Sara Hooker has spent the last few years doing her “work tour” at some of the most competitive labs in the AI industry. Now, he’s striking out on his own—and taking a rebellious stance against entrenched industry norms. Hooker and his co-founder, Sudip Roy, have raised $50 million for their startup, Adaptation Labs. The company is focused on the principle that effective, self-learning training methods—not large amounts of data and power—will lead to the best models.
“A lot of AI progress is: you build the biggest model, and you send the same model to billions of people around the world; it doesn’t matter what language, what industry, what business,” Hooker told the Observer. “This is a big departure from that.”
His ambitious bet caught the attention of Silicon Valley. The first round based in San Francisco was led by Emergence Capital Partners, with participation from Mozilla Ventures, Fifty Years, Threshold Ventures, Alpha Intelligence Capital, e14 Fund and Neo. Adaptation Labs declined to share details of its measurements.
Conventional wisdom in AI suggests that more computation leads to larger and more efficient models. But Adaptation Labs believes this approach has hit a wall. “One of our core beliefs is that no frontier AI lab is going to quadruple the size of its model next year,” Hooker said. “That means all bets are off.”
Instead, Hooker is interested in exploring architectural models that can learn continuously and adapt to workloads in real time as they interact with different environments. User feedback, for example, should quickly change the behavior of AI tools instead of being “lost in the dark.” The emphasis on efficiency in Adaptation Labs will also include testing ideas like “gradient-free learning,” which looks at alternatives to traditional training methods that rely on optimization algorithms to adjust model parameters in an effort to reduce errors.
Hooker and Roy are ready to disrupt the industry they have spent their careers developing. Both worked for years at Google before moving on to Cohere, where Hooker served as vice president of research, and Roy served as senior director of inference. Despite leaving Cohere last year, Hooker said he is “proud” of the work he has done, which includes a focus on the development of a multilingual model. But Adaptation Labs’ goals of enabling dynamic data, intelligence and collaboration can’t be realized within a traditional border lab, he said, because those areas are separated by different teams. “It’s very easy to start with that urgency to put everyone on the same page from the beginning,” he said.
Adaptation Labs isn’t alone in going against the grain. A growing chorus of voices in Silicon Valley has begun to question the industry’s leading assumptions. Yann LeCun, who recently left Meta to launch AMI Labs, has raised doubts about traditional scaling rules; so is David Silver, a former Google DeepMind researcher whose startup, Ineffable Intelligence, focuses on training self-learning models by using experience rather than feeding them data.
Hooker’s new funding is earmarked for team building. Adaptation Labs is currently hiring for 10 roles, some of which are based in global locations such as Turkey, Mexico and Brazil. Employees are also given a unique benefit: an “Adaptive Passport” that allows them to take an annual trip to a country they have never visited before. “We want to encourage people to not only explore, but we want to represent that we are a global technology company from day one,” Hooker said. Roy was the first to use the proceeds, using it for a trip to Costa Rica last year.
Hooker expects that the general laws of scaling will change as major AI developers face the reality that ever-increasing computing power brings diminishing returns. Algorithmic innovation will be the real driver of progress. “This is a year that will be very important,” he said.
