# Why AI Can Fold Proteins But Can't Tell You What to Eat | Abhishek Singh | TEDxBoston Source: https://www.youtube.com/watch?v=qgo_2wtRkvM Recap page: https://rapidrecap.app/video/qgo_2wtRkvM Generated: 2025-11-24T17:45:16.561+00:00 --- ## Quick Overview Abhishek Singh argues that while AI can perform complex tasks like folding proteins or generating media, it currently lacks the fundamental, evolutionarily ingrained mechanism of language that allows humans to cooperate at scale with strangers, which is the key differentiator enabling human collective intelligence, contrasting this with highly coordinated but less flexible systems like bee swarms or wolf packs. **Key Points:** - AI can generate videos, write software, and fold proteins, demonstrating immense capability in specific domains. - The speaker contrasts human cooperation (high scale, high heterogeneity) with animal cooperation (bees: high scale, low heterogeneity; wolves: low scale, high heterogeneity). - Human ability to cooperate at massive scale with strangers stems from language, which is used for constructing shared reality, not just describing it. - The speaker notes that current AI lacks this programming for a formal, reliable language that allows for complex collaboration through shared understanding. - The future of health involves AI leveraging massive, fragmented datasets (genomics, metabolomics, behavior) through federated learning to optimize personal health. - The solution requires building a collaborative layer that allows these diverse AI agents to communicate using a unified, formal language, similar to how protocols like HTML/HTTP standardized the web. ![Screenshot at 04:58: The slide contrasts two neural networks, illustrating the concept of decentralized machine learning over fragmented data, referencing Singh's PhD thesis work on 'Exchanging dreams instead of models for federated aggregation.'](https://ss.rapidrecap.app/screens/qgo_2wtRkvM/00-04-58.png) **Context:** Abhishek Singh presents at TEDxBoston, discussing the current limitations of Artificial Intelligence compared to human collective intelligence, specifically focusing on the role of language. He uses analogies involving bees, wolves, and the history of computing (CERN's information management, Bill Gates/Paul Allen developing BASIC interpreter) to frame the discussion around cooperation, scale, and heterogeneity in complex systems. The presentation culminates in a vision for personalized health driven by AI analyzing vast amounts of individual biological data. ## Detailed Analysis Abhishek Singh begins by acknowledging the current excitement around AI's capabilities, such as generating videos, writing software, and folding proteins (0:05-0:14). He immediately pivots to a core question: despite these miracles, why can't AI answer fundamental questions about personal health like diet recommendations? He frames the problem using a scale vs. heterogeneity graph, placing bees high on scale/low on heterogeneity (high coordination, low flexibility) and wolves high on heterogeneity/low on scale (specialized roles, intimate bonds). Humans, he argues, occupy a unique space, enabling large-scale cooperation with strangers by constructing shared realities through language (2:01-2:23). He emphasizes that language is not just for describing reality but for constructing it. He contrasts this with AI, noting that while AI agents are powerful individually (like the array of heads on shelves in 2:40), they lack the ingrained, formal language necessary for reliable, large-scale collaboration (2:51-3:38). He draws a parallel to the early web protocols (HTML/HTTP) that unified disparate systems, suggesting AI needs a similar unified language for its specialized agents (health, sleep, genetics analysis) to collaborate effectively (4:39-5:25). He concludes by noting that if we provide this unified, programmable foundation, we can leverage massive health datasets to optimize individual lives, ultimately enabling powerful self-organization and improvement in complex systems like human health (5:55-8:55). ### AI Capabilities vs. Human Cooperation - AI generates videos, writes software, folds proteins - Humans achieve massive-scale cooperation with strangers through language - Bees exhibit high-scale but low-flexibility cooperation - Wolves exhibit high-heterogeneity but low-scale cooperation. ### The Role of Language - Language is key to human cooperation, enabling the construction of shared reality, unlike animal instincts or earlier technologies like the internet protocols (HTML/HTTP). ### The AI Collaboration Problem - Current AI agents, though powerful individually (illustrated by data analysis tools like genomics, metabolomics), lack a unified, formal language to collaborate reliably across different data types and domains. ### The Solution - Create a unified, programmable language/interface (like CUDA for GPUs) that allows specialized AI agents (nutrition, sleep, DNA analysis) to build upon each other's insights and hypotheses, enabling collective intelligence at scale. ### Application to Health - Future personalized health relies on AI analyzing massive, fragmented datasets (terabytes daily) from wearables and tests to provide optimized, evidence-based recommendations for individuals. ![Screenshot at 0:04: Speaker Abhishek Singh introduces the talk at TEDxBoston.](https://ss.rapidrecap.app/screens/qgo_2wtRkvM/00-00-04.png) ![Screenshot at 0:52: Visual comparison showing bees existing at high scale/low heterogeneity and wolves at high heterogeneity/low scale on a 2D graph.](https://ss.rapidrecap.app/screens/qgo_2wtRkvM/00-00-52.png) ![Screenshot at 2:01: Slide illustrating humans bridging the gap, capable of high scale and high heterogeneity cooperation using language to construct reality, contrasting with bees and wolves.](https://ss.rapidrecap.app/screens/qgo_2wtRkvM/00-02-01.png) ![Screenshot at 2:40: Visual metaphor showing numerous identical heads on shelves, representing powerful but uncoordinated AI agents lacking a unified language.](https://ss.rapidrecap.app/screens/qgo_2wtRkvM/00-02-40.png) ![Screenshot at 4:58: Slide referencing Singh's PhD work on decentralized machine learning and federated aggregation over heterogeneous models, illustrating two connected neural networks.](https://ss.rapidrecap.app/screens/qgo_2wtRkvM/00-04-58.png)