pitermach,

First we had Mac get a GPT 4 image describing utility, now it’s NVDA’s turn. https://github.com/cartertemm/AI-content-describer/

jcsteh,

@pitermach Neat! Do I understand correctly that this doesn't allow you to ask questions, but the underlying API does? Or is the follow-up question part not exposed to the public API yet? Or perhaps you have no idea about any of this and I'm asking the wrong person? :)

jscholes,
@jscholes@dragonscave.space avatar

@jcsteh As far as I know, the API is stateless, so only implicitly supports follow-ups. I.e., if you want to ask a question about previous responses, you have to resend that history with the request, decorated with which party said what. This add-on doesn't support asking follow-up questions regardless. I suppose you could modify the initial prompt in the settings to ask something specific before the image had been recognised, or after the initial recognition before resending it. @pitermach

jcsteh,

@jscholes Interesting. I wonder if Be My AI does this or has some API that isn't publicly available yet. @pitermach

jscholes,
@jscholes@dragonscave.space avatar

@jcsteh From what I've seen, this is one of the most common questions on the OpenAI dev forums and other Q&A sites by far; people expect it to work like the web version of ChatGPT out of the box. So on one hand, statefulness would be a popular feature. On the other, it could decrease token usage and hence API revenue, or make token utilisation less predictable. There could also be questions about how long they store e.g. cached images within a session, although I think the data usage ship has well and truly sailed on this one. @pitermach

miki,
@miki@dragonscave.space avatar

@jscholes @jcsteh @pitermach All OpenAI APIs work this way, including Chat GPT. There's no state, you always send the entire conversation history. This is most likely what web Chat GPT (or the Chat GPT backend) does under the hood. Even if there was state, it would be an abstraction at best with little to no impact on token usage. You just have to feed the whole conversation history to the model every time to make it do anything useful, particularly in such a heavily multi-user setup. That's how transformers function.

jamminjerry,

@pitermach I really want to try this open ai thing for NVDA, but the open ai site isn't letting me add a payment method! it keeps saying you are making too many attempts. slow down. even the first time I tried to hit continue I got that stupid message! grrrrrrrrrr!

Jonathan,

@pitermach Do you have to pay $1 for the first three months, and then $5 every month after that?

Brynify,
@Brynify@dragonscave.space avatar

@pitermach LOL what? I'm sorry, I cannot provide visual descriptions or details about images.

  • All
  • Subscribed
  • Moderated
  • Favorites
  • random
  • Durango
  • DreamBathrooms
  • thenastyranch
  • magazineikmin
  • khanakhh
  • InstantRegret
  • Youngstown
  • ngwrru68w68
  • slotface
  • rosin
  • tacticalgear
  • mdbf
  • kavyap
  • modclub
  • megavids
  • osvaldo12
  • ethstaker
  • cubers
  • normalnudes
  • everett
  • tester
  • GTA5RPClips
  • Leos
  • cisconetworking
  • provamag3
  • anitta
  • JUstTest
  • lostlight
  • All magazines