Force the model to complete a plan and not do_nothing

[Captain-Bacon]

i'm using 3.5 as the model (no 4.0 API key yet!)

I'm constantly finding that the model gives itself a plan, and then doesn't execute the commands - it reverts to do_nothing. Now, I understand that sometimes it will need to do that to 'think', but it's about 90% of the output for me at the moment. It also forgets which commands to use.

To make matters worse it thinks it has completed the commands, and then we can't get back on track at all, or perhaps just as bad it reaches some local cache limit and clears the entire memory, goes utterly blank, and hasn't committed anything to longer term memory because it didn't reach an 'important' point to remember because it was mucking around.

It dilly-dallies around for about 50 prompts when I could ask ChatGPT 3, let alone 3.5 or 4 to write me a code snippet and it does it instantly - this is costing about 50,000 tokens every 10 minutes or so at the moment, and i'm getting absolutely nowhere!

Does anyone else have this problem?

[mylesmg]

I'm experiencing the same thing. It basically never finishes. I feel like there has to be an easy-to-implement, but genius, solution to this.

[Captain-Bacon]

it feels like there needs to be a UI or something where you can highlight a task and it then 'forces' that in prompt, or a way of injecting a prompt into the process without 'interrupting' as human feedback. Perhaps just having a 'negative' prompt that can be selected for a period of time that says something like 'stop testing, just ask the API what code there should be for x function and move on'.