Sara Hooker, former VP of AI Research at Cohere, is leaping into a new challenge. As the most visible albina Google Brain she wants to shake up the entire field of artificial intelligence. She co-founded Adaption Labs with Sudip Roy. Their vision is of AI systems that are always evolving and learning from the real world. This groundbreaking solution addresses the pain points of efficiently scaling large language models (LLMs). These models are what’s taken the AI world by storm over the past few years.
Adaption Labs has recently come into focus for its mission to increase access to AI research worldwide. Hooker aims to recruit research talent from underrepresented areas, such as Africa, thereby increasing the diversity within the field. In doing so, she wants to make sure AI models are helping the general population and not just the privileged few.
Closing the funding Adaption Labs earlier this fall successfully raised a first round of seed money in the range of $20 million to $40 million. This funding will supercharge their vision and creativity. Hooker believes the way we are going about scaling LLMs is totally backwards. He feels it doesn’t do enough to increase the common sense of the AI systems in question.
“There is a turning point now where it’s very clear that the formula of just scaling these models — scaling-pilled approaches, which are attractive but extremely boring — hasn’t produced intelligence that is able to navigate or interact with the world.” – Sara Hooker
The industry has always depended on massive models. Hundreds of businesses and developers have poured billions of dollars into building on the assumption that these bigger and bigger models will eventually deliver up general intelligence. This belief is increasingly being questioned. Richard Sutton, another Turing Award winner, calls Hooker a key player in the birth of reinforcement learning (RL). He calls her out for her important contributions to the development of state-of-the-art AI research.
In contrast to established companies like OpenAI, which requires customers to spend upwards of $10 million for consulting services on model fine-tuning, Adaption Labs seeks to democratize access to AI. What Hooker wants to highlight is that AI systems have the potential to learn from their environments more cost-effectively than existing practices do.
“We have a handful of frontier labs that determine this set of AI models that are served the same way to everyone, and they’re very expensive to adapt. And actually, I think that doesn’t need to be true anymore, and AI systems can very efficiently learn from an environment. Proving that will completely change the dynamics of who gets to control and shape AI, and really, who these models serve at the end of the day.” – Sara Hooker
Hooker’s vision isn’t about putting a different corporation in control, but redesigning the contours of who’s in control of AI development. She imagines an eventual landscape where non-traditional contributors are able to have an impact on the direction AI research and applications take. Through building user-oriented, flexible systems, Adaption Labs seeks to put more power in the hands of users and make AI more aware of individual needs.
These critiques of the generals’ big models have been coming from other industry leaders in chorus, too. Andrej Karpathy, an early employee at OpenAI, has expressed reservations about the long-term potential of RL in enhancing AI models. This growing skepticism highlights a broader discourse within the tech community regarding the best approaches to achieve meaningful advancements in artificial intelligence.
Adaption Labs is continuing to operate with a deep impetus towards effective learning from practice at-scale. This commitment drives the company’s mission as an open AI disruptor in the rapidly evolving AI landscape. The nonprofit performs services beyond their mission. Startup makes accessibility and adaptability a priority. Hearing these examples would inspire a new generation of intelligent systems designed to serve much wider and diverse audiences.