Building Self-Improving AI Systems That Work
Gaurav Sahu - Unified Sciences
Sept. 25, 2026, 3:30 p.m. - Sept. 25, 2026, 4:30 p.m.
ENGMC 11
Hosted by: Oana Balmau
Today’s AI systems talk really well, and are widely used in chatbots and many other applications. But can they do more than answer questions? Can they actually do research?
A scientific endeavour requires one to build a deep understanding of a domain, think of promising hypotheses, design experiments to test them, interpret the results, and separate the good ideas from the bad, all while maintaining rigor and reliability. The best researchers often “fail fast,” not by steamrolling carelessly through ideas, but by designing the fastest path to determine whether an idea works or not, and then using that evidence to guide the next step.
In this talk, I’ll discuss our work at Unified Sciences on building autonomous research systems that close this loop, and how they’ve made advances in GPU kernels, vector search, and long-context computing. I’ll show how AI systems can accumulate knowledge through experimentation and use it to push the frontier. I’ll also discuss why many agentic systems wander or optimize the wrong objective, and what it takes to make them pursue a research frontier coherently. Finally, I’ll briefly discuss how these ideas might extend beyond software to the physical sciences and enable breakthrough discoveries.
The central message is simple: moving from AI that talks about discovery to AI that makes discoveries requires disciplined, experimentally grounded, self-improving loops.
Gaurav Sahu is the Co-Founder and CEO of Unified Sciences, where he is building the infrastructure for AI-accelerated science. His research spans self-improving agents, language-model reasoning, and autonomous research systems, with a focus on accelerating scientific and engineering workflows with AI. He was previously a postdoctoral researcher at Mila – Quebec AI Institute and earned his PhD in CS from the University of Waterloo. He also manages the infrastructure of leading AI conferences like ICLR and is building tools to scale peer review.
