John Langford, a partner researcher manager at Microsoft Research in New York, said publicly on 6 November that AI Frontiers — the group he helped found — is finished, and that he is leaving the company. The core of his reasoning: Microsoft AI's leadership is betting on scaling rather than research, and he saw no path to getting his innovations seriously used there. The practical consequence: there will be no internships in the group next year. Langford wrote that he found other places for his team before departing.
The group was created inside Microsoft Research in October 2023, months after ChatGPT's launch, and grew to about fifty people. Its record is hard to argue with: it produced AutoGen, a widely used framework for coordinating multiple AI agents, and Fara, an agent that operates a computer to complete tasks. In early September the unit was moved into Microsoft AI as part of a reorganisation — the division that owns the consumer and Copilot products.
That makes this more than a personnel note. It describes a tension that has shaped the industry for two years: research budgets aimed at the question of how to build smarter models with less data and compute, against the pressure to turn the next order of magnitude of GPUs into shipping products. Langford has argued his position at length elsewhere — that a human child learns a language on a fraction of the tokens a large model consumes, that a Portia spider plans a hunt for an hour on a brain a million times smaller. If those efficiency gaps are solvable, scaling cannot be the only path.
Whether the closure is a verdict on Microsoft's research strategy or merely on one organisational form is an open question. Microsoft Research still runs labs in Redmond, Cambridge and Beijing, and the company has bent its product line aggressively toward AI. What stands out is the pattern: AI Frontiers is not the first unit to be pulled out of the research organisation or folded into product groups in the past eighteen months.
For the roughly fifty people involved, the matter is settled. For everyone else, Langford's own question remains: if the big labs commit to scaling, who pays for the research that might one day establish that scaling has stopped working?

