Multimodality and World Models: AI's Path to True Understanding

Resources

Resources & Recommendations

People Mentioned

• Christian Keller (Angel Investor, Startup Advisor, Product Lead at Meta Superintelligence Lab) - The main guest of the episode, discussing his experience building AI products and models for 15 years, including his work on PyTorch, Llama 3, Llama 4, and world models.
• Jan LeCun (AI Researcher) - Referenced for his argument that language models have built-in limitations and the need to train on video or multimodal models for AGI, and his firm belief in world models.
• Stanley (AI researcher, former head of AI at Uber, then OpenAI) - Author of "Why Greatness Can't Be Planned," whose argument about open-ended systems as stepping stones to AGI is discussed.
• Kaster (CEO of Applied Intuition) - Mentioned in a previous episode for discussing the movement towards "tiger teams" or small team models in organizations.
• John Waldman (CEO of Homebase, Stanford guy) - Mentioned in a previous episode for discussing how the product development process has changed, moving from 20-page PRD docs to clickable, lovable prototypes.
• Martin Reeves (Head of BCG's Henderson Institute) - Mentioned for his research indicating that only 3% of people are capable of both execution-oriented tasks and starting new things from a blank slate.
• Phoebe (Founder of Curato AI) - Mentioned in another episode for her work on eliminating bias in data collection for AI training.

Books

• "Why Greatness Can't Be Planned" by Stanley - This book by an AI researcher discusses how to achieve AGI through open-ended systems rather than predefined goals.

Tools & Software

• PyTorch - An open-source framework used to build AI models, described as the language researchers use to formulate AI models and a translation layer between functions and hardware. ChatGPT was built leveraging PyTorch.
• ChatGPT - An AI model that changed public awareness of AI capabilities and demonstrated how well packaged and accessible this technology could be.
• Google Assistant - Mentioned as an example of narrow, single-use case assistants that pre-dated the general approach of ChatGPT.
• Alexa - Mentioned as an example of narrow, single-use case assistants that pre-dated the general approach of ChatGPT.
• Llama 3 - An AI model that Christian Keller worked on, mentioned in the context of AI research and building AI products.
• Llama 4 - An AI model that Christian Keller worked on, mentioned in the context of AI research and building AI products.
• Gemini - An AI model used by Christian Keller's wife to create French language tests.
• DeepSeek - A Chinese AI model mentioned as an alternative to OpenAI models, particularly relevant in regions like Bhutan.
• Google Scripts - Used by Henrik to automate and summarize newsletters, later refined with the help of ChatGPT.

Organizations & Institutions

• Meta Superintelligence Lab - Where Christian Keller is currently a product lead, involved in advanced AI research.
• OpenAI - The organization that developed ChatGPT, mentioned for its approach to releasing a general technology.
• BCG's Henderson Institute - A think tank mentioned for its research on leadership and organizational capabilities.

Other Resources

• World Models - A concept pushed by Jan LeCun and supported by Christian Keller, focusing on creating models that can robustly predict changes of state in the world, rather than just the next most likely word.