From Maths to Models: A Winding Path into AI

Kid

As a kid, I spent most of my time solving maths problems, playing in chess tournaments, and practising music. I did well enough in each of these, but there was always a nagging feeling that I could never quite push over the hump into “true excellence.” In maths, I was invited to an IMO training camp but never made the team. In chess, I played on the national squad but was a dirty 1.e4 player and the weakest member of my school team. In music, I became technically proficient but never had the drive or passion to commit to the long hours of practice it really demanded.

During this time, I lost my mum to cancer. It was a painful period, and it planted in me a desire to work on healthcare-related problems. But I also felt like my skillset wasn’t well suited to medicine - especially with my poor memory recall, which I knew would be a problem. In the end, I stuck to what I knew: maths. It was a source of comfort.

I considered applying to study Engineering at university, but it felt like a high-variance option - I had no idea what to expect from the interviews. Cambridge maths admissions, by contrast, were known for being straightforward but tough. So I went for it and got into Trinity to study maths.

I still vividly remember stumbling across Andrew Ng’s Coursera course on ML back in 2012, while on a family holiday in Hong Kong. I was excited - like, “I just trained a net to recognise handwritten numbers, how cool is that?!” - but no one around me seemed particularly impressed, and I didn’t take it any further.

Uni

University maths was tough. In my first year, I was more excited about finally having the chance to play sports. Growing up, my weekends were always taken up by Chinese school and church, so I’d never had time. I jumped into rowing and basketball, and caught up on sleep during lectures. Unsurprisingly, I tanked - finishing with a 2:2.

In second year, I realised I didn’t see a future in pure maths. I had no desire to become an academic, and figured that for real-world work, there was only so much maths you really needed. I put in the effort to recover from my first year, managed to scrape a 2:1, and transferred into Engineering, hoping to do something more practical.

When I got to the Engineering department, I was overwhelmed by choice: quantum computing, nuclear energy, bridge building, medical imaging, cars, planes - you name it. I really enjoyed working in labs and collaborating with others. I took a wide range of modules but gravitated towards electrical engineering, AI, and medical applications. My favourite course was Computational Neuroscience - a sweet spot at the intersection of maths, medicine, and computing. Working on things I actually enjoyed paid off: I got a first.

Fourth year was disrupted by COVID. I’d been working towards a career in AI with a dissertation on causal inference and counterfactuals in clinical trials, but my supervisor was already stretched thin managing her group, and after the pandemic hit, contact hours disappeared. The internships and fellowships I applied to were cancelled, and I ended the year feeling pretty demoralised.

Career

After that, I wanted a break from AI but was still interested in research. I’d always been fascinated by AR devices like Google Glass and saw an opening at Microsoft Research in Cambridge to work as a research engineer supporting a team of optical engineers. The project involved designing multi-lens arrays made from 3D-tessellated hexagonal lenses, each with unique curved surfaces. To support this, we developed a CSG-based ray tracing library in Julia called OpticSim.jl, which I helped maintain and apply to prototype designs.

Unfortunately, Microsoft pulled funding for the role and I had to move on.

Around this time, I was curious about the food industry and had no clear direction, so I reached out to a friend who had helped start a food stall in Spitalfields Market selling biang biang noodles. I worked there for six months and quickly discovered that kitchen work is intense - something I’m happy to leave to the pros. After a near-death experience (a story for another time), I decided it was time to return to something more stable.

Many of my friends from the maths course had gone into quant finance. One of them referred me to Optiver, a high-frequency trading firm. I passed the interviews and was offered a spot in the next grad cycle, starting in six months. I used that time to intern at a systematic hedge fund, helping build signals for an FX strategy.

My time at Optiver was… intense. I started strong - mainly thanks to an ability not to panic under pressure - and was soon trading my own product on the main European index options desk. The trainee programme was exceptional, but once I graduated to the trading floor I increasingly found the workload exhausting. Intense screen trading during market hours followed by late-night research took its toll. Burnout crept in.

I also became disillusioned with certain aspects of the role and the company (something I might write more about in another post). After about 18 months, I left - without a clear plan for what to do next.

Break

2023 was a tough year. I was living alone in Amsterdam, far from friends and family, going through a difficult breakup, and not enjoying my job. I decided to take a six-month break to reset.

During that time, I learnt to box, ski, and scuba dive. I went camping in Switzerland and hiking in the Dolomites. Back in the UK, I picked up basketball and climbing again, and co-hosted some incredible dinner parties with my flatmate, the same friend I’d worked with at the food stall, who’d recently graduated from Le Cordon Bleu, complete with late-night runs to Billingsgate and Smithfield’s (iykyk).

I also caught the AI wave and started self-teaching out of curiosity. I began with the ARENA course from the AI safety community, which led me to an interest in mechanistic interpretability. That sparked a personal project focused on interpreting OthelloGPT.

I got in touch with a tiny startup working on agentic codegen and helped fine-tune a (questionable) diff model for them. I also reached out to a neuroscience professor at Oxford and began a collaboration training models on survival time data in long-term clinical trials.

I kept building. I replicated several research papers - most recently AlphaEvolve, and managed to achieve SOTA on a circle packing problem using my own implementation.

Interviews

Eventually, I felt ready to put myself out there and began applying for jobs. That’s when I realised just how crowded the space had become. A contact gave me some much-needed tough love: to succeed, I’d need to deliberately build up my public profile, portfolio, and network. This would take time.

Luckily, I already had a small network. Two referrals led to two promising interviews. The first was with a well-known codegen startup in San Francisco. They flew me out for a two-day onsite… which I spectacularly flunked, partly due to not being good enough, and partly due to being asked to book flights Saturday, arrive Sunday, and interview Monday-Tuesday. Brutal.

The second was at a drug discovery company in London, which was a much more enjoyable process. I made it to the final round but was turned down due to lack of experience for a mid-level role.

Now

So here I am - simultaneously encouraged and discouraged by these half-successes. At times, I’ve felt foolish or useless for attempting this career change. But I’m drawn to these problems in a way I can’t ignore, and I don’t see myself looking elsewhere anytime soon. So I stubbornly keep moving forward.

Right now, I’m working on patching some of the gaps that came up during interviews. I want to write up explainers on areas I was weak in, and build smaller projects that give me more experience. I’ve also come to realise that ML roles are becoming increasingly specialised, and I’d benefit from narrowing my focus - something I hadn’t seriously considered before. I have to keep telling myself that this is the difficult part, and it only gets better from here, but ultimately I feel lucky to be working towards something I’m deeply interested in.