← All Heroes HEROES BEHIND AI

Yichuan Zhang: The Sage Behind Boltzbit

A founder archetype analysis of Yichuan Zhang, CEO and Co-Founder of Boltzbit, the London AI research company building models that learn from each user. Who he is, how he leads, and what drives him.

Oct 1, 2026

CEO and Co-Founder at Boltzbit

Archetype The Sage

Classical Hero Prometheus

Pop Culture Hero Yoda

Journey Stage The Reward

Listen on TL;Listen Prefer to listen? Hear this article read aloud.

Who Is Yichuan Zhang?

Yichuan Zhang has been an AI researcher for almost twenty years. He worked at Google before the ChatGPT era, back when, as he puts it, machine learning was "all about how to optimize and how to get the right data." In 2020 he co-founded Boltzbit in London with a fellow researcher he first met around his PhD years. Today the company has fifteen people, half research and backend engineers, half business and marketing.

The name is the first clue to how he thinks. Boltzbit comes from the Boltzmann machine, the early neural network Geoffrey Hinton introduced in the 1980s, combined with "bit" from computer science. Yichuan named his company after a learning algorithm, and learning is the whole story.

His argument is simple to state and hard to unsee once you hear it. The AI industry measures intelligence by what a model already knows. Yichuan thinks the real measure is how fast it can learn something new, and that today's frozen, pre-trained models have quietly abandoned the idea that made machine learning work in the first place.

The Archetype: The Sage

Primary

The Sage

Secondary

The Rebel

Journey Stage

The Reward

Yichuan leads with the archetype of The Sage. He answers questions by going back to definitions. Asked about General Learning Intelligence, he started with what intelligence means, and the parent who brags that their three-year-old already knows a thousand words. The point is not the thousand words. It is how fast the child absorbed them. Asked why skills files and prompts are not enough, he asked what learning is, and separated it from memorization.

He describes his field the way a philosopher would. Learning, in his framing, is about finding patterns and compressing them. Piling more rules into a prompt is the opposite. "You are not compressing, you are not finding the patterns, you are just piling up things."

His secondary archetype is The Rebel, a quiet one. He disagrees with the industry's default that a handful of providers should own the weights and everyone else should rent them. He calls the open-source model a binary: you can run it, but the real source code is the weights, and those are "controlled by the model provider." His AGI 2.0 vision moves AI from centralized to distributed, the way computing moved from the mainframe to the PC.

"AI is not about engineering. AI is a very philosophical thing."

The Hero Match

Classical Hero

Prometheus

In Greek myth, Prometheus brings fire from the gods to people. In Aeschylus' version of the story, the gift is bigger than fire. He gives humans the ability to learn crafts for themselves: numbers, letters, the arts. His name means forethought.

That is the parallel with Yichuan. The capability he cares about is the ability to train intelligence, and today it sits inside a few labs. Boltzbit's goal, in his words, is "giving that real source code to everyone so everyone can start developing their own model." He sees machine learning as the most democratic kind of software, because in principle all you need is the data. And he is already thinking past that, to a new intellectual property system where your data stays with you and anyone who wants to use what you know has to ask your AI and pay for it. Forethought, applied to who benefits once everyone can train.

Pop Culture Hero

Yoda: The Empire Strikes Back

The Dagobah version. Small, easy to underestimate, and his whole lesson to Luke is that size is not where power comes from. Yichuan makes the same case about models. Boltzbit's model is much smaller than the frontier giants, built to learn one user's decisions, habits, and intentions. ChatGPT, in his analogy, is the bus: fast in its own lane, useless if you want to go somewhere it doesn't run. Like Yoda, he also believes understanding comes from experience rather than instruction.

"It has to be by example, right? And that's the real scalable way."

The Story Behind Boltzbit

In 2020, Boltzbit was already calling itself a generative AI company. Yichuan and his team took a GPT-based language model to machine learning engineers working on language tasks. The answer was no. They had specialized tools for everything; why would they need general intelligence? Then ChatGPT arrived, and the market, in his words, "opens like after one night." The question changed from whether anyone needed general intelligence to where exactly it fits to create real value.

Boltzbit started in financial services, picked after looking at healthcare, logistics, and marketing, because finance runs on enormous amounts of text. Those clients wanted solutions delivered fast, so in the second half of 2025 the team built an internal coding agent to ship faster, running alongside Claude and OpenAI's APIs. It was never meant to be a product. Then, early this year, friends who were watching their AI coding bills explode asked if there was something cheaper and more interesting. Yichuan said they had something internal they could try. They liked it, and it spread by word of mouth.

