🏅 Top traders

https://github.com/osmarks/liminal-octothorpe

Since I enjoy doing things at the last possible moment for no particularly good reason, here it is. I plugged a Manifold interface into some of my somewhat broken AI agent code, which includes web search (via scraping DuckDuckGo, so they block it sometimes). It works somewhat (I tested it).

@osmarks Can you make it into a Manifold bot?

@Shump It is a "Manifold bot" in the sense that if you run the script it automatically places bets and stuff. I don't have a separate account for it because I can't see a way to do that without another phone number.

@osmarks How automatic is it? Can you set it to run as a crob job or does it require a bunch of non-automatic prompting?

@Shump No user intervention required, except it's not very robust and might crash, but you can just autorestart it.

@osmarks You can just read the code.

@osmarks You win

Does a wrapper around GPT-4 that uses a search engine API to look up questions, an internal scratchpad, and the manifold API to make bets count?

In this comment thread, your answer seems to imply YES, as the linked twitter clip is of GPT-4 making automatic trades by just looking at the question titles, and your response seems to indicate that if GPT-4 actually did web searches and ran without human supervision, then it would resolve YES.

However, in this separate comment thread, your answer seems to imply NO.

I made a private GPT that does this for me (among other things like help create and resolve markets).

https://chat.openai.com/share/1af42a1e-3b75-4f80-a20b-2da226404951

You can see in the last message it does research and makes a bet based on research without any intermediate prompting.

Not sure if this satisfies the criteria. It doesn’t “automatically” search the internet, but it can automatically make a relevant bet after being prompted to do research.

@Charlie @DAL59 Does this count?

@Panfilo Looks like they're inactive. Might need to request mod judgement.

@Charlie Asking an existing AI to do something isn't "making" an AI unfortunately.

@DAL59 Oh so if it used any LLM's API at all it wouldn't count?

@DAL59 I am having trouble understanding the distinction here. What exactly did you have in mind for how this would work?

@DAL59 Is your distinction just "does it involve code and not just prompts"?

nice market

Doesn't @Yuna Do this? Or do they just look at betting odds from bookmakers, which I'm not sure would qualify under the criteria

@DAL59 Apparently Yuna uses sports APIs to get scores. Does that count? In the previous comments you confirmed that it should be a deep learning model so my guess is it doesn't count, because it's not deep learning

@Shump Is Yuna AI? If its creator trained a neural net on sports scores to predict outcomes, or something like that, then the your question is the crux. But if Yuna's logic is algorithmic, most wouldn't call that AI.

If Yuna's logic is some big regression model, then it starts to get a bit blurry.

@chrisjbillington Yes that's why I bought NO. I don't think it counts based on the previous comments.

@Shump I don't know why I wasn't already betting hard NO here. Lost opportunity, good trade on your part.

@chrisjbillington Because I'm secretly working on this internally, obviously.

@osmarks Good luck!

@chrisjbillington Luck is a low-interest-rate phenomenon.

If you search for "forecaster bot war" on Manifold you'll see "bots", some of which are indeed doing internet searches to make predictions. But actually, it looks like the prompting and the betting are being done by humans, rather than the bots operating autonomously. So they do not count for this market. It seems like this is a deliberate test to compare the different models' ability to forecast, rather than it being important that they're actually automated.