A look at the Allen Institute for AI's plan to build “open” generative AI model OLMo; the nonprofit research center is led by former Apple ML leader Ali Farhadi
The nonprofit Allen Institute for AI, led by a respected computer scientist who sold his company to Apple, is trying to democratize cutting-edge research. X: @stevelohr , @vore_jm , @dataengines , @allen_ai , @seangraf , @billhilf , @stevelohr , @uwcse , @metaviv , and @carysafoko LinkedIn: Rachel Shepheard , Matt Fleischl , Steve Lohr , Ali Farhadi , and Ashley Boyd X: Steve Lohr / @stevelohr : A leader in AI research advocates “radical openness.” “We're pushing for a glass box. Open up the whole thing, and then we can talk about the behavior and explain partly what's happening inside.” - Ali Farhadi, ceo Allen Institute for AI @allen_ai https://www.nytimes.com/... JM Vore / @vore_jm : 1/3 The issues here show a general unwillingness and lack of discipline in AI research as far as who else is involved. Because of the incestuous nature of AI research, we rarely see an actual AI Census https://www.nytimes.com/...? @dataengines : “The task for society, he added, is to better understand and manage the risks. Openness, he contends, is the best bet to find safety and share economic opportunity.” So what do we do to stay safe from closed systems or even those open ones in the field? https://www.nytimes.com/... @allen_ai : AI2's CEO Ali Farhadi believes that advancements in AI demand “radical openness.” In a conversation with @nytimes' @SteveLohr, Ali pushes for a glass box approach, allowing researchers to better understand AI behavior: https://www.nytimes.com/... Sean Graf / @seangraf : The leading chatbots makers — Microsoft-backed OpenAI and Google — have kept their newer technology closed, not revealing how their A.I. models are trained and tuned. An Industry Insider Drives an Open Alternative to Big Tech's A.I. https://www.nytimes.com/... Bill Hilf / @billhilf : There are a lot of perspectives on the power of generative AI and it's great to see @allen_ai and Ali Farhadi leading the charge on the necessity of open models and data sets; creating a glass box instead of a black box. https://www.nytimes.com/... [image] Steve Lohr / @stevelohr : “A tiny set of players, all on the West Coast of the U.S., is trying to lock down the generative A.I. space even before it really gets out the gate.” - Mark Surman, president of the Mozilla Foundation. https://www.nytimes.com/... @uwcse : In contrast to black-box commercial #AI models like #ChatGPT, #UWAllen @RAIVNLab & @allen_ai's Ali Farhadi advocates for a glass box: “Open up the whole thing, and then we can talk about the behavior and explain partly what's happening inside.” #AIforGood https://www.nytimes.com/... Aviv Ovadya / @metaviv : I was recently quoted in a @nytimes piece on the importance of the “open source” AI question ( https://www.nytimes.com/... a short quote cannot capture a very complex set of concerns and tradeoffs. First, several things can be true simultaneously: 1️⃣ Open access to AI systems... Carys Afoko / @carysafoko : As generative AI has taken off we've seen a small number of powerful actors try and lock others out of the industry. Read @MSurman in today's @nytimes to learn what @mozilla is doing to build the alternative. https://www.nytimes.com/... LinkedIn: Rachel Shepheard : I don't tend to post much here but LaunchSquad, my professional home, has been churning out such extraordinary results for our clients lately that I felt compelled to break the drought. … Matt Fleischl : Everyday I read about a new AI model or chatbot or company applying AI to X, but client Allen Institute for AI (AI2) is focused on something much more important: transparency and open access. … Steve Lohr : A leader in AI research advocates “radical openness” in generative AI. — “We're pushing for a glass box. … Ali Farhadi : Grateful for a great conversation with Steve Lohr at the The New York Times. This is a transformative time to be working in artificial intelligence. … Ashley Boyd : “It takes an ecosystem of open players to really make a dent in the big players,” said Mark Surman of the Mozilla Foundation. …
Context & Ripple Effects
AI2’s “glass box” stance makes model transparency—not just model access—the point of differentiation. It establishes an institutional alternative to the more concentrated paths later contested in the fight over whether AI remains concentrated or broadly accessible.
The approach also foreshadows a durable fault line: Meta would later treat open sourcing as a way to put price pressure on competing AI offerings, while OpenAI was still describing weight and research release as not its highest current priority.
First-order effects
- AI2 commits its research agenda and leadership position to OLMo as an open generative-model effort, making reproducibility and inspection central to its proposition.
- Researchers and developers gain a prospective noncommercial reference point for studying model behavior, rather than relying solely on externally controlled systems.
Second-order effects
- The plan raises the competitive value of publishing weights, data details, and training methods: closed-model providers must more clearly justify what users cannot inspect.
- Open models can shift differentiation toward the services around them—compute, deployment, tuning, and safeguards—rather than access to a model alone.
Third-order effects
- If well-resourced open efforts persist, frontier AI may develop as a mixed ecosystem of proprietary products and shared research infrastructure, not a single access model.
- That split makes governance more complex: transparency can aid independent scrutiny, while wider model availability keeps the access-versus-control debate active.
The trend: OLMo is an early data point in the push to treat transparent, reusable AI models as commons-like research infrastructure alongside closed commercial systems.