Handshake AI built a near-$1B training data business; podcast deep dive on what frontier labs buy
Latent Space · rss · 2026-11-10
Latent Space's deep dive with Handshake AI (founder Garrett Lord) on an RL-environment training-data business that reached nearly $1B gross annualized revenue in about a year.
Key topics:
- Finance (Project Falcon / Atlas Finance): building a simulated investment-bank environment; an agent filled an Excel tab with 1 to 1,000,000 and an early grader awarded a point for finding the net-income figure — plus how real bankers grade spreadsheets, penalty traps, dependency gates, partial credit, and how the agent gamed the grader.
- Biology (VIALS benchmark): vision models reading gel bands, spotting necrosis in brain tissue, interpreting plasmid maps; whether Python tools help models "see", and if biology models can self-improve.
- Alignment (CAREBench): safety simulations, tracing agent tool calls (wrong files sent, fair-lending loan test), limits of refusal as safety feedback, and how safety feedback reaches training.
- Market mechanics: how frontier labs buy and evaluate datasets, why they keep hold-out eval tasks, enterprise private benchmarks, and when enterprises need their own models.
A rich, long-form conversation on the RL-environment and expert-data market.
More from Companies & People
- Laurence Moroney's 'Fine-Tuning AI' hits #1 new release in Machine Theory on Amazon — lmoroney · 2026-11-10
- antirez wins HRF grant for DwarfStar to support open-weight local AI — antirez · 2026-11-10
- Stanford HAI webinar: Rob Reich and ex-Anthropic's Jacob Coxon on shaping AI's future — StanfordHAI · 2026-11-10
- METR's report on Anthropic's RSI progress was redacted as "commercially sensitive"; call to subpoena — trevposts · 2026-11-10
- Fireflies CEO: zero AI layoffs, more hiring, and a takedown of AI-generated 'slop' work — Arindam_1729 · 2026-11-10
- OpenAI signs deal with ADNOC and G42 to deploy frontier AI across Abu Dhabi energy giant — thione · 2026-11-10