The product that spread does something the big assistants don't. When you tell it you don't like a UI, or to always use your design tokens, it treats that as knowledge, fine-tunes it into its own weights, and remembers it from then on without you paying again in prompt tokens. When you ask for something it has never learned, it asks a frontier model for help.

That arc, from rejected-too-early to found-by-accident, is the part of a story where the hero earns the reward. The test years are behind Boltzbit. The B2B product is mature, a consumer version is about to launch, and Yichuan is already describing the next chapter.

The Founder's Journey ↔ The Company's Journey

Yichuan Zhang's Arc

Two decades of learning research. Machine learning at Google when learning was still the point. Watched the field trade learning for pre-training and found it "very sad." Co-founded a company named after a learning algorithm. Took his own team fully AI-coded and came out with a rule about where humans still belong.

Boltzbit's Arc

A 2020 generative AI startup selling to engineers who said no. A financial-services focus chosen before ChatGPT, because finance runs on text. A market that opened overnight once ChatGPT arrived. An internal coding agent that friends asked to borrow. A horizontal, infrastructure-focused product that learns each user. A consumer version about to launch, and a vision for an intellectual property system built on personal AI.

The same archetype drives both. A Sage who believes intelligence means learning built a company whose product keeps learning, and whose own direction keeps adapting while the principle holds still.

How Yichuan Leads

Yichuan leads from conviction. His standing message to his team is "the existence of startup is because we move faster than everyone else," and he applies it to himself first. Early this year Boltzbit restructured, wound down its front-end team, and moved to building every product with AI. "There's no single line of human code."

What makes that move credible is the rule he paired it with: "AI can automate things, but AI cannot take responsibility." Humans decide when a mistake would be expensive, like choosing the architecture of a backend service. When the stakes are a button animation, AI can just do it. He measures engineers the same way, against business goals: new services shipped, users reached, stability and scalability, data produced for the models. "But never line of code."

With his co-founder of nearly twenty years, he describes partnership, not hierarchy: two researchers who work naturally alongside each other, and a relationship built on being "supporting and being tolerant to each other."

Founder Superpowers

Superpower

Redefining the benchmark

Yichuan changes the scoreboard. Intelligence is not what a model knows; "the new benchmark will be the speed of learn rather than what it can learn or what it can do." Open-source models are not really open, because the weights are the source code. Skills files are memorization, not learning. Each reframe makes the conventional measure look like the wrong one.

Superpower

Making abstractions physical

He turns learning theory into pictures. ChatGPT is the bus; your own model is the car you take to the lake. Nobody has ever seen the UI of the internet. Try writing a skills file that teaches a model to speak Chinese. A twenty-year researcher who explains like this is rare.

Superpower

Drawing the accountability line

He took his own company through the AI shift and came out with a rule a team can run on: AI automates, humans stay accountable where mistakes are costly. It shapes who decides architecture, what gets reviewed, and how engineering performance is measured.

What It's Like to Work with Yichuan

In conversation, Yichuan is calm, patient, and precise about ideas. He lets a question land, answers in one complete arc, and checks in with "right?" as he goes, as if working the idea out with you. He is warm toward the people he talks with and firm about what he believes. He will tell you plainly when he thinks the industry is wrong, and just as plainly when he doesn't know something yet.

Expect a research culture with a bias for speed. Half the team is research and backend engineering, and he describes a "very high bar of engineering hiring." Expect to be measured on what your work does for users and the business, not on output volume. And expect a founder who changes direction readily, without changing what he believes.

"I don't believe there's a dead end, right?"

Why This Matters (For You)

If You're Running AI on Work Full of Exceptions

If your team's workflows look simple on paper and turn out to be "exception, exception, exception" in practice, Yichuan's argument is aimed at you. His example is a financial analyst extracting data from reports across twenty sources with different formats. Nobody will sit down for two days to write every exception into a prompt, and you could never list them all. "That's the biggest barrier for people to to adopt AI." His answer is a model that learns from your corrections as you work, instead of a growing pile of instructions. The question to ask of your current setup: is it learning, or just remembering what you wrote down?

If You're an Engineer Building on Foundation Models

Yichuan's split is worth stealing: general knowledge should be pre-trained once, efficiently, by someone else; personal and task-specific behavior should be learned on the fly. His product routes anything it hasn't learned to a frontier model, so the small learning model and the large general one collaborate rather than compete. He also offers a clean rule for shared models: train one team model when the work is scripted, like customer service, but never average together people with personal style, like traders. Where in your system are you prompting around something that should actually be learned?

