Copilot Daily and Podcasts and available to try in Copilot Labs.
  • MAI-1-preview, trained on roughly 15,000 Nvidia H100 GPUs, is in public testing on LMArena and will be phased into select Copilot text use cases in the coming weeks, with a form open for early developer access.
  • Initial LMArena results place MAI-1-preview around 13th for text tasks, trailing leading models from Anthropic, Google, OpenAI and others, as Microsoft emphasizes efficiency and cost-focused design.
  • Microsoft AI chief Mustafa Suleyman says the strategy centers on consumer-first, specialized models and a multi-model approach; the company notes an operational GB200 cluster and a multi-year roadmap.
  • Negotiations between Microsoft and OpenAI are deadlocked over an AGI cutoff clause, API and cloud hosting rights, and revenue/equity terms that would set Microsoft’s roughly 30%–35% stake, risking delays to OpenAI’s restructuring and a conditional ~$10 billion SoftBank funding round.