I run a YouTube operation built on high-quality, screen-recorded software tutorials. We’ve produced 75k videos (2–5 min each) in a couple of months using a trained team of 20 operators. The business is profitable, and the production pipeline is consistent, cheap and scalable.
However, I’m considering whether what we’ve built is more valuable as AI agent training/evaluation data. Beyond videos, we can reliably produce:
- Human demonstrations of web tasks
- Event logs, (click/type/url/timing, JSONL) and replay scripts (e.g Playwright)
- Evaluation runs, (pass/fail, action scoring, error taxonomy)
- Preference labels with rationales (RLAIF/RLHF)
- PII-safe/redacted outputs with QA metrics
I’m looking for some validation from anyone in the industry:
1. Is large-scale human web-task data (video + structured logs) actually useful for training or benchmarking browser/agent systems?
2. What formats/metadata are most useful (schemas, DOM cues, screenshots, replays, rationales)?
3. Do teams prefer custom task generation on demand or curated non-exclusive corpora?
4. Is there any demand for this? If so any recommendations of where to start? (I think i have a decent idea about this)
Im trying to decide whether to formalise this into a structured data/eval offering. Technical, candid feedback is much appreciated!