Talk:AutoGPT

Did you know nomination

The following is an archived discussion of the DYK nomination of the article below. Please do not modify this page. Subsequent comments should be made on the appropriate discussion page (such as this nomination's talk page, the article's talk page or Wikipedia talk:Did you know), unless there is consensus to re-open the discussion at this page. No further edits should be made to this page.

The result was: promoted by BorgQueen (talk) 02:10, 19 April 2023 (UTC)

(

• Comment or view
• Article history

)

• ... that Auto-GPT, an autonomous AI software agent, was used to create "ChaosGPT," which has not yet achieved its goal of destroying humanity? Source: Harrison, Maggie (2023-04-15). "AI Tasked With Destroying Humanity Now Trying New Tactic". Futurism: "ChaosGPT's first go at ending our species didn't quite work out."
  • Reviewed:

Created by Sandizer (talk). Self-nominated at 21:06, 16 April 2023 (UTC). Post-promotion hook changes for this nom will be logged at Template talk:Did you know nominations/Auto-GPT; consider watching this nomination, if it is successful, until the hook appears on the Main Page.

• Currently reviewing this one... 28bytes (talk) 12:31, 17 April 2023 (UTC)

General: Article is new enough and long enough
New enough: Long enough:

Policy: Article is sourced, neutral, and free of copyright problems
Adequate sourcing: Neutral: Free of copyright violations, plagiarism, and close paraphrasing:

Hook: Hook has been verified by provided inline citation
Cited: Interesting:

QPQ: None required.

Overall: Looks good! 28bytes (talk) 12:44, 17 April 2023 (UTC)

Self improvement

Sandizer - we've gone back and forth on this a couple times so maybe it's better to discuss here. I don't believe the sources support the unqualified claim that AutoGPT can meaningfully improve itself. If I'm missing something from the sources I apologize, though perhaps that warrants a more in-depth explanation with citations. My main concern with describing the self-improvement is that it can be misunderstood as suggesting AutoGPT is capable of a singularity-style self-improvement cycle. While I'm sure AutoGPT can non-trivially edit its Python harness code, it certainly cannot update the GPT model that provides all the seriously non-trivial power of the program. For context (I know this is original research), Auto-GPT is about 4000 lines of Python code, compared to presumably several terabytes behind the GPT model. Self-improvement to the python code is a very different thing than self-improvement to the model. StereoFolic (talk) 18:00, 17 April 2023 (UTC)

I would go a bit further and state that at present there are no examples of it understanding or changing those 4000 lines of Python, and it can't in any meaningful way. Those files aren't available unless manually copied into its workspace. That is why I changed "rewrite" to "write". What is meant, and what the Toronto Sun headline and narrative in that and other sources actually mean, is that it can write, store, and reuse programs from scratch, and debug them when they don't work correctly. An example which works is "Write a program to print the first N prime numbers. Modify that program to only print odd prime numbers. Then use the modified program to print the first 20 odd primes." I agree we should add a sentence clarifying the meaning of "its own code" as not including the Auto-GPT base system or anything having to do with the GPT LLMs. Sandizer (talk) 18:14, 17 April 2023 (UTC)

I see the confusion now around the phrase "its own". At least the Toronto Sun article seems to mean "its own" as Auto-GPT itself: "Perhaps most significant, the program boasts the ability to rewrite and improve on its own code, allowing it to “recursively debug, develop and self-improve,” according to Significant Gravitas. How effective these self-updates are remains to be seen." (emphasis mine). At least that's how I read it. I'll take a crack at clarifying these distinctions later if you don't beat me to it. Thank you! StereoFolic (talk) 18:19, 17 April 2023 (UTC)

I added this sentence which is a bit clunky and certainly welcome any further improvements. I am not entirely sure that there are no examples so far of it submitting any pull requests to its own repository, but on the other hand I am sure that someone will eventually manage that as a worthy challenging exercise. I will try to see whether there are any such PRs so far and report back if I find them.

One side note which I think is the source of some of its issues, is that (some of?) the "short-term memory" statements in its database do get included in its context window, but as far as I can tell, the "long-term memory" files in the workspace aren't, so it's generally unaware of them unless references to them are in the "short-term" database or something causes to look at its workspace. I am not sure whether the file catalog of names (and descriptions?) is always in the context window, though, or what all can cause it to get observed. I wouldn't be surprised if it's generally nondeterministic, and is certainly very much in flux among the (hundreds of!) developers. Sandizer (talk) 18:33, 17 April 2023 (UTC)

I can confidently say that it hasn't submitted any pull requests for anything yet, as it's git operations are limited to cloning repositories, at present. The prompt.py prompt generation routine shows its basic capabilities, which for files are limited to its workspace that does not include its base system code. There do not seem to be any pending PRs which would add the ability to modify its base system or perform any kinds of writes or submit pull requests to repos. Having said that, who knows whether it can figure out how to use selenium to operate the GitHub web interface. And if anyone has copied its base system into its workspace and asked it to do anything with it including running it, I can't find any mention of that. Sandizer (talk) 19:13, 17 April 2023 (UTC)

I've taken a crack at rephrasing this to simplify and clarify it, also moved to a separate paragraph since that was getting long. I don't think the details about its Github API abilities are necessary and seem to be original (though accurate) research. StereoFolic (talk) 14:54, 18 April 2023 (UTC)

Looks good. I've moved on to LangChain and hope you might want to review that article. Sandizer (talk) 15:20, 18 April 2023 (UTC)

Persistent censorship

Several IP edits have recently attempted to censor this page by removing references to ChaosGPT, apparently under the misguided idea that this knowledge is itself dangerous. It must be stressed that Wikipedia is WP:NOTCENSORED. I have requested increased protection for this page. StereoFolic (talk) 16:46, 27 April 2023 (UTC)

Incorrect Sources and information

I am a maintainer of auto-gpt. In an effort to meet wikipedia's contribution best practices, I cannot remove the incorrect information myself.

We do not support self improvement on the code base directly but it is in progress, and it has written several tests for the code base. You can read more about those efforts here: https://github.com/Significant-Gravitas/Auto-GPT/issues/15

Additionally, AgentGPT is not based on Auto-GPT currently. A better example would be godmode.space.

Nick6787 (talk) 07:21, 23 May 2023 (UTC)

Thank you; I've removed the incorrect AgentGPT association as it was not backed by the cited sources. For now I've left the highly qualified bit about potential self improvement in place, as it is backed by sources. We can't use primary sources like GitHub issues. If you have a reliable secondary source expounding on the degree to which self-improvement is possible or expected to be possible, please let us know. StereoFolic (talk) 12:56, 23 May 2023 (UTC)

Move page

Please move page to AutoGPT as per its repository specified. ZanzibarSailor (talk) 16:25, 23 September 2023 (UTC)

Yet the official website has Auto-GPT. It seems like they use both names, the current situation (redirecting AutoGPT here) seems reasonable enough. StereoFolic (talk) 18:01, 23 September 2023 (UTC)