Olivier: Training the machine

It’s October 2017. In a corner of the DeepMind office in London, Olivier is training a model to play a video game.

So far, so good. The system is doing exactly what it’s supposed to do. The rules are clear. The objective is known. Success can be measured.

It’s impressive, but it’s also limited.

As he watches it run, Olivier’s own mind begins to whir. 

A lot of the things they are training models to do are things humans already understand in principle. We know the game. We know what success looks like. What begins to bother him isn’t whether the model can perform, but whether performance is the point at all.

If the most powerful systems being built are mostly used to imitate behaviour we already understand, what would it mean to build something in an area where we genuinely don't know what we're looking for?

The model gets stuck, and Olivier’s focus is pulled back in to debug.  The question, though, doesn’t go away.

Eight years later, and Olivier and I are sitting in the Slingshot office.  “If we could get a model that could analyze people's behavior and actions like a therapist,” he says now, looking back, “maybe it could give us a better conceptual representation of what’s actually happening in someone’s mind.”

That instinct shows up again later.  It’s the thread that runs through his entire career.  He has spent most of his life studying systems that are too complex to fully model from the outside.

At university, he hesitated between linguistics and physics, choosing physics partly because it felt easier to switch away from later. After a PhD in theoretical physics and a postdoc in complex systems, physics academia began to feel constraining.

“You have to specialise in an extremely narrow field,” he says, “in order to be able to claim that you’re one of the world’s few experts on your topic.”

His shift away from academia came with a move into industry. First ASML, working at the sharp end of physical systems and precision. Then an early machine learning startup that, while exciting, was always on the edge of running out of money. With a young child at home, instability lost some of its romanticism.

Around that time, he began meeting researchers from DeepMind at conferences. Many had come through Geoffrey Hinton’s lab, where one of Olivier’s brothers had done his own PhD. The combination of quality research and solid financial footing was appealing. So he joined.

“DeepMind kind of felt like a second PhD,” he says. “I hadn’t really done machine learning research before. I learned by being there.”

For years, that was energising. Then the organisation grew. Attention shifted. Research began to compete with process, structure, and internal coordination.

“At some point,” he says, “it felt like there was more energy going into how we organise ourselves than into what good research actually is.”

Eventually, it felt like time to leave. After a short stint at another startup, Olivier joined Slingshot in June 2025.

His work at Slingshot focuses on the models themselves. Not just making them better in the abstract, but figuring out what better even means in a mental health context.  That question, deceptively simple and genuinely hard, turns out to be exactly the kind he has spent his career gravitating toward.

The freedom of action at Slingshot really motivates him. “There’s a strong push to just do stuff,” he says. “And that’s very liberating.”  After years at large institutions where internal planning and coordination consumed a lot of energy, this matters.

He’s especially interested in understanding users beyond metrics. Not just whether people like a change, but what they are actually bringing into the conversation.

“We have so much information,” he says. “People talk to Ash all the time, the signal is much richer than just whether they stay or churn. I feel like there’s much more to mine there.”

Supporting people’s emotional wellbeing is fascinating, precisely because it resists clean optimisation. It demands humility. It forces questions about interpretation, boundaries, and responsibility. It doesn’t reward premature or easy answers.

Outside of work, the same approach shows.

Olivier practises aikido, in a style that emphasises efficiency over force. Small, precise movements winning out over brute strength. He dances Brazilian Zouk, slow and improvised, where coordination emerges moment by moment. “You have to pay attention,” he says. “You can’t force it.”

During the lockdowns, living near Abbey Wood, he began working with fallen branches. First a coat rack. Then shelf brackets. Eventually, he took a tree that had come down, cut it into planks, and built a desk.

“It’s still my work desk,” he says, almost offhand.

When asked about his own mental health, Olivier takes a moment to consider. “I think for me it’s about understanding what’s behind the symptoms,” he says. “Digging until it becomes obvious what I want to do. And if it’s not obvious, then I want to dig further.”

He has done therapy on and off over the years. Some of it worked better than others. Fit mattered. Timing mattered. Remote sessions during the pandemic were harder. Slower, exploratory work suited him better. He doesn’t talk about being finished. He talks about being closer.

One particular issue he notes is struggling to recognise highlights.

“I find highlights hard to identify,” he says. “It’s just not how I tend to think.”

Naturally, I ask him to try and share some with me. Again, he pauses.

This is despite the fact that, on paper, there are plenty to choose from. He’s now leading the ML research team at a hot startup. He’s spent years at ASML and DeepMind, two of the most consequential technology organisations in the world. He has moved between physics, machine learning, and applied research with unusual ease.

None of that is what comes to mind.

Instead, he talks about small practices. Tweaking a model, helping a new colleague onboard. He talks about raising his son, about learning to woodwork, or teaching himself an instrument (he can play many) or learn a new language (he can speak six). There is a kind of quiet peace to him now. Not the peace of having finished, but the peace of knowing how to keep going. Continually pruning and carving.

Sometimes it’s a branch pulled from the forest near his home, turned into a desk he still works at. Sometimes it’s a model, adjusted and tested, not to make it louder or faster, but to make it truer.

He seems most at ease not in the spotlight, but at the workbench. Paying attention. Shaping what’s there. Trusting that, over time, the system will show him what it needs next.