The departure
According to a report by IT之家, citing an article from Windows Latest, John Langford, a longtime AI scientist and cofounder of the AI Frontiers lab, has left Microsoft after more than 14 years. The news was reported on November 14, 2026. Langford is described as a prominent machine learning researcher, known for his contributions to contextual bandits and for his work on exploration, learning-problem reduction, and offline policy evaluation.
According to the source, the reasons for the departure lie in the lack of a clear path for applying research results and the difficulty of gaining serious attention for innovative ideas. His team of about 50 people had been reorganized in September and placed within the Microsoft AI division led by Mustafa Suleyman. The disagreements concern the research direction, the channels for bringing innovation to market, and the very manner of the reorganization.
The scale issue
Langford acknowledged that increasing model size improves performance and that this path offers more predictable returns for companies. His criticism is not aimed at scalability itself but at its exclusive centrality: focusing only on scaling up, he argues, shifts attention to data and reduces the room to explore new methods.
The specific point of friction is the 'full-bandwidth Transformer' research published by Langford in August. The study proposes that, when generating the next token, the model reuses intermediate internal states in addition to the tokens already emitted, reducing information waste. In experiments, a model with 1 billion parameters reportedly achieved performance close to that of a standard model trained on about 1.5 times the data, with a limited additional computational cost per token. According to the account, Microsoft AI instead prioritizes expanding model scale.
An echo in the industry
The source notes that on October 25, 2026, ByteDance's Seed team published research called Recycler, which follows a similar approach in reusing hidden states. According to the report, this method would outperform models with two or three times the parameters in reasoning and mathematics. This is an indication of how the issue of reusing internal states is also emerging outside Microsoft.
It remains to be verified what actual impact these research efforts have on the two companies' industrial strategies and whether the disagreements described have broader consequences for the organization of research at Microsoft AI. The material does not contain official reactions from Microsoft or Langford beyond what is reported, nor details on any subsequent developments.
