Tech
OLMo Released Every Checkpoint. Its Lead Joined Microsoft.
Hanna Hajishirzi's last arXiv submission came seven days before she joined Microsoft Superintelligence. OLMo's model flow expertise now runs inside a closed lab; AI2 has said it cannot afford frontier-scale open model work anymore.
Hanna Hajishirzi joined Mustafa Suleyman's Superintelligence team at Microsoft in late March, weeks after AI2 CEO Ali Farhadi departed the same organization.
Seven days before she left, she posted TurnWise to arXiv on March 17. The paper showed that 10,000 multi-turn training conversations lift OLMo 3 by 12 percent on multi-turn evaluation, against single-turn-only baselines.
She and Noah Smith, both University of Washington faculty, co-led OLMo's research direction; the training team at AI2, including Luca Soldaini and Dirk Groeneveld, remains in Seattle. OLMo's model flow released every training stage, dataset, and intermediate checkpoint needed to reproduce each run from scratch. The models have been downloaded more than 10 million times.
Suleyman called her "one of the most cited researchers of natural language processing in the world, full stop." Microsoft's Superintelligence unit was built to reduce the company's dependence on OpenAI for frontier models, and Hajishirzi had published exactly the training transparency that process requires.
AI2 board chair Bill Hilf told GeekWire the pullback from frontier-scale work was financial: "The cost to do extreme-scale open model research is extraordinary." He called it "really hard to do extreme-scale model work inside of a nonprofit."
Hajishirzi retains a University of Washington faculty position and co-leads a $152 million NSF and Nvidia grant with Smith. That grant covers open AI tools for scientific research in climate, conservation, and health. Frontier LLM training is not its scope.
What Moves, What Stays
The departure exposes a narrowing in who can afford to verify frontier AI openly. OLMo's model flow required human expertise to design a full training disclosure and an organization willing to cover frontier compute costs as a public good. Hajishirzi moves the scientific ambition to Redmond; Hilf has said the compute commitment is no longer viable for a nonprofit.
AI2's next major OLMo release is not yet announced. When it arrives, the model flow documentation will answer the question her departure leaves open: whether the decision to publish everything was hers, or the institution's.