The Decoder profiles 27 neolabs betting on new AI research paths

The report categorizes 27 of these labs into six groups based on their primary technical bets, with a focus on overcoming limitations in continual learning, data efficiency, and learning from experience. For example, Adaption and Core Automation aim to build models that update their weights during deployment, while Ineffable Intelligence seeks to learn from environmental rewards rather than human examples. Flapping Airplanes focuses on learning from scarce data in fields like robotics.

Financially, these labs are attracting massive capital despite having few public products. Nvidia reportedly invested $5 billion in Safe Superintelligence, and Jeff Bezos’ Prometheus raised $12 billion at a $41 billion valuation. The article notes that while big labs have more compute and data, neolabs may succeed by filling research gaps that established corporations ignore due to internal competition or the innovator’s dilemma. Some neolabs, like TypeSafe AI, have already seen their innovations quickly adopted by giants like OpenAI, suggesting their value may lie in proving expertise or serving as acquisition targets.