Generalist
GeneralistRemoteTask-basedPosted Nov 11
Paid per taskThis opportunity lists pay per task, rather than an hourly rate.
What the work looks like
AI models that read images are often bad at counting, and they get worse when counting takes reasoning. You'll build tasks that expose that weakness:
- Taking your own original, unedited photos of real-world scenes, mainly outdoor ones (streets, parking lots, building facades, parks, markets) that hold many countable objects
- Writing 3–5 counting prompts per image that make the model reason, not just count: conditional filters ("How many red cars are parked in the lot?"), purpose-based reasoning ("How many objects require seatbelts to operate?"), spatial relationships ("How many cars are parked next to an empty space?"), telling similar objects apart, and prompts whose correct answer is 0
- Making sure every prompt has exactly one correct answer that 9 out of 10 people would agree on right away, then checking and recording that answer
- Outlining (masking) every counted object in an annotation tool so each answer can be checked object by object
- Aiming for image and prompt sets where the model gets the count wrong
Pay is per task, based on how many objects your hardest prompt counts: $60 (0–24 objects), $70 (25–49), $80 (50–79), $100 (80+).
What this covers
- Original photography of natural, unedited scenes (no stock, social media, AI-generated, or edited images)
