Having a folder where you can put txt files and autoGPT will eventually review them / work on them step by step. (ARV.) #720
GoMightyAlgorythmGo
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
We could have different folders like:
Suggestions
Issues
Ideas
Observations
useful_tools
There could be some logic that he uses or he can change or he can fall back to if things go wrong or the user can edit. Maybe different workflows or using numbers to go to simple tasks. 1 do super simple step we know you can do -> 2. do super simple step -> write number to file to make sure you did them -> 3 etc.
This way even a bad language model or once that is constantly interrupted by crashes, errors, userfeedback and shutdowns can make lasting progress in anything. :) Even writing a book or, coding, researching for a metalerning system, making a app, earning money, or other things can use organization and smart strategy's. This way the Users intelligence and massive overview (gpt 10 so to say) can also be integrated/used.
The usefulness of this is autoGPT can experiment and the user can see what he is doing. For example if AutoGPT is having a txt file with "best_restaurant_list.txt" and it lists good restaurants but you told him to code something. Then you might know he might be confused or you need to change a promt or talk to autoGPT. Also this way he can get past his small working memory since he might be able to write stuff down but to effectively remember it is also a challenge. So just following dumb easy systems like 1 complete then go to step a1 complete now go to step 3.2.5 and so on. This way he can also work around complex tasks even if he uses free/worse language models that can be run for free on a personal pc to save OpenAI token cost. (even if costs are okay it still saves lots of costs and can be made to run autoGPT in free mode to experiment/test things or when you have more time. :)
And then you can make a txt and list some things and autoGPT might be able to use them, example:
-Use textbased workflowcharts (example: Do this -> then do THAT if a then do ->THIS if b then do -> THAT etc.) He can name the workflow chart "saving userfeedback.txt" or "after_crash_remembering.txt" and kinda make them activate when he detects sertain things. Like lets say a crash happens. He goes trough the ususal agents crash happened? Y/ N agent: Y -> Executing "after_crash_remembering.txt" Content 1) review the "last task" txt. (that he always has to update when he starts a new task) -> review log.txt -> then -> then -> then -> if / else -> this or that -> end go back to number so and so or execute next workflow chart.
-I suggest you log what you do by with time and date so you can later use those list to review how long you worked and review effitiency
-use chains of agents to make numberical plans and then let the next agent refine it and so on. Example: 1.0 Research things 2.0 code things 3.0 Organisation -> next agent 1.1 make a new agent and ask to research, 1.2 google things, 1.3 explore capabilitys etc. then the next agent -> 1.2.1 document the results and sort them by topic, 1.2.1 check if previous searches have already been made to avoid repeating the same search -> 1.2.1.1 use a [] list of keywords to ruthly guess if a search is novel ... usw
Have 1 agent review that the steps dont go into other areas example: Agent 1.0 + 1.2 + 1.2.1 + 1.2.1.1 is this still in the scope of 1,.0 "research"? usw
-try to use .py code to read txt if you cant access them otherwise
-use little programs' to reduce token usage
-document metrics about the project
-document if you reach milestones if you cant reach them break them down or use smaller milestones
Stuff like that (not very organized or cohesive but it is just a example) also numbering might help 1, 2, 3
Beta Was this translation helpful? Give feedback.
All reactions