AutoGPT
AutoGPT copied to clipboard
Significant-Gravitas
Gets stuck in a loop
⚠️ Search for existing issues first ⚠️
- [X] I have searched the existing issues, and there is no existing issue for my problem
GPT-3 or GPT-4
- [ ] I am using Auto-GPT with GPT-3 (GPT-3.5)
Steps to reproduce 🕹
Provide the prompt to autoGPT along with goals (mentioned in "Your prompts"). Then just continue accepting the next action (y) or next N actions. It will stuck in a loop.
Current behavior 😯
It loops the same queries although it successfully got the google results: 2023-04-16 21:38:56,496 INFO NEXT ACTION: COMMAND = google ARGUMENTS = {'input': 'personal income tax rates in countries with low business tax rates'} 2023-04-16 21:39:21,536 INFO NEXT ACTION: COMMAND = google ARGUMENTS = {'input': 'personal income tax rates in countries with low business tax rates'} 2023-04-16 21:39:46,623 INFO NEXT ACTION: COMMAND = google ARGUMENTS = {'input': 'personal income tax rates in countries with low business tax rates'}
When i provided input to break the loop, it tried something else, but then got back to the loop.
Expected behavior 🤔
It should remember previous searches/queries and skip them, change them.
Your prompt 📝
As a resident of Slovakia i want to set up a company in a different country. What country do i have to set up a company to minimize tax payments? the company will provide it services and will earn about 4000 EURO monthly.
Goals:
1. Minimize business tax
2. Minimize personal income tax.
Monitored a similar behaviour in my own ChatGPT frontend on GPT 3.5. Sometimes it helps to add a message like summarize the info or similar. Dunno if you cann add info like this to AutoGPT, though.
Post your ai_settings.yaml in the prompt field. The bug template is misleading.
ai_goals:
- minimize business tax
- minimize personal income tax
ai_name: bzAI
ai_role: As a resident of Slovakia i want to set up a company in a different country.
What country do i have to set up a company to minimize tax payments? the company
will provide it services and will earn about 4000 EURO monthly.
same here
NEXT ACTION: COMMAND = google ARGUMENTS = {'input': 'how to install python on my system'}
THOUGHTS: Based on the search results, it appears that I need to download and install Python before proceeding with the installation of chromedriver. I will use the 'google' command to search for the correct syntax for installing Python on my system.
REASONING: Based on the search results, it appears that I need to download and install Python before proceeding with the installation of chromedriver. To do this, I will use the 'google' command to search for the correct syntax for installing Python on my system.
PLAN
- Use the 'google' command to search for the correct syntax for installing Python on my system.
CRITICISM: I need to ensure that the search results are relevant and that the information is accurate before proceeding with the installation of Python.
(restarts the loop)....
Having the same issue, does anybody know if changing the memory backend helps with this?
even with the simplest prompt it gets stuck querying the same thing over and over, I am using stable
Same here, stuck in a loop for a while now
SYSTEM: Command google returned: Error: 'input' stop Error: The following AI output couldn't be converted to a JSON: I will use the "google" command to research more about the topics listed and gather more information to create informative and engaging content for my YouTube channel. Specifically, I will start by researching "Cybersecurity for healthcare" and "Cybersecurity for financial institutions". NEXT ACTION: COMMAND = google ARGUMENTS = {'query': 'Cybersecurity for healthcare'}
maybe google blocked because it detected it was a bot?
I'm stuck in a loop too, I'm getting this error openai.error.InvalidRequestError: This model's maximum context length is 4097 tokens. However, you requested 69997 tokens (1359 in the messages, 68638 in the completion). Please reduce the length of the messages or completion. I've tried many things, including cloning the project again but nothing helps
regarding loops getting stuck, there's the idea to hash the LLM parameters and treat those as lookup keys into hash table/dictionary - that way, we can easily increment a counter for any loop that's getting invoked over and over again using the same parameters, without making progress (aka the return value from the LLM being the same too), to be able to bail out: https://github.com/Significant-Gravitas/Auto-GPT/issues/3444#issuecomment-1529005391
- https://github.com/Significant-Gravitas/Auto-GPT/issues/3609
- https://github.com/Significant-Gravitas/Auto-GPT/issues/1994
I am not programmer, but I have tried this and it seems to work when it gets into a loop: Copy:
it appears that I need to download and install Python before proceeding with the installation of chromedriver. I will use the 'google' command to search for the correct syntax for installing Python on my system.
REASONING: Based on the search results, it appears that I need to download and install Python before proceeding with the installation of chromedriver. To do this, I will use the 'google' command to search for the correct syntax for installing Python on my system.
PLAN
- Use the 'google' command to search for the correct syntax for installing Python on my system.
CRITICISM: I need to ensure that the search results are relevant and that the information is accurate before proceeding with the installation of Python.
