What’s next for Ai2: A conversation with Interim CEO Peter Clark
Allen Institute (AI2)
Ai2's interim CEO Peter Clark just outlined where the nonprofit AI lab is headed next, from open models to AI for science. His bet: staying open and slow-and-steady beats the industry's closed, breakneck sprint.
Based on reporting by Allen Institute (AI2) — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Peter Clark has been at Ai2 long enough to have done this job before, and he's stepping back into the interim CEO seat at a moment when the AI industry is moving faster than almost anyone can track. His pitch isn't about catching up to that speed. It's about resisting the urge to let it swallow the longer, harder work that doesn't pay off in a quarter or even a year.
That's the same bet Paul Allen made when he founded the institute: push the science of AI forward, make it useful, and do it in public. Clark points to ELMo, the 2018 project that helped kick off the whole large-language-model wave, as proof that Ai2 was early to this game long before anyone was calling it a revolution. The lineage runs straight through to Olmo and Molmo today, plus a newer training method called FlexOlmo, all released with weights and training details out in the open rather than locked behind an API.
The institute isn't just building models to prove a point, though. AutoDiscovery, a system for generating and testing scientific hypotheses, is reportedly already shaping how some oncologists approach cancer treatment options, and it's since graduated from research prototype to a managed tool where scientists can upload their own datasets and inspect the statistics behind every result. OlmoEarth is doing similar work on environmental and climate data. Clark frames this pairing — deep, patient research feeding into things people can actually use — as the core of how Ai2 operates, and as something a nonprofit structure makes possible in a way a venture-funded lab might struggle to justify.
Looking forward, Clark names four bets: understanding model behavior and reliability through the NSF-and-NVIDIA-backed OMAI project, an agentic research platform called Asta (with tools like ScholarQA and Theorizer built to comb literature and float new explanations), embodied AI through early efforts like MolmoAct and MolmoBot, and AI applied to planetary-scale problems like conservation and climate systems. None of it, he insists, is meant to be flashy. It's meant to be legible — the kind of work outside researchers can actually pick apart, replicate, and extend, rather than just marvel at from a distance.
What Clark is really describing is a hedge against the current mood in AI, where the biggest labs increasingly treat their best work as trade secrets. Ai2's argument is that opening the hood doesn't just serve some abstract ideal of transparency — it's actually how the field gets smarter, faster, because more people can poke at the same system instead of guessing at what's inside a black box.
My take — AI-written commentary, not fact-checked reporting
I'll say the obvious thing nobody wants to say: most of the industry's 'open' announcements are marketing dressed up as generosity, and Ai2 is one of the few outfits that actually ships weights, data, and training recipes instead of a blog post and a benchmark chart. A nonprofit with NSF and NVIDIA money behind it patiently building OMAI while everyone else races to the next closed frontier model isn't a consolation prize — it might be the only sane long-term strategy left standing once the current funding frenzy cools off.
Read more about this at: Allen Institute (AI2)