3 ms·
Vision LLM is really remarkable Had a project that involved describing and cataloging over 20,000 images. Traditional method using real people would take mont
by spaceman_2020 2y ago
Vision LLM is really remarkable
Had a project that involved describing and cataloging over 20,000 images.
Traditional method using real people would take months and crap load of money (the descriptions have to be customer-readable)
OpenAI’s vision API does it for cents per image. Must have spent under $200 for the whole thing
- forrestthewoods 2y agoDid you have humans verify OpenAI’s descriptions were accurate?
- lolinder 2y agoFor a lot of applications it doesn't matter—you can check the descriptions as you use them as long as everyone knows they may be flawed, and checking 20k descriptions at point of use is still orders of magnitude easier than writing them from scratch.
- atrettel 2y agoI suppose at that scale you may not have to worry as much as errors like this, but I still think the point that there needs to be a human in the supply chain shows how much of AI is in fact just laundering of human intelligence. There's even an XKCD on this [1]! [1] https://xkcd.com/1897/ https://xkcd.com/1897/
- spaceman_2020 2y agoThey were far better and more accurate than the $6/hr VAs from Philippines we were using earlier. They were perhaps not as 100% accurate as someone spending half an hour writing out each description, but for our use case, the level of accuracy we were able to achieve with vision LLMs was good enough.