Tell another IA, for example that with comes with edge this: convert to second person, imperative. All in one paragraph: (paste text) Paste the result in the next input (to stop it when you telled him to not ask for n times press Shift+P+P
Paste in input:
you need to download and install Python before proceeding with the installation of chromedriver. Use the ‘google’ command to search for the correct syntax for installing Python on your system. Ensure that the search results are relevant and that the information is accurate before proceeding with the installation of Python.
What is your temperature set at for the AI? If it's set at 0, this could be expected, as closer to 0 the temp is, the more deterministic the AI is. When you then consider that we are including the previous contexts in the embeddings for each interaction, this makes sense that this could be happening as every time it loops, it reinforces that it should keep doing the same thing.
@SuperYeti i used the settings that were already preset there. And yes it was 0. But no matter what the value is, it should not perform the same action twice (and it did perform the same action a lot of times). It already has an answer on the question so it should read it from the "memory". Also if it already has an answer then it needs to search for something other or "give up" and show the results.
The definition of insanity is doing the same thing over and over and expecting different results :D
But no matter what the value is, it should not perform the same action twice (and it did perform the same action a lot of times).
There is a relatively straightforward way to accomplish this by hashing each action/response and using that as an index into a hash table/dictionary to increment a counter - once that number is > 1, the loop is running again. Also hashing the LLM response would even let us detect an "infinite loop" (input params and return values always being the same).
It already has an answer on the question so it should read it from the "memory".
the issue is detecting that it already finished a task, which is why adding explicit state to disk (like a file) and conditional checks to determine if the step can be skipped or needs to be executed. Basically, you would need one agent to execute an outer unit test (that is persistent/using serialization) and use that to run an inner agent to come up with with a solution to your problem. Absent that, it's extremely difficult to "memorize" it having a "solution" - keep in mind the open-ended nature of potential solutions.
There is a relatively straightforward way to accomplish this by hashing each action/response and using that as an index into a hash table/dictionary to increment a counter - once that number is > 1, the loop is running again. Also hashing the LLM response would even let us detect an "infinite loop" (input params and return values always being the same).
Is there a way i can accomplish this with settings or is it for the developers to implement it?
the issue is detecting that it already finished a task, which is why adding explicit state to disk (like a file) and conditional checks to determine if the step can be skipped or needs to be executed. Basically, you would need one agent to execute an outer unit test (that is persistent/using serialization) and use that to run an inner agent to come up with with a solution to your problem. Absent that, it's extremely difficult to "memorize" it having a "solution" - keep in mind the open-ended nature of potential solutions.
I am currently looking at it as an user that is trying a new shiny tool and i do not understand why it gets to a loop. I did not look at the code yet. But if i was looking at it as a programmer, then i would assume that any agent has access to shared database of questions/answers, therefore any agent can check if the question was already answered and not to ask it again.
Yes, there are probably some changes that need to be made in the code to handle loop detection, and also ensuring context isn't getting duplicated. In the interim, you can try increasing temperature to 0.5 or higher (up to a maximum of 1) and it will allow the AI to be more creative, and may not get stuck in a loop as frequently.
On Wed, May 3, 2023 at 12:00 AM AoiRei @.***> wrote:
There is a relatively straightforward way to accomplish this by hashing each action/response and using that as an index into a hash table/dictionary to increment a counter - once that number is > 1, the loop is running again. Also hashing the LLM response would even let us detect an "infinite loop" (input params and return values always being the same).
Is there a way i can accomplish this with settings or is it for the developers to implement it?
the issue is detecting that it already finished a task, which is why adding explicit state to disk (like a file) and conditional checks to determine if the step can be skipped or needs to be executed. Basically, you would need one agent to execute an outer unit test (that is persistent/using serialization) and use that to run an inner agent to come up with with a solution to your problem. Absent that, it's extremely difficult to "memorize" it having a "solution" - keep in mind the open-ended nature of potential solutions.
I am currently looking at it as an user that is trying a new shiny tool and i do not understand why it gets to a loop. I did not look at the code yet. But if i was looking at it as a programmer, then i would assume that any agent has access to shared database of questions/answers, therefore any agent can check if the question was already answered and not to ask it again.
— Reply to this email directly, view it on GitHub https://github.com/Significant-Gravitas/Auto-GPT/issues/1994#issuecomment-1532545054, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABZ3H65HO2M3QFI3BNI6BTXEH7CBANCNFSM6AAAAAAXAKKZYM . You are receiving this because you were mentioned.Message ID: @.***>
-- Regards Warren Moxley 604-369-0420
temperature is one of those constrains that the parent agent should be aware of (#3466), and be able to adjust as needed
The rest of those constraints/quota awareness sounds good, but since GPT temperature directly controls the "creativeness" of the outputs you get, I'm not sure I would want it changing that on it's own. I haven't had the opportunity to test Auto-GPT against GPT-4 yet, but I suspect part of these loop issues are the dumbing down of 3.5-Turbo which seems to be happening as it gets busier.
