Stop Renting Your Company’s Intelligence with Dr Yichuan Zhang, Boltzbit
Send us Fan Mail Summary Project Flux hosts James and Yoshi speak with Yichuan Zhang, CEO and co-founder of Boltsbit, about general learning intelligence and his vision for AGI 2.0. Zhang explains why intelligence should be measured by the ability to learn, how contextual data and ongoing feedback could enable more adaptive AI, and how Boltsbit combines task modes with continuously learning models. Key Topics Yichuan Zhang traces his path from software engineering and AI research in Edinburgh and Cambridge to Google and founding Boltsbit in 2020. He connects Boltzmann machines and the unification of learning and thinking to later deep learning breakthroughs, including AlexNet and generative language models. Zhang distinguishes three learning capabilities: memory and retrieval, learning patterns from examples, and creating from instructions; he argues today’s AI still has room to improve at learning from examples. He describes AlphaGo Zero as evidence that AI can produce discoveries beyond its training examples when supported by sufficient compute and data. His Infinity Parameter LMs paper explores dynamically changing model parameters and generating new experts as a model produces tokens; he frames it as an early step toward general learning intelligence. Zhang argues data is use-case-specific and contextual, and that intelligence should be judged by the speed and capability of learning rather than stored knowledge or current skills. Boltsbit offers general chat, widget creation, document work, and coding modes, with a learning console for users to guide what the agent learns and how it responds. He describes Boltsbit’s approach as combining open-source models and applying continuous fine-tuning, while favoring models of roughly 40 billion parameters or fewer for practical hosting. Zhang argues organizations should build control and adaptability around AI rather than rely entirely on rented intelligence, and predicts model training will become cheaper while valuable domain data remains differentiating. On the future of AI, Zhang says people will still supply judgment and approval, even as AI makes building and deploying models much more accessible. Chapters 00:00 Meet Yichuan Zhang: from AI research to Boltsbit 03:46 Hinton, Boltzmann machines, and deep learning’s rise 08:05 Three kinds of learning and what AI still lacks 11:01 AlphaGo Zero and AI’s capacity for novel discoveries 12:46 Dynamic experts and the Infinity Parameter LMs paper 16:54 AGI 2.0: measuring intelligence by learning speed 18:12 Why data needs context, not just scale 22:04 General learning intelligence and human experience 26:23 Why world models do not solve contextual learning 28:46 Boltsbit’s four modes and continuous learning 32:15 Feedback, enterprise workflows, and model choices 36:00 Small models, AI ownership, and the cost of renting 40:07 What may change by 2030, and where human judgment stays 44:21 Open-source models, training, and the closing conversation Links & Stuff 🔗 Yichuan on LinkedIn: www.linkedin.comhttps://www.linkedin.com/in/dr-yichuan-zhang-bz1886 🏢 Boltzbit:https://boltzbit.com/ www.linkedin.com/company/boltzbit Support the show Subscribe to the Project Flux Newsletter! All our podcasts commentate on the weeks hottest news brought to you by the Project Flux Newsletter! Get a step ahead on the world of AI and tech, subscribe and get the weekly newsletter to your inbox every Monday! Sign Up to Newsletter Project Flux is independently owned, funded, and operated by its founder. Any external references or features are for community benefit only and do not imply endorsement, control, or ownership.
