What if AI could build better versions of itself — without human experts in the loop? That’s the core idea behind SIA (Self-Improving AI), the debut open-source project from Hexo Labs.
In this episode of the VMblog Expert Interview Series, David Marshall speaks with Kunal Bhatia, CEO and co-founder of Hexo Labs, fresh out of stealth mode. Kunal breaks down how SIA works, why it’s outperforming specialized AI agents on benchmarks like MLE-bench and LawBench, and what it means for the future of scientific research and human civilization.
What is SIA (Self-Improving AI)?
SIA is an open-source framework and research paper that demonstrates how an AI system can autonomously build and improve itself. Rather than specializing for one type of task, SIA is a general-purpose self-improving system — give it any problem, and it iterates on itself until it solves it, then keeps going. On MLE-bench, it didn’t just reach #1 — it beat its own previous record repeatedly, similar to how Michael Phelps kept breaking his own world records.
Why open source?
Kunal argues that one of the biggest risks in AI isn’t the technology itself — it’s the concentration of power among a handful of closed labs. Open-sourcing SIA is a deliberate move to make powerful self-improving AI accessible to a broader research community, encouraging competition and contribution rather than monopolization.
Accelerating superintelligence, not achieving it
Hexo Labs’ mission is to accelerate the transition to superintelligence — not to define or claim it. Kunal draws a compelling parallel: just as hunter-gatherers couldn’t have imagined a farming society, we can’t fully picture what a world with superintelligence looks like. What we can do is help get there faster.
Partnering with frontier scientists
Hexo Labs is already working with researchers at Stanford, Oxford, UC Santa Barbara, and other leading institutions — applying SIA to hard problems in quantum technologies, materials science, new energy, defense, and biotech. The Hexo Labs Grant Program extends this further, offering researchers a mix of capital and GPU compute infrastructure to accelerate their experimentation.
The ethical question
Is self-improving AI dangerous? Kunal believes it’s a net positive — but acknowledges the real concern is power imbalance. The antidote, in his view, is openness: more contributors, more competition, and more people with access to the technology.
🔗 Learn more about Hexo Labs and SIA: HexoLabs.com
