Recently, renowned AI figure François Chollet and Zapier co-founder Mike Knoop have jointly established the Ndea Research Laboratory. This lab focuses on a distinctive mission: to develop artificial general intelligence (AGI) capable of genuine invention and innovation.
Ndea's primary research direction involves integrating program synthesis with deep learning, aiming to create more adaptable AI systems. The lab intends to accelerate scientific progress through this innovative approach, targeting the development of AI technologies that can innovate independently. To achieve this goal, Ndea is assembling a distributed team dedicated to researching program synthesis.
Chollet brings extensive experience in AI, having architected the Keras deep learning library, which has provided robust support for AI development at companies like YouTube, Waymo, and Spotify. During his nearly decade-long tenure at Google, Chollet created the ARC dataset, a benchmark for measuring AGI advancements that significantly impacted how AI researchers assess their work.
The establishment of Ndea coincides with Chollet and Knoop transforming the $1 million ARC prize competition into a formal non-profit organization. Insights from this competition seem to influence Ndea's research methods. Despite high scores achieved by companies like OpenAI in ARC-AGI benchmarks, they still require substantial computational power and fall short of human-level performance in tasks easily mastered by children.
Ndea's foundation is rooted in a deep understanding of current AI limitations. While modern language models excel in pattern matching and specific tasks, they struggle with solving open-ended problems and efficient learning. Chollet stated on social media that although today's deep learning AI is impressive and economically valuable, it is constrained by inefficiencies in learning and adaptation. Ndea aims to build AI that can truly invent, adapt, and innovate through alternative approaches.
The technical strategy at Ndea revolves around program synthesis, an approach involving searching discrete programs to interpret observational data, potentially requiring fewer examples than traditional deep learning methods. By combining program synthesis with deep learning techniques, Ndea hopes to create systems that can both recognize patterns and perform formal reasoning. The lab's name, inspired by the Greek concepts of intuitive understanding (ennoia) and logical reasoning (dianoia), reflects its dual AI research approach.
Ndea's creation also highlights a trend in the AI field where prominent researchers leave large tech companies to establish independent labs. Chollet’s move aligns with others, such as former OpenAI chief scientist Ilya Sutskever, who are exploring alternative paths for AI development. Knoop will step away from daily operations at Zapier to fully commit to Ndea, bringing expertise in managing global distributed teams and AI-driven automation.
While program synthesis holds significant potential, it faces considerable computational challenges. Ndea believes that guiding the program search process with deep learning can overcome these hurdles. The lab compares the current state of program synthesis to deep learning in 2012—immature yet showing great promise, with major AI labs beginning to explore its applications.
Beyond AI development, Ndea plans to apply its technology across various scientific fields, positioning itself as a “factory for rapid scientific advancement.”
