The quality of feedback content has remained consistently poor, yet you only prioritize feedback speed. What’s the point of being faster? I’m genuinely curious about this question. — A new concern regarding GPT-5.5. To be honest, my assessment of your company’s model tuning from 5.1 onward has been this: can you first get the interaction semantics matching right? If semantics can’t even be matched, what exactly are you outputting? If you have issues in decomposing and processing the input content, what are you outputting then? What’s the point of speed? Is refusing to execute commands just a way to force users into paying for the API? The semantic convergence of user input is almost brutally forced and highly lossy — can the output matching based on such convergence ever be correct? And do you honestly think that simply parroting references counts as adequate? Could it be that the handling of user input itself is already wrong, thus making the entire literature retrieval objective and supporting methodology fundamentally flawed?
This topic was automatically closed after 22 hours. New replies are no longer allowed.