Microsoft AI Veteran of 14 Years Departs, Citing Difficulty Turning Innovation into Products and Lack of Serious Consideration for Scaling Research

Microsoft 14-year AI veteran departs: says innovation is hard to ship, cost-reduction research not seriously considered

IT Home, November 14 news: Tech media outlet Windows Latest published a blog post today (November 14), reporting that an AI scientist who worked at Microsoft for more than 14 years and co-founded the AI Frontiers lab, 约翰 · 兰福德 (John Langford), has left the company. He said the key reason for leaving was that he could see no path to real-world application for research results, and that innovative ideas were even hard to get seriously considered.

According to public information checked by IT Home, 兰福德 is a senior and influential researcher in the machine learning field, having long worked on machine learning and learning theory research. His important contribution is advancing Contextual Bandits; with collaborators, he has done representative work on exploration algorithms, reductions of learning problems, and offline policy evaluation, providing a theoretical and methodological foundation for machine learning systems to improve decisions using incomplete feedback.

兰福德's roughly 50-person team underwent restructuring in September this year and was merged into the Microsoft AI division led by Mustafa Suleyman, but there were major disagreements over research direction, channels for shipping innovation, and the way the team was reorganized, so 兰福德 decided to leave.

兰福德 acknowledged that scaling up models can indeed improve prediction results, and that for companies, the returns from this route are easier to predict clearly. But what he opposes is treating scaling as the only core, which would over-concentrate the work on data and squeeze out room for exploring new methods.

兰福德 personally leans toward research innovation, while the Microsoft AI division places more emphasis on scaling up AI models. The dispute mainly concerns the "full-bandwidth Transformer" research 兰福德 published in August this year.

That research argues that when a model generates the next token, it does not only use previously output tokens but also reuses internal intermediate states, reducing information waste. In the paper's experiments, a 1-billion-parameter model could, with less training data, approach or match the performance of a conventional model trained on about 1.5 times the data, while the additional computation cost per token was very small.

The report noted that ByteDance's Seed team released the Recycler research, which adopts a similar approach, on October 25 this year; the method reuses hidden states in a similar way and, in reasoning and mathematics, can surpass models with two to three times as many parameters.

Related articles

• No subsidies, no shifting costs to others: Microsoft promises to pay its own electricity bills for AI expansion

• Microsoft officially releases .NET 11: faster development, more complete AI tools

• AI boom collides with resource disputes as Microsoft's under-construction data center in the Netherlands faces protests

• Report: ByteDance to broadly restrict employees from using Claude, Codex and others internally at their own expense

• Sources say ByteDance Seed core lead 周畅 leaves to start a business in physical AI

• Not ByteDance: MediaTek lands second major AI chip customer, mass production in 2028