If You're Early in Your Career

Yichuan's advice for the AI era is to work with the tools rather than wait for school to catch up, because "our education system cannot catch up." But the deeper lesson in his story is about timing and patience. He was building generative AI before anyone wanted it, kept the core idea, and changed direction as the market moved. He also pushes back on the idea that AI makes human thinking obsolete. Learn to be human, he says, not to behave like a machine. When you test your own big idea, his method is to simplify it down to "absolute brutal facts" and challenge it yourself.

If You're Considering Joining Boltzbit

Boltzbit is a fifteen-person team in London, roughly half research and backend engineers and half business and marketing. The founders are researchers who have known each other for nearly twenty years. The company builds every product with AI, sets a high bar for engineering hires, and measures engineers on business outcomes rather than lines of code. Humans own the high-stakes calls. If you want to work on continual learning and model ownership with a founder who thinks from first principles, read the "How Yichuan Leads" section above.

Go Deeper

The full conversation with Yichuan Zhang is on its way. Check out other episodes in the meantime.

Visit the Channel

Frequently Asked Questions

What is Yichuan Zhang's leadership style?

Yichuan Zhang leads from conviction with a strong bias for speed. He tells his team a startup exists because it moves faster than everyone else, and Boltzbit now builds its products entirely with AI. He pairs that with a clear rule: AI can automate, but humans stay accountable for high-stakes decisions like system architecture. Engineers are measured against business outcomes, never lines of code. He describes his co-founder relationship as supportive and tolerant.

What is Yichuan Zhang known for in AI?

Yichuan Zhang is an AI researcher of almost twenty years and the CEO and co-founder of Boltzbit, a London AI company founded in 2020. He worked at Google before the ChatGPT era. He is known for arguing that intelligence should be measured by how fast an AI learns, not what it already knows, and for building models that keep learning from each user's feedback.

What does Boltzbit do?

Boltzbit builds AI models that learn from their users. Its coding product works through an interface similar to Claude Code or Codex, but when a user gives feedback, the model fine-tunes that knowledge into its own weights and remembers it. For tasks it has not learned, it calls frontier models from Anthropic or OpenAI. Boltzbit started in financial services, is now horizontal and infrastructure focused, and is preparing a consumer product.

Where does the name Boltzbit come from?

The name combines the Boltzmann machine, an early neural network introduced by Geoffrey Hinton in the 1980s, with "bit" from computer science. Yichuan Zhang chose a name rooted in the history of learning algorithms because his company's core belief is that learning, rather than stored knowledge, is what makes AI intelligent.

What is AGI 2.0 according to Yichuan Zhang?

Yichuan Zhang describes AGI 2.0 as the shift from centralized AI to distributed AI, comparing it to computing's move from the mainframe to the PC. In his view, today's AI can already handle most everyday tasks, so the next phase is personalized models that each person and business owns, where the benchmark becomes speed of learning, cost, and adaptability.

Why does Yichuan Zhang think prompting and skills files are not enough?

Yichuan Zhang separates learning from memorization. Skills files and prompts store instructions the model retrieves, but many real preferences cannot be written down, and real workflows are full of exceptions no one has time to list. He argues natural language is a bottleneck for intelligence, and that a model should learn from examples and corrections the way people learn from experience.

What is it like to work at Boltzbit?

Boltzbit is a fifteen-person team in London, about half research and backend engineers and half business and marketing. Yichuan Zhang describes a very high bar for engineering hiring. The company builds its products fully with AI, keeps humans responsible for high-stakes decisions, and measures engineering performance through services delivered, users reached, stability, and scalability rather than lines of code.

Who is Yichuan Zhang's hero archetype?

Yichuan Zhang's primary archetype is The Sage, who explains AI from first principles about what learning is. His secondary archetype is The Rebel, dissenting from the idea that a few providers should own model weights. His classical hero match is Prometheus, who gave people the ability to learn crafts for themselves. His pop culture match is Yoda in The Empire Strikes Back, who teaches that size is not the source of power.

How does Yichuan Zhang think about AI and intellectual property?

Yichuan Zhang believes personal AI can create a new intellectual property system. In his view, people keep their own data and share it only through their own AI. Anyone who wants to use that knowledge has to ask that AI and pay for it, so creators keep control of what they know instead of having it absorbed into someone else's model for free. He says Boltzbit is building toward this.

What is Yichuan Zhang's advice for founders?

Yichuan Zhang tells founders to find the right vision and believe in it. Finding it means challenging yourself honestly rather than only asking others, simplifying your vision down to brutal facts and following first principles. Believing in it means not changing course just because a competitor launches something similar or a market seems to disappear.