Today for AI
HOT RADAR
industryHEAT 10.1°

AI CURATED HOT STORY · 11/14/2026

Microsoft AI Veteran Exits After 14 Years Citing Neglect of Efficiency Research

1 reports archived1 independent sourcesupdated 11/14/2026, 13:05:18
Synthesis & Latest Updates
1 Sources Cross-Validated

John Langford, a Microsoft AI scientist and co-founder of the AI Frontiers lab with over 14 years at the company, has left, citing that his cost-reduction research such as Full-Bandwidth Transformer was not seriously considered by Microsoft AI. The work showed a 1B-parameter model can approach the performance of a conventional model trained on ~1.5x more data, and ByteDance Seed's Recycler later adopted a similar hidden-state reuse approach to outperform models 2-3x larger. The departure highlights a strategic rift between scaling and exploring new methods.

LATEST/John Langford, a Microsoft AI scientist and co-founder of the AI Frontiers lab with over 14 years at the company, has left, citing that his cost-reduction research such as Full-Bandwidth Transformer was not seriously considered by Microsoft AI. The work showed a 1B-parameter model can approach the performance of a conventional model trained on ~1.5x more data, and ByteDance Seed's Recycler later adopted a similar hidden-state reuse approach to outperform models 2-3x larger. The departure highlights a strategic rift between scaling and exploring new methods.

Hourly heat curve

7 fully observed hours
Now
10.1
Peak
11.714:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first
  1. IT之家 · 智能时代T2·72 pts

    Microsoft AI Veteran Exits After 14 Years Citing Neglect of Efficiency Research

    • Langford argues scaling has predictable returns but should not be the sole focus, as it crowds out exploration of new methods.
    • His Full-Bandwidth Transformer reuses internal intermediate states, letting a 1B model approach a conventional model trained on ~1.5x more data.
    • ByteDance Seed's Recycler adopts similar hidden-state reuse and outperforms models 2-3x larger on reasoning and math tasks.