Y'know, this would be a great project for Mozilla, if they have the resources for it. They're already doing that crowdsourced voice training data thing.

I bet CloudFlare has the bandwidth for it.

And yet they use reCaptcha. I hit it all the time accessing sites when I was in France.

As if IP addresses are not shared. IPv6 isn't present enough.

and you don’t think that will be quickly used to make a tool to break the captchas?

No. You cycle through new types of challenges that the model can't solve yet.

For example, one week of stop-sign recognition, another week of pedestrian labeling, etc.

Building those models sounds like a lot of continuous and boring work, and is probably best done by a company. Unless you find some crowd-sourced way of building those models.

Exactly, see the other comment in response to mine. If Mozilla can do this for speech recognition, then we can do this for image recognition/transcription tasks.