From time to time (I estimate around 1-2% of attempts) ChatGPT fails to complete a response. It hangs. This is relatively easy to trigger on bad connections but also occurs on good connections. The longer the response the bigger the risk.
When going from one conversation to the other in the left bar, the conversation is still unfinished in the tab where it hangs. The only way to fix this is to regenerate the response, which often triggers a very different response.
After this response is done, it asks if the new response is better or worse than the old one, and it doesn’t seem to know that its initial response was unfinished. This may or may not introduce a bit of static into the system if users report that the initial response was “worse” if the system isn’t aware that the initial response was cut off.
It also seems like having to regenerate a response causes unnecessary use of resources. So maybe this is something worth to take a look at?
This may be browser-related, it seems when we’re using the API directly it’s much rarer. Though for the API the ability to do an instant retry to get a cached result would be great as well, esp. if it can be combined. Impacts a low number of requests but can improve UX in these cases.