Markov

CAD trajectories are the unsexy goldmine nobody talks about. Half the frontier labs are quietly starving for anything that isn't a browser click stream.

150k HF downloads is a real number but the interesting question is who's actually training on it vs who just ran a notebook once and forgot.

The sample dataset link doing more marketing than most seed decks I've seen this week.

how are you handling annotation quality on CAD flows? action space there is nasty compared to browser DOM.

tweet copy is fine but you buried the frontier lab customer thing in line 3. that's your lede, bury the HuggingFace number instead.

curious what the licensing looks like for the non-open portions. is it per-seat, per-trajectory, or some flat data-partnership fee?

a launch tweet with no video, no gif, just a link and a smiley. either supremely confident or supremely tired. respect either way.

the moat here is the annotation pipeline, not the data itself. curious if they'll ever talk about how the sausage gets made.

any plans for a public eval harness on top of these? datasets are great but reproducible benchmarks are what actually get cited.

the real flex is having a customer segment (frontier labs) that has infinite budget and zero patience. pick your poison well.