01
Synthetic data
Hard problems with verified answers, built around the skills today’s models still lack.
Why it matters
Models have read what people wrote. They haven’t done what people do. That’s what we build.
What we build
Problems to solve. Worlds to act in.
02
Environments
Real tasks, start to finish. Graded on what got done, not what was said.
30–60 min
Synthetic data
1–2 hours
Environments
4+ hours
Long environments
How long today’s frontier models take on one task. Even a failed try teaches them something.
What we cover
First, general. Then, specific.
General intelligence is ready now. The subjects below are coming soon.
Ready now
General intelligence
Reasoning, planning, tool use, and staying on track across long tasks. The base every subject builds on.
01
Coming soonMathematics
Olympiad and research-level problems, with full proofs, not just answers.
02
Coming soonSoftware engineering
Whole codebases, hard bugs and contest-level algorithms.
03
Coming soonReal-world physics
Messy, many-step problems. No textbook shortcuts.
04
Coming soonChemistry
Reaction mechanisms, materials and drug discovery, at graduate level and up.
05
Coming soonBiology
Genetics, cell biology and protein structure, at research depth.
06
Coming soonMedicine
Clinical reasoning on hard cases: diagnosis, treatment, evidence.
Where we aim
Too easy teaches nothing. Too hard teaches nothing.
Models learn at the edge of what they can almost do. We build right at that edge, and keep moving it out as models improve.
Watch it learn
Lost at first. Then it learns.
A real learner, running live. Every try gets a score, and what scores well gets repeated. Swap the maze for real work, and you have our environments.
How we work
Checked.
If we can’t verify it, it doesn’t ship.
New.
Made from scratch, not scraped.
Yours.
Exclusive on request: once a set is yours, it’s never resold or reused.
Let’s build what your model learns next.
Tell us where your model gets stuck. We’ll show you tasks that fit.