
Microsoft AI veteran John Langford left the company on 30 October 2026 after struggling to get promising research seriously considered, according to his account. Five days earlier, China's ByteDance published a study exploring a similar approach to making artificial intelligence models more efficient.
In a 6 November blog post, Langford criticised the research priorities he encountered after Microsoft transferred his AI Frontiers group into its Microsoft AI division.
His departure highlights tensions between developing larger AI models and exploring alternative architectures.
Why John Langford Left Microsoft
Langford confirmed his departure on 31 October after 14 and a half years at Microsoft. He helped establish AI Frontiers within Microsoft Research in 2023. The group had around 50 researchers, according to his estimate.
Its projects included AutoGen, a framework for coordinating AI agents, and Fara, a computer-use agent. Researchers also explored new training methods and model architectures.
AI Frontiers moved into Microsoft AI in early September. Langford criticised what he described as the priorities of leaders previously associated with Inflection AI.
'The core philosophy there is scaling, and aggressively not research,' he wrote. He added: 'I could not get innovation into use (or even seriously considered) when I was in MAI.'
Langford declared AI Frontiers 'over'. However, Microsoft's 14 September update said the group had joined Microsoft AI and would continue its research. Its website still describes work on efficient language models and AI agents.
That statement preceded Langford's criticism and was not a response to it.
How ByteDance's Research Compares
On 25 October, ByteDance Seed published 'Recycler: Cost-Free Latent Reasoning via Hidden State Recycling'. Its method feeds information from a model's previous hidden state into the next processing step instead of relying solely on generated tokens.
ByteDance reported that Recycler outperformed certain models with two to three times as many parameters, particularly on reasoning and mathematics tasks.
The approach resembles the 'Full-bandwidth transformer' paper Langford co-authored in August. Microsoft's researchers also tested feeding previous hidden states back into a model to retain information between processing steps.
Their one-billion-parameter models matched or approached standard transformers trained on approximately 50% more tokens. However, these findings measure different things. ByteDance compared performance across model sizes, while Microsoft examined efficiency relative to training data.
Langford welcomed ByteDance's broader experiments. The similarities do not establish copying or prove either approach would deliver identical results in commercial systems.
Could Microsoft Have Saved Money?
Langford argued that ByteDance's experiments demonstrated the potential value of alternative model designs. He suggested that achieving comparable performance with a model half the size could substantially reduce training and inference costs.
Those savings remain hypothetical. Parameter counts, training requirements and operating expenses are related but different measures.
Neither study proves Microsoft could halve its AI spending or suffered financial losses because of its research priorities.
Where Langford Went Next
On 12 November, Langford announced he had joined AI research company Core Automation, describing it as 'very willing to try new things'.
His departure raises a wider question for AI developers: how much should they invest in more efficient architectures rather than relying primarily on increasingly large models?











