Dr Yichuan Zhang is co-founder and CEO of Boltzbit, a London AI research lab building generative models that keep learning after they are deployed.
On 16 September 2026 he and co-founder and CTO Dr Jinli Hu, with two co-authors, posted a working paper on arXiv, “Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data”, describing a way for a model to adapt from live interactions while its stored parameter footprint stays constant.
The paper follows the public launch of the company’s live learning approach on 30 July, and Boltzbit’s website lists TP ICAP, Liquidnet and QUOD Financial among its customers. He tells Business Matters why frozen models are a structural flaw, why sovereignty has to sit with the user and why founders must be comfortable being misunderstood.
What do you currently do at Boltzbit?
As co-founder and CEO of Boltzbit, I lead our long-term strategic plan, core research direction and commercial execution.
As a firm, we are pioneering General Learning Intelligence (GLI), an architectural shift towards what we call “AGI 2.0”. Instead of relying on static, multi-billion-dollar pre-trained models that stop learning in deployment, we build generative AI systems capable of continuously adapting their weights directly from live user interactions and production environments in real time.
My day-to-day is split between guiding our research, driving our product roadmap and product development, and working with deployment partners across financial services, data infrastructure and technology. A significant part of my focus is ensuring that our theoretical breakthroughs translate directly into reliable, high-performance infrastructure that businesses and individuals can trust with their most critical workflows.
What was the inspiration behind your business?
My background spans academic machine learning research at the University of Cambridge and the University of Edinburgh, alongside applied research. Across all these environments, one fundamental structural flaw became glaringly obvious: the industry has been treating learning as a massive, one-off event during training, rather than an ongoing part of daily use.
Modern foundation models store hundreds of billions of static parameters but, once deployed, they stop learning. They cannot learn or accumulate new skills from user interactions alone.
As a result, keeping these models relevant demands exorbitant, continuous retraining cycles that only a tiny handful of centralised tech giants can afford. This creates massive centralisation of power, vendor dependency and an unsustainable trajectory.
What does your new working paper set out?
In September we published a working paper, “Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data”, written with my co-founder Dr Jinli Hu. It describes a compact hypernetwork that turns runtime data into a low-rank modulation of a shared base network, while a Bayesian belief over the generator is updated throughout a session.
The stored parameter footprint stays constant while the effective weights are generated dynamically, which is why we describe the parameters as infinite. The aim is for a model to benefit from the facts and corrections users give it without expanding the context window or relying on retrieval alone.
Who do you admire?
I have immense respect for Professor Geoffrey Hinton. My PhD research built on his earliest work on the Boltzmann machine, one of the contributions recognised by his Nobel Prize.
He was among the first to prove that neural networks can learn within a unified framework grounded in principled Bayesian probability rather than just ad hoc heuristics. His persistence in foundational AI research continues to inspire me to hold firm to the technical direction we chose at Boltzbit from day one. Many people did not understand us at the beginning.
I also admire Yann LeCun for his willingness to challenge industry orthodoxy in public. Whether you agree with every specific viewpoint or not, tech needs leaders who resist groupthink, question prevailing assumptions and look past hype cycles.
Looking back, is there anything you would have done differently?
During the initial explosion of generative AI, the overwhelming momentum in the market pushed everyone in the same direction: fine-tuning off-the-shelf foundation models, optimising prompt engineering, expanding context windows and building retrieval-augmented generation (RAG) pipelines.
Looking back, we could have started exploring commercial applications of AI much earlier. The biggest lesson I have learnt from my years of building Boltzbit is that market demand is the true driving force behind cutting-edge technological innovation, not the technology itself.
From the start, my co-founder and I have committed our entire focus to dynamic-weight learning and continuous adaptation to make AI not only smarter but cheaper and more accessible. That was always the right call, but trusting our original scientific conviction sooner may have allowed us to bypass those industry detours entirely.
What defines your way of doing business?
Three core principles define my approach.
The first is first-principles technical rigour. There is still a wide gap in AI between what marketing promises and what products really deliver. For example, many tools currently on the market claim to adapt in real time when, in reality, they are simply querying static files while the model remains frozen. Given our research heritage, we hold ourselves to high technical standards and let our theoretical foundations speak for our commercial products.
The second is true user-level sovereignty. There is an urgent, necessary debate today around digital sovereignty and who gets to influence the trajectory of AI. This matters profoundly because allowing a tiny handful of centralised labs to hold the monopoly on how intelligence evolves is dangerous for innovation, privacy and economic resilience.
But we define these concepts differently. We believe real sovereignty must happen at the user level. So everything we build at Boltzbit aims to give users the ability to hold the pen on how their models evolve, turning AI into a durable, private asset that they genuinely own.
The third is smart infrastructure over raw scale. Expanding server farms to retrain static models is an expensive substitute for genuine architectural ingenuity. Very few companies can sustain that capital burn, and even fewer should have to. We focus on building fundamentally leaner, smarter systems and turning theoretical science into viable commercial applications.
What advice would you give to someone starting out?
Have the courage to build where the consensus is not looking, and develop the stamina to endure the quiet middle between the founding idea and commercial traction.
When you start a company, especially in deep tech, there is enormous pressure to jump on whatever narrative is currently raising money or dominating headlines. But hype cycles change constantly. If you build on market sentiment, you will pivot yourself into irrelevance.
Defensibility comes from finding a deep, structural problem that you understand better than anyone else, and staying the course even when you are out of step with the prevailing trend.
As a founder, your biggest asset is resilience. You have to be comfortable being misunderstood for a long time and resilient enough to handle the setbacks.
