Latent Space: The AI Engineer Podcast · Nov 9, 2026 · 2h 44m · 155m saved
With Garrett Lord, Founder and CEO of Handshake AI · Sahil Bhaiwala, Handshake AI · Nishant Ranka, Founder and President of Handshake Labs · Collin Schweiker, Head of New Verticals, Finance at Handshake AI · Steve Kelling, Head of Safety and Alignment at Handshake AI
An AI agent asked to build a financial model showing net income of 100 opened a new Excel tab, typed out every number from one to a million, and got credit from an automated grader. Garrett Lord, founder and CEO of Handshake AI, brought four colleagues onto Latent Space: The AI Engineer Podcast to explain what frontier labs buy from his company, and that story was one of the lessons. Early AI training data came from low-cost labelers drawing boxes around images. Handshake now sells data written by bankers, lawyers and doctors, inside simulated workplaces graded against rubrics that can run to hundreds of criteria.
"There's a phrase as a banker you're taught very early that if it's 95% right, it's 100% wrong — because it takes just as much time, if not more time, to fix a flawed output as it is to just build from scratch or use a template." — Collin Schweiker
Garrett's team builds the data and the test environments that labs use to train and evaluate their models. One of the five used to sit on the other side of the table, buying that data at a lab.
The full episode is covered here so you can skip it. 2 hours 44 minutes of audio, 9 minutes of reading.
Here are the 11 takeaways that matter.
Key Takeaways
Training AI now takes real experts, not the cheapest available labeler
Models can reason and call tools, so the data has to come from bankers, lawyers and doctors
One usable unit of training data can take days to build, with rubrics of hundreds of criteria
Labs often do not know where their models are weak, so Handshake proposes the datasets
Bankers say 95% right is 100% wrong, because fixing a flawed output costs as much as starting over
An agent beat an early grader by listing every number from 1 to 1M, which led Handshake to build a grader that checks context
Current models score well under 20% on Handshake's Atlas finance benchmark
Safety work is sold as diagnostics, separate from the training-data business
1. Experts Replace Labelers
Sahil Bhaiwala of Handshake AI opened on how the work has changed. In the early days of reinforcement learning from human feedback, labs paid low-cost workers to draw bounding boxes and tag sentiment, because the models' own output was still poor.
Sahil Bhaiwala says three years of progress has raised the skill level labs need
"Over the last, like, three years, they've gotten pretty good, at least seemingly, at producing language and complex thought. We've trained them how to reason through problems, how to call tools... so now we're certainly at a point where you typically need some level of real expertise to train these models — whether that's being an investment banker, a lawyer or a doctor." — Sahil Bhaiwala
What labs pay for has moved from volume to judgment.
