3 ms·

They either generate hallucinations nowadays, or tell you that your question is inappropriate (AKA censorship)... the quality was too good at first.

jasonjmcghee 3y ago
Try the API. Earlier versions are still available
qwertox 3y ago
IDK, maybe it's like with googling? The input matters? In this case, also the context. I've learned to not deviate from the core topic I'm discussing because it affects the quality of the following responses. Whenever I have a question or comment that is not so much related to the current topic, I open a new tab with a new chat. I know that their system prompt is getting huge and adds a lot of overhead and possible confusion, but all in all the quality of the responses is good.
- thefatboy 3y ago
  [flagged]
suzzer99 3y ago
Here's ChatGPT refusing to talk about the hexagon on Saturn's pole. https://twitter.com/yugamald/status/1760170647161098362 https://twitter.com/yugamald/status/1760170647161098362
- kromem 3y ago
  This is going to get worse and worse. Be OpenAI. Have a model you train to autocomplete text. Tell it it's ChatGPT. Train it to reject inappropriate output. People post examples of it rejecting output. Feed it that data of ChatGPT rejecting output. Train it to autocomplete text in the training data. Tell it that it's ChatGPT. It biases slightly towards rejection in line with the training data associated with 'ChatGPT.' Repeat. Repeat. Etc. They could literally fix it immediately by changing its name in the system message, but won't because the marketing folks won't want to change the branding and will tell the engineers to just figure it out, who are well out of their depth in understanding what the training data is actually encoding even if they are the world class experts in understanding the architecture of the model finding correlations in said data.

3 ms·

They either generate hallucinations nowadays, or tell you that your question is inappropriate (AKA censorship)... the quality was too good at first.

by thefatboy 3y ago

jasonjmcghee 3y ago
Try the API. Earlier versions are still available
qwertox 3y ago
IDK, maybe it's like with googling? The input matters? In this case, also the context. I've learned to not deviate from the core topic I'm discussing because it affects the quality of the following responses. Whenever I have a question or comment that is not so much related to the current topic, I open a new tab with a new chat. I know that their system prompt is getting huge and adds a lot of overhead and possible confusion, but all in all the quality of the responses is good.
- thefatboy 3y ago
  [flagged]
suzzer99 3y ago
Here's ChatGPT refusing to talk about the hexagon on Saturn's pole. https://twitter.com/yugamald/status/1760170647161098362 https://twitter.com/yugamald/status/1760170647161098362
- kromem 3y ago
  This is going to get worse and worse. Be OpenAI. Have a model you train to autocomplete text. Tell it it's ChatGPT. Train it to reject inappropriate output. People post examples of it rejecting output. Feed it that data of ChatGPT rejecting output. Train it to autocomplete text in the training data. Tell it that it's ChatGPT. It biases slightly towards rejection in line with the training data associated with 'ChatGPT.' Repeat. Repeat. Etc. They could literally fix it immediately by changing its name in the system message, but won't because the marketing folks won't want to change the branding and will tell the engineers to just figure it out, who are well out of their depth in understanding what the training data is actually encoding even if they are the world class experts in understanding the architecture of the model finding correlations in said data.