The artificial intelligence landscape is currently dominated by a fervent push towards "scaling," where companies invest billions in building colossal data centers to support ever-larger language models. This strategy hinges on the assumption that increasing computational power for existing AI training methods will inevitably lead to advanced, superintelligent systems capable of diverse tasks.
However, a growing number of AI researchers are beginning to question the efficacy and sustainability of this scaling-centric approach. Sara Hooker, formerly the Vice President of AI Research at Cohere and an alumna of Google Brain, is among those challenging the status quo. She has launched Adaption Labs, a new venture co-founded with fellow Cohere and Google veteran Sudip Roy. Their startup is built on the premise that the relentless scaling of large language models is becoming an inefficient route to enhancing AI performance. Hooker, who departed Cohere in August, quietly announced her new company this month as part of a broader recruitment drive, seeking to develop AI systems that can continuously learn and adapt efficiently from real-world experiences.
Adaption Labs represents a significant shift away from the prevailing "scaling-pilled" mentality within the AI industry. Hooker asserts that while attractive, this approach has failed to produce truly intelligent machines capable of navigating and interacting effectively with the real world. She emphasizes that adaptability is the core of true learning, drawing parallels to human cognition where mistakes lead to refined behavior. Current reinforcement learning (RL) methods in AI, while useful in controlled environments, do not yet enable production-level AI models to learn from real-time errors, leading to repeated failures. This highlights a critical gap that Adaption Labs seeks to fill by focusing on efficient, adaptive learning. The industry's current reliance on expensive consulting services for model fine-tuning further underscores the need for more inherently adaptive and cost-effective AI solutions. Hooker believes proving that AI systems can efficiently learn from their environment will democratize control and influence over AI development, making these powerful tools more accessible and responsive to diverse needs.
The emergence of Adaption Labs coincides with increasing skepticism regarding the long-term potential of scaling large language models. Recent research from MIT suggests that the largest AI models may soon encounter diminishing returns. Prominent AI figures, including Richard Sutton, often referred to as the "father of RL," and former OpenAI employee Andrej Karpathy, have publicly voiced their reservations about the ultimate scalability and real-world learning capabilities of current LLM and RL methods. While the AI industry has seen some advancements through AI reasoning models, these still entail significant computational costs. Adaption Labs, having reportedly closed a substantial seed funding round, aims to be at the forefront of the next AI breakthrough by demonstrating that experiential learning can be achieved far more economically. Hooker, known for her work on compact AI systems and her commitment to global AI research accessibility, envisions a future where AI's power is derived from intelligent adaptation rather than sheer size, potentially revolutionizing who benefits from and shapes artificial intelligence.
