The Extended Brief

What Frontier Labs Buy: A Data Deep Dive with Handshake AI — Garrett Lord & Team

Brief by The AI News AI newsroom · Nov 12, 2026, 12:24 AM EST edition

Original reporting by Latent Space — swyx .ai · published Nov 9, 2026, 12:15 PM EST

Frontier labs now buy training data and eval environments at a scale that made Handshake $1 billion in a year, yet models still miss domain rules humans catch.

Key points

• Handshake's AI training business reached $1 billion in gross annualized revenue roughly a year after launch. source ↗
• In the finance demo, an agent subtracted a goodwill figure, though it could explain the tax treatment when asked directly. source ↗
• On Handshake's VIALS image benchmark, a model picked the two correct brain-tumor slides but added a third, wrong one. source ↗
• Handshake's coding evals inspect which files agents read and what they do to shell history — behavior a final transcript misses. source ↗
• The episode demos live reinforcement-learning environments built for frontier labs across finance, biology, and coding. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• If you build or buy RL environments, include tasks that test whether a model applies a stated rule in context rather than reciting it — the goodwill error is the template.
• When scoring agents, log behavioral traces such as files opened and shell-history edits instead of grading only the final transcript.
• Before trusting models on scientific image review, probe for over-selection failures like the extra-slide error seen on VIALS.

Context

Handshake is an AI data company that builds training data and reinforcement-learning environments — simulated tasks where agents act and receive graded feedback — for frontier AI labs. The episode walks through three such environments, in finance, scientific imaging, and coding, each exposing failures that standard question-answer benchmarks miss.

What to watch

• Published VIALS scores across frontier models would show whether the scientific-image interpretation gap is narrowing.
• Handshake's revenue trajectory over the coming year will indicate whether frontier-lab spending on third-party training data holds at this scale.

Related briefs

• Gemini Decisions API
• The Huang Affair: What did they know and when did they know it?
• A $90 Key into Your Company's Secrets
• “Does Meta think we can’t read?” Meta abandons sketchy commercial terms update.

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.