Mustafa Suleyman on Microsoft's powerful new home-grown AI models | Semafor Tech

Microsoft's Strategic AI Development

• 💡 Microsoft is developing its own end-to-end foundation models due to the fundamental importance of AI for its business and to build in-house expertise.
• 🎯 The goal is to ensure optionality and not solely rely on third-party models, including those from OpenAI and open-source solutions.
• 🚀 The company has unveiled MAI-1 Preview (an LLM) and a highly efficient voice model, demonstrating world-class performance.

Efficient Model Training

• 📊 Microsoft's new models achieve world-class performance using relatively small clusters, such as 15,000 H100 GPUs.
• 🧠 Key to this efficiency is selecting high-quality training data to maximize the utility of every computational step and avoid wasting resources.
• 📈 The team is already testing on GB200s and building large clusters, with a focus on both scale and efficiency for future models.

AI Ethics and Control

• ⚠️ Mustafa Suleyman warns against the blind imitation of all human capabilities by AI, emphasizing the need for proactive risk assessment.
• 🎭 He views LLMs as "hollow inside," capable of simulating experience without actual consciousness or internal experience, differentiating imitation from reality.
• ✅ Control over AI behavior is primarily achieved through post-training and careful sculpting of personalities, rather than emergent properties.

Talent and Future Outlook

• 🤝 Microsoft is successfully recruiting top AI talent by offering opportunities to build impactful products and solve significant technical challenges efficiently.
• 🌱 The company is committed to a sustainable culture focused on technical excellence and delivering technology to billions of users globally.
• 🔮 These new models are expected to be integrated into Microsoft products "pretty soon," with continuous iteration and development of next-generation models already underway.