September 7 at 8AM, a new model fine-tune or retrain or reparameterization hit API gpt-3.5-turbo-0613, causing major cognitive loss when following system programming instructions, accompanied by a significant increase in token production rate to let us know it was an architectural change. The AI couldn’t even write four-word summaries for ChatGPT titles without errors. And then the model was hit again with further stupidity with more developer apps not working a week ago.
The result is you can tell it to write simply a tweet with three parts and get nothing like what you asked for (and because I could demonstrate 0301 actually working in that post, they had to mess with that model also).
Today below. Did I ask for two tweets from gpt-3.5-turbo? No.
The go back to the quality I showed September 11 from 0301 in that post on multiple calls. Now 0301 also doesn’t follow the multi-part output instruction - and makes it even worse:
The only way you’ll likely succeed is in putting an instruction (not a rule or programming that you need followed) that the AI would follow from a user into a user role. And what can you do with system? “You are ChatGPT”.
API users not paying for GPT-4 are obviously OpenAI’s adversary.