AI researcher and industry insider Ali Farhadi doesn’t mince words when it comes to open-source artificial intelligence: “Openness is the only way forward,” he says.
Farhadi is a professor at the University of Washington as well as the CEO of the Allen Institute for AI (Ai2), one of the most active nonprofit AI research institutes in the nation. The institute has three main areas of focus: Open Ecosystem, empowering others with AI capabilities; AI for scientists, helping them do their work by mitigating the information overload in the science industries; and AI for the Planet, providing tools to protect the environment.
Ai2 has become a leading voice in the industry’s fierce debate over how open new AI technology should be. The organization has shaped groundbreaking projects such as foundational AI research (including the first iteration of a language model, eLMo), and Ai2 spin-off startup Xnor.ai (which was eventually bought by Apple).
Farhadi was a third year assistant professor at UW when Ai2’s founder, Paul Allen, called and asked him to join. Despite being advised to focus on securing tenure at the time, he embraced the role.
“It was such a unique opportunity that I had, I didn’t need to think twice. So I jumped on it, and it’s been sort of a game changer in my career,” Farhadi says. “In retrospect, this was the best decision I’ve ever made, not only to help my tenure and my research, but also I’ve learned so much doing this that defined my career.”
Farhadi joined Ai2 in 2015, and five years later the organization sold Xnor.ai (which he cofounded), which specialized in low-power, edge-based artificial intelligence tools, to Apple. The Xnor.ai deal was around $200 million, and Farhadi actually left to join Apple for a stint.
He came back as CEO of the Allen Institute in 2023 to advocate for an open approach to AI. “It was an easy decision, and felt like going back home, but also the timing was sort of critical, given how AI was evolving,” Farhadi says.
Under Farhadi’s leadership, AI2 launched OLMo, a series of open language models. Unlike other proprietary models, which often operate behind limited APIs, OLMo offers complete openness by releasing full model weights, training code, training logs, training metrics in the form of Weights & Biases logs, and inference code.
Every model has publicly accessible code, weights, and 500-plus intermediate checkpoints, supported by tools to trace back to the exact data that was used at that point during training. Underpinning OLMo’s success is the Dolma data set. Publicly available on Hugging Face, it contains a mix of web content, academic publications, code, books, and encyclopedic materials.
On November 26, Ai2 released the latest version of OLMo, OLMo 2, which it calls “the best fully-open models to-date.” OLMo 2 is a family of 7B and 13B models, trained on up to 5 trillion tokens.
“To me, open source means I should be able to grab what you’ve developed, understand it, change it, and build upon it,” Farhadi says. “Without that, we’re crippling the potential of researchers and developers.”
Farhadi says he is acutely aware of the risks posed by AI, but he insists secrecy is not the solution. “Having these black box solutions where people don’t know much about it would only increase the chances of unintended damage,” he adds. “The biggest problem with AI today is that we don’t fully understand these systems. Opening them up allows millions of practitioners to participate in solving those gaps.”
This story is part of AI 20, our monthlong series of profiles spotlighting the most interesting technologists, entrepreneurs, corporate leaders, and creative thinkers shaping the world of artificial intelligence.
