Watch the video, transcript and sources
https://www.youtube.com/watch?v=P6jkXpDZRXM
- John Langford says AI Frontiers is over after its September move
+ Shared research infrastructure and a path to use remain the concrete questionsAn AI research group inside Microsoft sounds like a place with a future. John Langford, one of the people who helped organise AI Frontiers, says the group is over. In his November 6 post, he describes moving into Microsoft AI in early September, trying to place his colleagues and leaving. His account also explicitly mentions people left behind. About fifty was his estimate of the group's size; it is no firing tally.
The group began after ChatGPT arrived. Langford thought a lone researcher could arrive too slowly with baselines too weak to demonstrate a useful idea. AI Frontiers tried to organise stronger experiments, and its work included AutoGen and fundamental model research. His argument is that this collective effort was producing useful work despite severe compute bottlenecks.
The reorg changed the environment. He describes a scaling-only focus and says research was unwelcome. That is Langford's characterisation. He also acknowledges the business argument for scaling: its returns are more predictable than the returns from uncertain research. The company has its own public research ambitions. Suleyman's superintelligence announcement expressed those ambitions a year earlier, in November 2025; it establishes no current reply to this group's experience.
One example of the research is the Full-bandwidth transformer. The design feeds a previous top-layer hidden state into the next input alongside the sampled token. Its authors report improvements in tested models. Langford uses related research to argue for cheaper capability, then asks what halving costs could be worth. This episode establishes no such industry-wide saving. Parameter counts, training work and serving invoices need separate measurements.
The practical developer warning is clearer. AutoGen's current repository says Microsoft provides no bug fixes, security fixes, new features or support for the archived project. Its README points users toward Microsoft Agent Framework and migration guidance. The archive warning predates Langford's post. A departure and an archived repository have separate chronologies; check the maintenance status before copying an old tutorial into a production system.
The pipeline detail is the most concrete part of his organisational criticism. AI Frontiers developed a monorepo shared by default across Microsoft Research, including a model pretraining pipeline. Langford knew of at least four pipeline efforts; the other three were outside repositories shared across MSR. That statement says nothing about identical implementations or a measured waste bill. It does show why changing a reporting line and sharing infrastructure are different jobs.
He says he could not get innovation into real use, or even seriously considered, inside Microsoft AI. Returning to Microsoft Research felt like facing the same obstacle. The group had built shared infrastructure, and he had lost confidence in the route from an experiment to a product. That is the specific tension behind his exit account.
Verdict: NEEDS REVIEW
I'd want a research-to-product path and a maintenance plan before calling this consolidation a success. Langford's account gives me reasons to question both. Other founding leadership members “seem” to have reached a similar conclusion, in his wording. We have no verified departure roster and no evidence that all Microsoft research shut down.
What does your team share by default: research code, training infrastructure, or mainly the calendar invitations?
Sources
Visual credits and use boundaries
Full-bandwidth transformer is by Xi Wang and colleagues, paper v2, under CC BY 4.0. The exact source excerpt is reformatted; the accompanying model diagram is conceptual.
Mustafa Suleyman's authentic portrait is cropped from Village Global's November 17, 2022 Grove photo, with background removal and resizing, unflipped, under CC BY 2.0. Original Flickr source. It identifies Microsoft AI only; no authorship, closure order, current reaction or endorsement is implied. The public article recording keeps source text unchanged, with sidebar/comments hidden, larger article reflow and editorial cursor/highlight/camera emphasis.
Spider-Man/Imgflip template 110133729 is resized with a process-commentary caption; it supplies no identical-code evidence or actual executive reaction. Fantasia, Disney/@disney GIPHY, with Disney Tumblr provenance, is a silent, resized short excerpt with a narrator-commentary caption. The GIF shows a broom being brought to life; no multiplication claim is made. Both are brief transformed process commentary. Hosting is not a licence or legal clearance and grants no blanket reuse.
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.