Techstrong TV — Tue, Aug 4

When AI Starts Building AI

Alan Shimel with Kunal Bhatia

Self improving AI is no longer a research curiosity. Kunal Bhatia, CEO and co-founder of Hexo Labs, joins Alan Shimel on TechStrong TV to explain how SIA, the first self improving AI system, trains both the harness and the weights of another agent at the same time. Furthermore, Kunal argues that enterprises must own their own intelligent stack before the frontier consolidates. Kunal opens with his path. He is on his third AI startup and has been building in AI for more than a decade. Consequently, he brings a rare view of how quickly the ground shifts, from an early Alexa style device in 2014 to a general purpose learning app built on GPT-2 in 2020. Alan then digs into Hexo Labs. It started as a consulting business that grew a strong contract research pipeline. As a result, Kunal and his co-founder productized what they had learned into an agent that trains other agents, and they shut the profitable services business down to focus on the platform. Kunal walks through SIA. In addition, he explains that if the meta agent points at itself, the target agent becomes the meta agent and the loop starts to recursively self improve. Therefore, teams get an AI that behaves like a full time researcher and can push new weights, not just tweak the harness. Alan and Kunal close on the strategic stakes. Meanwhile, Satya Nadella, Alex Karp and a wave of open weight releases from China all point at the same shift. Enterprises must own their intelligent stack, invest in the loop from data through model weights to production context, and use self improving AI to close the talent gap. Learn more at hexolabs.com.