Constant looping and no execution

[blksqd]

Hey guys, I#ve been running prompt experiments for the past 4 days by modifying the prompts little by little and using different versions of Auto-GPT but I keep running into the same issue: The AI makes a plan and "promises" execution but then never happens. Even when it says in the console that the "action" was successfully executed. Any idea why this happens?

[monxas]

I have the exact same behavior. I had it running for plenty iterations, with very simple golas like create an html file with a title and a text about flowers. it never actually delivers

[MPLScyberguy]

Agreed! I've run 3-4 scenarios over the last 24hrs and most seemed to get pretty far. But then they get caught in a "loop" trying to locate a file, create a folder, add the file to the folder, execute the file... (and then repeat). ChatGPT suggested quitting the application (which I then lose - not sure why memory isn't working) and they suggested a prompt to stop it, I've tried correcting the action by human feedback to the next step, but the program crashes every time I try anything to get out of the "loop".

If anyone can figure this out please reply. If I do, I will update this as well. Thx!

[blksqd]

I ran a few more tests by creating a custom agent to feed the information into the loop but I keep running into the same issue. The latest version solved the problem with the JSON bracket that was showing a constant error but in that version the responses and overall chain seems to be shorter and not so deep as the previosu version with the error...

I love where the project is going but we are obviously a bit far from something reliable.

[sharpwang]

I found the same problem. Auto-GPT will fall into a loop and repeat the steps previously performed.

[bjornfix]

I think I'm getting a basic understanding about why this happens through my experiments. It's called hallucinations as far as I have understood. To prevent them, you need to guide the AI better. It's the only way to prevent this behaviour as far as I can understand.

As soon as you understand the following, much more will become clear to you.

You are not dealing with a regular piece of software that you just give commands to. You are dealing with something that attempts to behave like a human would. It makes mistakes and it tries to understand what went wrong.

You give guidance to the AI. It tries to understand what are your intentions. Using the roles to clearly define what the AI is supposed to be is important. Defining precise goals is important. And defining what the AI is not allowed to do is important as well.

Now, I might be totally wrong in my assumptions here, but as soon as I stopped thinking about this like a regular piece of software, things started progressing in a more fruitful way.

As of now, I still don't have a working bot as I am hoping to make, but I can see from the dialogue between the bot and it's agents that I am getting closer.

P. S. I have seen some clear proof of it trying to learn when it changes behavior after a number of attempts to repeat the looping behavior. I have also seen it depends very much on the instructions given in the goals. I use the last goal to give it a way out if everything else fails. Like, if you get stuck, I want you to do this.