AI Can Learn Your Business. A Cambridge Researcher Says Stop Renting It

Most firms are renting their AI. A Copilot licence here, a ChatGPT Enterprise seat there. Dr Yichuan Zhang, CEO and co-founder of Boltzbit, reckons that's a bigger risk than most of us have priced in.

🎧 Short on time? Hit play: Watch the full episode on YouTube

He's not someone who wandered into AI last year. He did his PhD in machine learning at Edinburgh, under a supervisor who'd worked with Geoffrey Hinton, then a postdoc at Cambridge, and he founded Boltzbit in 2020, before ChatGPT. So when I put it to him, "what's actually wrong with renting?", I expected a careful answer. I didn't get one.

His argument is simple. There are two ways to get intelligence into a business. You hire someone, or you rent it from OpenAI or Anthropic. Hire someone and you have an agreement, and their salary is fixed. Rent it and the price, the terms, and what happens if sanctions or politics get in the way all sit with somebody else.

That's a fair point, and a slightly uncomfortable one if your whole AI plan is one big licence.

I'll be honest, my first thought was that this sounds expensive. Yoshi pushed back harder: Microsoft runs it for you, it plugs into everything, and most firms don't have the people to build anything. Also fair.

Yichuan's answer is that there's a middle ground. Don't train a model from scratch.

But don't be a pure wrapper either, one price rise away from trouble. Build your own harness around open-source models, keep the option to switch, and let the system learn from your work. That's what Boltzbit does itself. It fuses open-source models and keeps fine-tuning them, and it mostly sticks to models of around 40 billion parameters or fewer, because they're cheap enough to host and, tuned well, very capable. (He gets into all of this from 36:00 in the episode.)

His analogy for it has stuck with me. We don't all drive sports cars, and we don't all take the bus. Different jobs, different vehicles. Most of what we ask AI to do doesn't need the biggest model on the market.

He told us about one of their financial services clients. Their staff spend hours pulling entity names out of emails: banks, bond buyers, all with short forms, typos and lines that don't match up. Boltzbit asked them to write the rules down so the AI could just be prompted. They couldn't. As the client put it:

That's most of the work I know in our industry. The knowledge that matters sits with people who've seen a thousand variations of the same problem, not in a prompt. If AI is going to help with that kind of work, it has to learn from your context. And your context is exactly what a rented model doesn't have.

His prediction for the next couple of years is that training models gets cheap, close to a commodity. What separates firms after that is the data they own. Construction firms are sitting on decades of it. Most of it is a mess, but it's ours.

👉 Watch the full conversation with Dr Yichuan Zhang on YouTube

Yoshi asked him how far AI could take a firm that wanted to build its own model. Very far, he said. About as easy as building a website on Wix. But there's a line:

Which is roughly where I keep landing on this podcast. The tools get easier every month. Someone still has to sign it off, and own it when it goes wrong.

We also get into Hinton and Boltzmann machines, AlphaGo's famous move 37, why he thinks intelligence should be measured by how fast you learn rather than what you already know, and Boltzbit's Infinite-Parameter LLMs paper, which is about models that keep learning after they're deployed.

Why it matters: If your AI strategy is one big licence, your costs, your terms and a lot of your know-how sit with one supplier. You don't need to build a model. You do need a plan for owning the bits that make your firm different: your data, your feedback, and your decisions.

Not subscribed yet? Subscribe to the Project Flux newsletter for the week's AI and tech news for project delivery, every week.

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.