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

At 22, most people are still figuring out what they really want to do with their careers. Singaporean student Mathew Vanherreweghe already has his dream job: he’s rebuilding AI models from the ground up at the heart of Silicon Valley.

The firm he works at, Pebblebed Ventures, even gave him a big title: AI research scientist. It’s a paid internship role for now, but Vanherreweghe expects it to turn into a full-time position once he gets his US work visa.

Image credit: Arsal Ysfin

But why does a VC firm need an AI researcher? For Pebblebed, having an in-house researcher aligns with its intention to “get [its] hands dirty as to what’s happening on the field,” says co-founder and partner Tammie Siew.

The company was launched in 2022, and its founding team includes OpenAI co-founder Pamela Vagata and Facebook AI Research founder Keith Adams. In addition to providing capital, the VC firm also works with its portfolio companies on technical matters, such as writing production code and training foundational models.

In the three years since ChatGPT made transformer architecture the dominant standard, major AI labs have poured billions of dollars into scaling transformer-based models to ever-larger sizes. However, since late last year, this approach has begun to yield diminishing returns, prompting a search for alternatives.

For Pebblebed, this means hiring Vanherreweghe to prove whether the transformer architecture can be rebuilt to make it cheaper to train and easier to build on.

The 22-year-old didn’t land this hot seat by chance. In May 2025, he published a paper proposing a new neural network architecture – the structure that controls how the model processes data – based on an old math theorem.

See also: The Asian brains forging AI models at OpenAI, Google, Meta

The new network could match the performance of a transformer-based model working on financial predictions, despite being around 10x smaller, which means it requires less computational power and time to train.

That caught Pebblebed’s eye.

“Most efficiency work takes an existing model and tries to make it smaller. Mathew is asking a more fundamental question: What if we built the architecture differently from the ground up?” Vagata and Adams say in a joint response to Tech in Asia.

While the results remain preliminary, the work represents a growing question about whether AI’s “bigger is better” paradigm has alternatives worth pursuing. As one research published by UNESCO shows, smaller models tailored to specific tasks can cut up to 90% of energy consumption.

From theory to practice

A math major at the National University of Singapore (NUS), Vanherreweghe says he learned programming through college courses after getting excited about what he could build with AI during the early days of ChatGPT.

Smarter math for leaner AI

Et tu, Liu

What comes next