The 22-year-old student rethinking AI

Welcome to Tech in Asia’s daily newsletter, your essential dose of Asia’s tech and startup buzz. Not yet a subscriber? Register here. Got a story tip? Send it to editors@techinasia.com.

In focus

• ​​Ambitious Singaporean student working on the next AI breakthrough in Silicon Valley
• ​​A past look at the growing prevalence of tinkerer VCs
• Grab-backed Superbank readies itself for public listing

Hello reader,

2011 was a definitive year for my career. I was 22 and had just started an internship role as a lab assistant at my university, where I studied English education.

Later in the year, I completed a three-month teaching internship at my alma mater and decided that teaching was not for me. When I graduated the next year, I landed a job at a local tech and gaming magazine instead, and I never looked back.

I’m sure many of us have no idea what to do with our life or career at 22 years old.

That’s not the case for Mathew Vanherreweghe. The National University of Singapore student already got his dream job of rebuilding AI architecture with Silicon Valley heavyweights.

It’s not everyday you see a Singaporean student working full-time with the co-founders of OpenAI and Facebook AI Research. I break down how Vanherreweghe got to this point and what sort of breakthrough he and the team are trying to achieve in today’s Top Story.

In essence, it’s another case of folks challenging the “bigger is better” paradigm held by AI giants. But this time, it involves an ambitious youngster and an old math theorem from the 1950s.

The work is still in the early stages, but if it scales, it could make training AI models cheaper and more accessible beyond the handful of companies that can afford billion-dollar compute budgets.

Glenn Kaonang, journalist

Top Story

NUS student joins OpenAI co-founder to rethink how AI models work

Photo credit: Mathew Vanherreweghe. Illustration by Arsal Ysfin / Tech in Asia

In May 2025, Mathew Vanherreweghe published a paper proposing a new neural network architecture that was based on an old math theorem. Despite this network being 10x smaller, it could match the performance of a transformer-based model working on financial predictions.

From our archives

Trending news