Microsoft Research's AI Frontiers group, about 50 people, is gone after a September reorg into Microsoft AI, and its founder has left
- AI Frontiers, a roughly 50-person group created inside Microsoft Research in 2023 after ChatGPT's release, was reorganized into Microsoft AI in early September and is now over; founder John Langford and the other founding leadership members left.
- Langford says MAI's ex-Inflection leadership runs a scaling-only philosophy that is deliberately not research, so the main work there is getting and refining data rather than new methods.
- AIF built the Autogen orchestration framework and the Fara-7B computer-use agent, and its NYC members worked on the Next Latent objective, Dion3 (an improved Muon) and the Full Bandwidth Transformer.
- Langford points to ByteDance experiments where a Full Bandwidth Transformer variant was comparable to a model twice the size, and asks what halving training and inference cost is worth against industry investment approaching $1 trillion.
- There will be no AI Frontiers internships in the coming year, and Langford rules out returning to MSR because research there has no path to real use and the 'do good work and good things happen' deal broke.
Hacker News 의견들
More context here: archive.is/AaY5O. The original hunch.net link got hugged to death like it's 2008 and you just lit your LAMP stack.
Took me a second on the title. This is about a group literally named AI Frontiers, not the frontier of AI research coming to an end.
Dedicated labs have never been great at commercializing their own tech, the skillset is different from running a product business. Still, latent reasoning and separating intelligence from knowledge are wide open.
This is happening at every big company now. Transformer scale-out is a straight highway and nobody is allowed to take the exit. Meta basically fired the guy who said transformers won't scale.
Almost everything LeCun said LLMs couldn't do turned out wrong, so letting him go was the right call.
When your company has paid billions to OpenAI, any activity that doesn't dance in support of that investment reads as disloyalty.
Where's the money for 10 trillion and 100 trillion parameter runs? If scaling were really the road to god in a box, compute demand would be infinite and nobody building for it would carry junk debt ratings.
100 trillion parameters makes no sense against a roughly 30 trillion token dataset. You want something like 10x tokens per param, so 300 to 1000 trillion tokens, which is nonsense.
Honestly I barely notice gains since 4.6 for coding work. Same mistakes, same strengths. The best improvement lately was Haiku 5.5, cheap workhorse for implementing plans.
The headline is bait for AI skeptics. It's 50 people with a grandiose name. The interesting question is whether MS has given up on AI research entirely because its money elsewhere is a better bet.
Curse of money. A company with a market cap near the trillions decides 50 researchers are too expensive.
If MS dropped something promising, doing it outside makes sense. There's investor money for this and it could attack the moat of the frontier labs.
Microsoft is a rough place to work right now, there are a hundred stories like this across the company. They've offered voluntary redundancy to a big share of staff.
Every large tech company is like this at the moment, maybe except OpenAI and Anthropic since they're still building something new.