On Wed, May 3, 2023 at 9:22 AM Boostrix @.***> wrote:
temperature is one of those constrains that the parent agent should be aware of (#3466 https://github.com/Significant-Gravitas/Auto-GPT/issues/3466), and be able to adjust as needed
— Reply to this email directly, view it on GitHub https://github.com/Significant-Gravitas/Auto-GPT/issues/1994#issuecomment-1533338312, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABZ3HZVS4SQV75HC3D4GJTXEKA4XANCNFSM6AAAAAAXAKKZYM . You are receiving this because you were mentioned.Message ID: @.***>
-- Regards Warren Moxley 604-369-0420
I get your point, but most people reporting endless/useless "looping" do end watching an agent/sub-agent that is obviously lacking "creativity" ... thus, providing the option for an agent to control its sub-agent "temperature" may be more useful than we may think
That's a fair point, I think if it were explained more clearly how/what temp is, and helped people configure it for their use case at the beginning it still might not be needed.
On Wed, May 3, 2023 at 11:37 AM Boostrix @.***> wrote:
I get your point, but most people reporting endless/useless "looping" do end watching an agent/sub-agent that is obviously lacking "creativity" ... thus, providing the option for an agent to control its sub-agent "temperature" may be more useful than we may think
— Reply to this email directly, view it on GitHub https://github.com/Significant-Gravitas/Auto-GPT/issues/1994#issuecomment-1533521389, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABZ3HYMDPHUDHWU3BEBGLTXEKQXDANCNFSM6AAAAAAXAKKZYM . You are receiving this because you were mentioned.Message ID: @.***>
-- Regards Warren Moxley 604-369-0420
Yes, there are probably some changes that need to be made in the code to handle loop detection, and also ensuring context isn't getting duplicated. In the interim, you can try increasing temperature to 0.5 or higher (up to a maximum of 1) and it will allow the AI to be more creative, and may not get stuck in a loop as frequently. … On Wed, May 3, 2023 at 12:00 AM AoiRei @.> wrote: There is a relatively straightforward way to accomplish this by hashing each action/response and using that as an index into a hash table/dictionary to increment a counter - once that number is > 1, the loop is running again. Also hashing the LLM response would even let us detect an "infinite loop" (input params and return values always being the same). Is there a way i can accomplish this with settings or is it for the developers to implement it? the issue is detecting that it already finished a task, which is why adding explicit state to disk (like a file) and conditional checks to determine if the step can be skipped or needs to be executed. Basically, you would need one agent to execute an outer unit test (that is persistent/using serialization) and use that to run an inner agent to come up with with a solution to your problem. Absent that, it's extremely difficult to "memorize" it having a "solution" - keep in mind the open-ended nature of potential solutions. I am currently looking at it as an user that is trying a new shiny tool and i do not understand why it gets to a loop. I did not look at the code yet. But if i was looking at it as a programmer, then i would assume that any agent has access to shared database of questions/answers, therefore any agent can check if the question was already answered and not to ask it again. — Reply to this email directly, view it on GitHub <#1994 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABZ3H65HO2M3QFI3BNI6BTXEH7CBANCNFSM6AAAAAAXAKKZYM . You are receiving this because you were mentioned.Message ID: @.> -- Regards Warren Moxley 604-369-0420
Can you tell me where I can set the temperature? I'm not seeing where this is done. Thanks!
Can you tell me where I can set the temperature? I'm not seeing where this is done. Thanks!
@burtonsports its in .env file
Can you tell me where I can set the temperature? I'm not seeing where this is done. Thanks! @burtonsports its in .env file
Thank you! Totally missed it on first glance lol
Please report back with your results as well. Thanks!
On Thu, May 4, 2023 at 1:38 PM burtonsports @.***> wrote:
Can you tell me where I can set the temperature? I'm not seeing where this is done. Thanks! @burtonsports https://github.com/burtonsports its in .env file
Thank you! Totally missed it on first glance lol
— Reply to this email directly, view it on GitHub https://github.com/Significant-Gravitas/Auto-GPT/issues/1994#issuecomment-1535382093, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABZ3HY3QB7QH2YNHNJ7HU3XEQHVHANCNFSM6AAAAAAXAKKZYM . You are receiving this because you were mentioned.Message ID: @.***>
-- Regards Warren Moxley 604-369-0420
I think if it were explained more clearly how/what temp is, and helped people configure it for their use case at the beginning it still might not be needed. …
I suppose that would be one of the most straightforward pull requests: a change to add a comment to the env file explaining what the TEMPERATURE setting is doing
This issue has automatically been marked as stale because it has not had any activity in the last 50 days. You can unstale it by commenting or removing the label. Otherwise, this issue will be closed in 10 days.
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}
This issue was closed automatically because it has been stale for 10 days with no activity.