Aditri Bhagirath: Ex-Meta, Stanford, Carnegie Mellon
Aditri Bhagirath traces a career that runs from undergraduate research in Tom Mitchell's machine learning lab at Carnegie Mellon, through four years at Meta building predictive models to surface high-quality internal content, to a Stanford master's in CS focused on AI. She grew up in New Delhi as the daughter of two dentists, barely knew what a for loop was when she arrived at Carnegie Mellon, and is now a machine learning research engineer at Anton, an AI shopping startup.
She breaks down core AI concepts for a general audience, distinguishing AI, machine learning, and AGI, and argues we have loosely reached a form of AGI while true AGI may still be a couple of steps away in both compute and architecture. Along the way she discusses her ACL paper on modeling user preferences for personalization, world models, robotics, quantum computing, and her central message: you do not need to be technical to build powerful things with AI.
“AI should be thought of as a tool and a thought partner, not necessarily just to completely offload your brain to and rely on cognitively as the primary source of ideas and inspiration.”
“I think a lot of people have this misconception that you have to be super technical to build something, but I don't really think that's the case.”
“I'm really excited about training models to essentially replicate the soul of an artist.”
Key takeaways
- 00:01:10After Carnegie Mellon she spent about four years at Meta building predictive models that flagged high-quality, important information for thousands of internal employees.
- 00:03:05She is now a machine learning research engineer at Anton, an AI shopping startup that interprets natural-language queries like finding the perfect couch for a Brooklyn apartment under $3,000.
- 00:06:40Her ACL paper studied pairwise user preferences to predict future preferences far more accurately than the prior state of the art, with applications across e-commerce personalization.
- 00:09:56She noted that despite its power for scaffolding entire systems, Claude Code had only around 300,000 users worldwide at the time of recording.
- 00:13:00She argues AI has loosely reached AGI for many tasks but that true AGI is still a couple steps away, pending advances in quantum computing and model architecture.
- 00:13:44She is bullish on small, on-device language models from companies like Liquid AI whose architectures are inspired by biological systems such as earthworm brains.
- 00:20:32Her advice for founders using AI coding tools is to trust but always verify, keeping strong evals in place and monitoring system inputs and outputs.
- 00:24:50She closes with the message that you never have to be technical to use AI, citing lawyers, marketers, and doctors who have 10x'd their abilities with it.
Chapters
- 00:00:00Welcome and the highlight reel
- 00:00:38Carnegie Mellon and Tom Mitchell's lab
- 00:01:10Four years at Meta
- 00:03:05Now at Anton, shopping with AI
- 00:03:36Growing up a dentist's daughter in New Delhi
- 00:06:03The ACL paper on personalization
- 00:09:19AI coding tools and world models
- 00:12:25Have we reached AGI?
- 00:14:21AI 101: AI vs ML vs AGI
- 00:19:54Advice for founders: trust but verify
- 00:21:41Robotics and quantum computing
- 00:24:50You don't have to be technical
