-
Notifications
You must be signed in to change notification settings - Fork 43.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
still looping. #3855
Comments
me too loops a lot |
Me three. I was optimistic when I saw them claime V3 worked reliably, but I found it just as hopeless as previous versions in doing a basic search for information and writing it to a text file. It overcomplicates everything, gets into a loop about specifics that are unrelated, and eventually errors out. |
Do you use redis? Does it keeps using the same index? BTW, my pull request #3914 deals with looping it by stopping when it happens (the user can direct it to do something else). |
No, it's not just you. I will +1 that they royally screwed up this project somehow. I've been messing around with it for a little over a month now, and in my experience the best performing version by far has been 0.2.2 + unsandboxing the file_operations.py. With that version combined with a reddit user named stunspot's workflow, I was able to get it to research and code its own modules that actually functioned (it just couldn't execute them by itself; they had to be ran manually). You can see my thread on it here: https://www.reddit.com/r/AutoGPT/comments/12ohmmo/uhhhits_a_little_terrifying_how_easy_it_was_to/ Now? It does nothing but loop endlessly on research. I just gave it 10 prompts (y -10), right? The given task was to make a disposable vape vending machine business plan. Not only did it just go to 5 webpages and find nothing of interest...but it then went to those same webpages AGAIN, completely wasting a bunch of tokens. It routinely does this now for some reason, yet the 0.2.2 version with a modified file_operations.py never did that. It had its bugs, sure, such as its inability to execute python as I stated earlier. But in terms of having an efficient workflow + actually getting stuff done, it's no contest. Here's hoping they course correct for the next release, because until then I'll be sticking with 0.2.2. I'm using GPT-4, btw. --gpt4only |
Agreed. All it does is seems to endlessly generate plans for doing various
abstract stuff, 90% of which is unrealistic at present. It's like a
government, but worse!
Someone leading this project needs to get 5 of their best developers who
are dedicated to the project, and get them focused on some basic use-cases
which are practical, useful, and not too abstract, such as...
Here's 3 off the top of my head...
1. Creating a basic website based on a supplied topic that could be
uploaded via FTP and just work. Simple HTML/CSS/JS with a form, a topic,
some generated images, that's all.
2. Taking a topic (from human, some keywords, sample data) then have it go
out and researching it, scrape some data (eg. ebay product titles,
descriptions, meta), interpret some data (eg. articles, wikipedia) then
generate an output with some columns in CSV format.
3. Generate a novel with 10 chapters, first establishing beginning, middle,
end, then fleshing it out, generating some images, etc.
…On Sun, May 7, 2023 at 4:26 PM Mitman93 ***@***.***> wrote:
No, it's not just you. I will +1 that they royally screwed up this project
somehow. I've been messing around with it for a little over a month now,
and in my experience the best performing version by far has been 0.2.2 +
unsandboxing the file_operations.py. With that version combined with a
reddit user named stunspot's workflow, I was able to get it to research and
code its own modules that actually functioned (it just couldn't execute
them by itself; they had to be ran manually). You can see my thread on it
here:
https://www.reddit.com/r/AutoGPT/comments/12ohmmo/uhhhits_a_little_terrifying_how_easy_it_was_to/
Now? It does nothing but loop endlessly on research. I just gave it 10
prompts (y -10), right? The given task was to make a disposable vape
vending machine business plan. Not only did it just go to 5 webpages and
find nothing of interest...but it then went to those same webpages AGAIN,
completely wasting a bunch of tokens. It routinely does this now for some
reason, yet the 0.2.2 version with a modified file_operations.py never did
that. It had its bugs, sure, such as its inability to execute python as I
stated earlier. But in terms of having an efficient workflow + actually
getting stuff done, it's no contest. Here's hoping they course correct for
the next release, because until then I'll be sticking with 0.2.2.
I'm using GPT-4, btw. --gpt4only
—
Reply to this email directly, view it on GitHub
<#3855 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AEYVIAUTIWG242OFPT3S573XE46BJANCNFSM6AAAAAAXXLSG7I>
.
You are receiving this because you commented.Message ID:
***@***.***>
|
I believe I mispoke on the version #, btw. It was 0.2.0 I had tremendous success with, not 0.2.2. Whatever the version before they introduced DevTools for searching instead of just using the "browse_website" command. I just confirmed it. Check this out. This is the worthless output I got from 0.3.0: Using Browser: chrome
DevTools listening on ws://127.0.0.1:61724/devtools/browser/443d8fc4-ecec-4d8e-ae96-1d2d8430a5c5
DevTools listening on ws://127.0.0.1:62192/devtools/browser/d90fd7aa-1359-45d4-9f76-066dd4352a65
DevTools listening on ws://127.0.0.1:62519/devtools/browser/19ffd3a4-8d29-445d-b7bc-32b465656364
Not going to paste the whole thing because it's pretty long, but yeah. It basically just visited 5 websites, says "The previous searches did not provide specific information on disposable vape vending machines. Trying a different approach may yield better results."...AND THEN IT DOES THE SAME EXACT THING AGAIN. IT EVEN REVISITS THE SAME 5 WEBSITES. Now, you ready for this? Here's that older version I was talking about. Same exact prompt. Same exact given goals. Same everything, but the version: NEXT ACTION: COMMAND = browse_website ARGUMENTS = {'url': 'https://www.harvin.eu/blog-news/disposabe-vape-vending-machines/', 'question': 'What is the business overview of disposable vape vending machine businesses?'}
Again, not going to post the entire thing because it's super long, but let me just say...WOW. It's already successfully researched several areas of the business, drafted preliminary proposals, and is currently looking for reputable suppliers. In other words: IT'S ACTUALLY DOING WHAT I TOLD IT TO DO! IT'S ACTUALLY WORKING!! Woohoo! Whenever they stopped using the "browse_website" command in favor of that Devtools thing...that appears to be what is causing all the search loop headaches. |
If you can, you should consider sharing some of the settings that work/don't work for you for debugging purposes - if you are concerned about sharing "secrets", you can obviously change the topic/keywords of your research. I am primarily interested in the underlying workflow to see what can be improved to get it stop unnecessary looping, and to work out a scheme for tasking/workspace recovery (continuation) that works. |
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. |
This issue was closed automatically because it has been stale for 10 days with no activity. |
Which Operating System are you using?
MacOS
Which version of Auto-GPT are you using?
Latest Release
GPT-3 or GPT-4?
GPT-4
Steps to reproduce 🕹
scrape google for refernces to xxxx and creat a text document called xxxx, update it as you find results
Current behavior 😯
loops endlessly checking the itegrity of the reference document, doesn't stre any info and loooooooooooops
Expected behavior 🤔
a text doc with info on it about subject
Your prompt 📝
No response
Your Logs 📒
No response
The text was updated successfully, but these errors were encountered: