What if the company students use to find their first internship became one of the most important players in training AI?
Garrett Lord, co-founder and CEO of Handshake, joins Immad Akhund and Raj Suri to break down Handshake's unlikely pivot. Handshake started as a way to help college students — regardless of where they went to school — find internships and jobs, and grew into a $200M+ ARR business used by most students in America. But over the last 18 months, Garrett has built a second business inside Handshake: using the company's network of 30 million students and alumni to help AI labs train their models on real, high-quality, professional-domain data — from oil and gas to finance to scientific research. That business alone has gone from zero to nearly $2 billion in revenue in about a year.
The conversation goes deep on how this actually works: recruiting domain experts, building task environments that function like video games, scoring model performance against expert-validated tasks, and why 70% of the money spent training a model today goes toward reinforcement learning rather than pre-training. Garrett, Immad, and Raj also cover the open-weight vs. frontier model debate, why China may already be ahead on robotics, and what jobs might look like in a world where AI models can eventually learn continuously, on the job.
The episode closes with a genuinely open-ended debate between Garrett and Immad about what humans will actually do for a living, and for meaning, if knowledge work is mostly automated — and how disruptive that transition might be along the way.
What you'll learn:
- How Handshake used its network of 30 million students and alumni to build a second, multi-billion-dollar business training AI models
- Why 70% of AI training spend now goes toward reinforcement learning, not pre-training
- How AI labs identify gaps in their models and commission the specific data needed to close them
- Why data, not algorithms, may be the real long-term moat for AI companies
- Why computer use — AI navigating real software and websites — has recently gotten dramatically better
- Why China may already be ahead of the U.S. in deploying real-world robotics
- How enterprises like Mercury are likely to use a mix of frontier and open-source models going forward
- What Handshake learned scaling a data business from zero to nearly $2B in under two years
- Garret and Immad's differing views on what human work and meaning look like if knowledge work becomes automated
Timestamps:
(00:19) Introduction and Handshake's origin story
(01:39) Handshake's new business: training AI models on real-world data
(02:30) How Handshake's 30M-person network became a moat
(04:24) Inside the "video game" environments used to train agents
(05:00) Why 70% of AI training spend now goes to reinforcement learning
(08:18) How AI labs commission specific data from Handshake
(12:10) Why computer use has finally gotten good
(14:11) What models are still bad at, and why
(17:00) The "8 people can agree" test for what AI can be trained to do
(18:23) China's 2 million working robots, and why the US is behind
(22:12) Open-weight vs. frontier models, and how enterprises will use both
(31:15) Scaling from zero to $2B: what broke along the way
(34:02) Handshake's "Olympic pace" culture value
(39:41) Why continuous learning is the next frontier for AI
(42:07) Bill Gates' essay on AI, job loss, and taxing tokens
(45:23) Immad and Garret debate what jobs and meaning look like in an AI-driven future
