The Allen Institute for AI spent years building the most transparent AI model in existence. Then Microsoft hired the team that made it.
Ai2 (source), the nonprofit research lab founded by the late Paul Allen, created OLMo, a family of AI models that released everything to the public. Not just the model itself, but the training data, the code, the intermediate checkpoints, the evaluation tools. In an industry where companies treat their models like trade secrets, Ai2 designed OLMo so any researcher on earth could see exactly how a frontier AI system gets built.
Beginning with the departure of Ai2's CEO Ali Farhadi in March, followed two months later by COO Sophie Lebrecht and a string of key researchers including project lead Hanna Hajishirzi, according to GeekWire (source). They joined the team building what Microsoft AI CEO Mustafa Suleyman has called "Humanist Superintelligence." (https://microsoft.ai/news/towards-humanist-superintelligence/)
Ai2 had just brought online a federally funded supercomputer called OMAI (source), backed by $75 million from the National Science Foundation and $77 million from NVIDIA. The hardware runs NVIDIA's latest Blackwell Ultra GPUs, built specifically so open-source researchers could compete with corporate labs on raw compute power. Weeks after the cluster went live, the people it was built for were gone.
OLMo wasn't just another open-source model that lets you download the weights and call it a day. The team released over 30 distinct components (source) on Hugging Face, from base models to training data mixes to reward models. Ai2's own research estimates that 82% of a typical training effort (source) is exploratory work that never makes it into the final product. When that work is shared publicly, every GPU hour benefits the entire research community. When it stays inside a corporation, it benefits one company.
Clem Delangue, CEO of Hugging Face, said (source) OLMo gives the AI community "full visibility into a state-of-the-art large language model in order to confront the problems with existing LLMs."
Eric Horvitz, Microsoft's Chief Scientific Officer, is a founding member of Ai2's Scientific Advisory Board (source). He publicly praised OLMo and called himself "enthusiastic about getting OLMo into the hands of AI researchers." His company then hired the researchers who built it.
Microsoft's reasoning tracks with what came out during the Musk v. Altman trial, where testimony revealed the company feared being too dependent on OpenAI (source). Recruiting a proven team of AI researchers is a natural hedge. But doing it by gutting a nonprofit that the federal government just backed with $152 million puts some weight on what "humanist" actually means in practice.
Peter Clark, one of Ai2's founding members, has stepped in as interim CEO. Researcher Iz Beltagy returned to Ai2 (source) "at this pivotal moment," and the lab is actively rebuilding. But replacing a CEO, a COO, and the lead researchers behind your flagship project is not a quick fix.
The federal government bet $152 million on the idea that open AI research could hold its own against corporate labs. Microsoft answered that bet by hiring the people, not funding them. That's the real vulnerability in open-source AI right now: the work is open, but the people doing it can always leave. A nonprofit can build the most transparent model in the world, share every training run, release every checkpoint, and still lose everything in a single hiring cycle because it can't match Big Tech compensation. If Ai2 rebuilds and the OMAI cluster delivers on its promise, this will be a speed bump. If the talent doesn't come back, it'll be the most expensive proof yet that open-source AI's biggest problem isn't compute. It's payroll.
