Will there be a fully self-sustaining Auto-GPT agent in 2023?
167
564
2K
resolved Jan 1
Resolved
NO

Context:

The Github repo says:

This program, driven by GPT-4, autonomously develops and manages businesses to increase net worth.

I don't think this is the case right now - if anything, the model is spending down its budget and/or getting itself into rabbit holes.

Will there be an Auto-GPT using GPT-4 that is fully self-sustaining, i.e. is taking continuous actions every day for at least 30 days without using up its budget or other non-renewable resources?

(Evaluating this market will require some judgment calls.)

Get Ṁ1,000 play money

🏅 Top traders

#NameTotal profit
1Ṁ493
2Ṁ64
3Ṁ60
4Ṁ53
5Ṁ51
Sort by:

"taking continuous actions every day for at least 30 days without using up its budget or other non-renewable resources"

Couldn't this resolve yes by spacing out the time intervals between actions it's allowed to take, and starting with a high budget? Really, the only requirement is that it doesn't run out of money, which you could do by starting with a huge pool of money.

predicted YES

@jonsimon Actually though I would like an answer to this question. What does "taking continuous actions" mean concretely? An action per hour? Per second? @stuhlmueller

predicted YES
predicted YES

If there is an open source program that is technically similar to Auto-GPT (but is not the same) which fulfills the above criteria, does this resolve as yes or no?

@RobertCousineau Resolves as yes

bought Ṁ10 of NO

I'm so confused on why this is so popular. I looked at the code and it literally just wraps prompts around GPT4...so it's literally just GPT4 with prompt injection (or GPT3.5 if you don't have the premium OpenAI plan). What am I missing? Am I crazy?

@PatrickDelaney It's not just prompt injection, you're missing that it lets the program perform actions other than generating text, like downloading files, browsing the web, running files, etc. The prompt injection makes the model select an action for a task, the code runs the action and gives the output to the model so it can determine the next task and action.

predicted NO

@horse I've seen a few threads on Twitter about Auto-GPT since I posted my previous comment. I'm still a bit confused about what it can do, as I'm seeing a lot of claims about its capabilities without a lot of evidence to back them up. I also saw a demo for a centralized search/LLM agent that seemed to be super easy and intuitive to use. It's possible that I'm not the target audience for Auto-GPT, but it's hard to tell when there's so little information available. I'd love to learn more about Auto-GPT, so if you have any resources you can share, I'd be grateful. This is what I found. https://www.phind.com/

Is a human allowed to do anything besides set the initial prompt? E.g. can they provide additional prompts to get AutoGPT unstuck if it gets into a loop?

@Benthamite Yes, if it's part of the system - e.g. AutoGPT reaches out to the human when it's stuck - and it's a relatively small part, that's OK. If it's an out-of-the-loop intervention or it has to happen after every few actions it's not OK. For borderline cases I'll err on the side of calling it not autonomous.

@stuhlmueller Would autoGPT need to pay the human in this case?

@MartinRandall No, doesn't matter why the human does it

bought Ṁ65 of YES

How confident that there is at least one business opportunity with GPT-4-in-2023-pickable low-hanging fruit (seems certain, pretty sure I can think of examples myself) and that someone will find a way to apply GPT-4 to pick it (seems possible)? I guess 24% seems underpriced.

I feel like dropshipping is probably fully automatable. Like, to the point where you probably don't even need a LM to do it, but you could. Auto-GPT looks like it has limited ability to interact with the web, but there's other tools that make it easier for GPT to do that.

I feel like for this to be trivially accomplished, the underlying model needs to be significantly better and roughly 10x cheaper.

Comment hidden