Microsoft Research’s AI Frontiers group is gone. One of its founding leaders says the roughly 50-person lab was folded into Microsoft AI (MAI) in early September, and that its leadership team has since left. The account comes from a personal post by that leader, which got wide attention on Hacker News. He says MAI, which is run by the former Inflection team, follows a philosophy that’s “scaling, and aggressively not research.”
He also confirms that the group’s researchers won’t be taking interns next year. “So AI Frontiers is over,” he writes.
⚡ The short version
- What happened: In early September, AI Frontiers (AIF) was reorganized out of Microsoft Research (MSR) and into Microsoft AI.
- Why it fell apart: MAI focuses on building bigger models and better data. It doesn’t fund new research ideas. The author also says there was “plenty of evidence that they just did not really want us.”
- Who left: The author says every other member of AIF’s founding leadership team seems to have reached the same decision he did.
- Why it matters: One of Big Tech’s best-known research labs has openly picked scaling over new methods, and the researchers who disagreed have walked out.
🧪 What AI Frontiers actually built
AIF started in 2023, right after ChatGPT launched. The idea was simple. One researcher working alone couldn’t keep up with rival labs that put large, focused teams behind their projects. So the founders tried to organize research into coordinated efforts, somewhere between top-down and bottom-up. Compute was often in short supply.
Even so, the group shipped work people noticed:
- AutoGen, a widely used framework for coordinating multiple AI agents
- Fara, a computer-use agent that operates software the way a person would
- Next Latent, a training objective designed to help models extrapolate
- Dion3, an improved version of the Muon optimizer
- Full Bandwidth Transformer, which uses recursion to get more power out of a model
AIF also built a shared monorepo that every Microsoft Research team could access by default. It held a pretraining codebase for edge-scale models. That sounds like a small detail, but the author says he knows of at least three other internal efforts building their own model pipelines, none of them in shared repos.
📉 Scaling vs. research
The author treats this as an argument about strategy, not just a reorg story. He’s fair about the case for scaling. Bigger models do make better predictions, and to a business, the payoff from scaling looks far more predictable than the payoff from research.
He still calls a scaling-only focus “both misdirected and not fun.” His evidence is ByteDance experiments suggesting that a variant of the Full Bandwidth Transformer might match a model twice its size. If that holds up, it roughly halves training and inference costs for the same capability. With industry spending getting close to a trillion dollars, efficiency gains like that are worth a fortune.
What stands out to me is the timing. Compute bills keep climbing and power is turning into a real bottleneck. That’s exactly when architecture and optimizer research tends to pay off most. Cutting that research loose now is a bet that brute force alone will stay cheap enough to win.
🏛️ Why he didn’t go back to MSR
The author gives two reasons for not returning to Microsoft Research:
- No path to real use. He couldn’t get new ideas adopted, or even seriously considered, while inside MAI. That makes him doubt research from MSR has much better odds.
- The deal is broken. MSR’s unwritten promise used to be “do good work and good things happen.” He says AIF did good work and was still “non-consensually reorged away.”
He expects more of the same “until MSR leadership develops a will to effectively organize to meet the moment.”
🔭 What to watch
- Talent flows: Researchers who focus on methods now have one fewer big-company home. Expect startups and research-friendly labs to recruit them.
- Open-source projects: If you depend on AutoGen or related Microsoft agent tools, keep an eye on maintenance activity and roadmap updates.
- Microsoft’s model strategy: MAI is building its own models alongside the OpenAI partnership. A scaling-only approach puts the competitive weight on data and compute.
- The efficiency argument: If methods like the Full Bandwidth Transformer prove out somewhere else, Microsoft will have funded the early work and handed the payoff to someone else.
The author ends on a hopeful note. He says it’s easy to picture a future Microsoft that becomes attractive for research again “if the people in control can develop a little bit more wisdom.” The full post and the Hacker News discussion have more detail.