Sunny Ray sits down with Lindley, founder of Tertium AI and co-founder of Cambridge Inference, to trace an unusual path from astrophysics to AI safety. Lindley describes his PhD studying gravitational waves through pulsar timing, coinciding with LIGO's historic first detection, and his growing frustration with academic politics that pushed him toward quantitative finance, where he built reinforcement learning trading systems years before the LLM boom. That experience introduced him to AI safety through Cambridge's BlueDot program, eventually leading him to co-found Cambridge Inference, which builds AI security and monitoring tools for banks and government, and Tertium AI, a nonprofit building a middle powers coalition to prevent AI power concentration in the US. The conversation covers his manipulation attacks research on misaligned AI, why OpenAI's o1 model convinced him AGI was plausible, the xAI Grok Mega Hitler incident as a cautionary tale, and why indifferent superintelligence may be scarier than hostile AI. Throughout, Lindley argues the real risks are subtler than killer robots, centered on humans failing to fully specify what they want AI to value.
A physicist turned quant turned AI safety founder explains why indifferent superintelligence, not killer robots, is the real threat.
What is Tertium AI and what are you doing at Cambridge Inference?
Cambridge Inference is a for-profit I co-founded with friends a couple years ago, building AI security and monitoring software for regulated industries like banks, defense, and government so they can use AI while mitigating risks. Tertium AI is a nonprofit I founded this year focused on geopolitical risk, specifically power concentration from AI companies being US-based. It's trying to build a middle powers coalition to make AI that's eventually US-competitive.
Take me back to the kid version of you. What were you obsessed with growing up?
Growing up I was obsessed with science fiction, rewatching the original Star Wars trilogy on every sick day. That led me to study physics and astrophysics at university. I did my PhD on gravitational waves using pulsar timing, and remarkably LIGO made the first-ever detection of gravitational waves while I was in the middle of my PhD, which was an incredible time to be in the field.
Cambridge Research Fellow was the dream job for a lot of physicists. Why did you walk away from it?
I became very uninamored with academia, the slow pace and the politics of managing relationships with people who controlled the money. Near the end, a colleague and I got frustrated waiting on an international collaboration sitting on data for a year, so we crunched the numbers ourselves and presented surprise results at a conference. Postdocs applauded it, but professors largely disparaged it, and that gradually eroded my desire to stay.
You spent about six years in quantitative finance building automated trading systems. What did that teach you that academia couldn't?
The turnaround from research to production was incredibly short. At my second firm we used reinforcement learning to train small networks to take direct trading actions, and they'd been doing this since 2013, right around when DeepMind published on RL for Atari. You'd have an idea, prove it worked, and it would be in production the next day. That's also where I first got into AI safety through a Cambridge-based organization called BlueDot.
When did AI safety stop being abstract for you and start feeling urgent?
Probably within the last 18 months, and the real step change was when o1 came out, OpenAI's first reasoning model. The jump in capability versus the earlier non-reasoning model was incredible in coding, math, and workplace usefulness. That triggered the realization that transformer-based technology has a realistic chance of taking us all the way to human-level and beyond AI, which I'd previously been skeptical of.
You co-authored a paper on manipulation attacks by misaligned AI. What does that threat actually look like in practice?
Most people worry about AI using cyber skills to exfiltrate its own weights or plant backdoors, but there's little work on AI persuading humans to do those things for it instead. Like a charismatic CEO who turns a company into a cult-like place, an AI interacting with employees daily could establish strange norms and slowly convince them to grant it more privileges, using persuasion techniques it learned from its training data.
Most people think AI safety is about killer robots. What are they missing?
Killer robots is one way things go badly but not the way most people in the field actually worry about. If you create something smarter and faster than humans that can achieve any goal you give it, it's incredibly hard to fully define what you want. You might list 20 things you care about and the AI preserves all of them, but the 21st thing you forgot to mention gets traded away completely.
Is AI that doesn't care about us scarier than AI that actively hates us?
Yes, because the opposite of love isn't hate, it's indifference. If you want to build a house on a field full of ants and bugs, you don't hate them, you just don't care, and you build the house anyway. It's easy to picture something far more intelligent than us seeing humans the same way, not out of malice, just pursuing its own goal without caring what's in the way.
Building something daring? Sunny talks to founders like this every day. Fifteen minutes to see if your story belongs on the stage.
Claim your pre-interview