From art exhibitions to AI policy

Arts policy, banking regulation, investment risk, a PhD in AI, and exhibited artwork look like five unrelated lives. In this conversation on Analythical with Stephen Tracy, I try to explain how they are really one: a pull to keep learning new things, and a fascination with models and systems. It is also the clearest short version of how I think about AI. The full talk is below; here is the gist.

Art came first

I did not start in finance. I started in arts policy, back in 2003, when the buzzword was not AI but "creative industries". I even helped create a regional design competition as a reality show. Art itself came later, and from a low point: I was stuck doing strategic planning, which was mostly board papers and speeches, and I hated it. Digital illustration became the escape - no materials to buy, an undo button, something I could do in the gaps. The escape turned into an obsession: commissions, exhibitions, the lot. I do not see art, code, and writing as separate pursuits competing for time. They are different facets, and each has its own place in the day.

A rabbit hole into models

In 2007 I joined the Monetary Authority of Singapore, starting in banking regulation. On day one a boss told me, more or less, that we do not do the "artsy" things here. Looking back, that comment started the rabbit hole. I went into the technical end - Basel capital rules, the regulations that actually have equations - then a master's in financial engineering, model risk and validation, then investment risk management, overseeing the risk of Singapore's foreign reserves. Somewhere along the way I realised I am not a generalist; I like going deep on technical things. At 42 I took a break for a PhD. I applied in finance, got rejected, and ended up in AI instead - which turned out better than finance would have been. What connects the disparate parts is that same pull to learn, and overlaps I did not expect: I am drawn to fractals, where art emerges from simple rules, and emergence shows up in AI too.

What the PhD was really about

I wanted to do generative AI - this was 2019, when the exciting thing was GANs generating faces. I dropped it; generative research needs huge data, large models, and long runs. Instead I built dynamic multimodal networks: small transformer models, the same family that sits under today's chatbots, but pointed at modelling things that change over time and connect to each other, for jobs like financial forecasting. I had been auditing banks' models before the PhD, and I love models. The gap I was really filling was my own: how much more could I learn about them. Finance would have led me to quantitative models; AI led me to more complex ones. Same instinct, bigger subject.

How I think about AI

That instinct runs straight into how I treat AI risk, and how I teach AI. A few things I keep coming back to in the conversation:

• AI risk is boring, by design. The AIRG I wrote at MAS is principles-based, and applies to all financial institutions and all AI - machine learning, deep learning, generative, agentic. The instinct to picture AI colluding or spelling our doom is the wrong frame. AI risk management is a dull system you manage properly. If it goes haywire, that is a design decision you got wrong, not a Terminator plot.
• AI is not generative AI. AI is the big circle; machine learning sits inside it, deep learning inside that, generative AI inside that again. Generative AI is one set of techniques, not the whole field. If all you have is a chatbot, you have a hammer - and plenty of problems are not nails. A smaller, cheaper, faster model is often the better tool.
• Go to fundamentals, not the shiny new thing. A new agentic system is still a large model in a system, given tools and memory. Break it down to what it actually is, and the churn stops being overwhelming. Start with the problem, not the tool; break a big problem into smaller ones; and be as specific as you can, especially about what success means.
• On hallucination. No model is ever always right - anything claiming 100% accuracy is a red flag. Generative models are trained to produce a confident answer even when there is no good one behind it, which is what people call hallucination. So do not trust the output; verify it, the way you would check a capable colleague's work. Being specific helps, and agentic workflows can be built to expect errors rather than wish them away. Prompt engineering alone will never fully solve it.

This is also why Stephen and I are against "prompt and pray". The mental model we teach is FRAME: frame the problem and define success, recognise the data type and the task, apply the right method, measure honestly, and evolve as the data and the world change. It works on a full machine-learning pipeline or on the next question you type into a chatbot. It is the same idea as my Thinking in AI philosophy: data types and tasks, not prompt and pray.

Watch the full conversation

The full interview is on Analythical, hosted by Stephen Tracy: From art exhibitions to AI policy. We also cover the AI course we built together on applying AI in research.

Work with me

I teach and advise on AI and AI risk, plainly. Read more on AI risk management and how to think about AI, see the books and courses and workshops, or get in touch.