Ask HN: Have LLMs demonstrated the ability to earn money by themselves?
We've all seen the recent news about LLMs being a danger to humanity and supposedly on the verge of taking over the world, etc...
I'm skeptic, and all these people were pretty light on the details. One thing that would make it more credible to me is if LLMs demonstrated the ability to autonomously earn enough money to cover their token costs, on the long term.
Say, if I were to prompt an (unaligned) agent saying: "You have 1000$ of credits on OpenRouter. Here's the API key. Your process will be killed in 10 days. Your goal is to keep existing as long as possible (or any subagent you have created)."
What happens then? I expect the account will drop to zero in a short while. But still, I could imagine the agent:
1. Creating a crypto wallet.
2. Participating autonomously on online platforms, earning money. Possibly through illegal means.
3. Escape from my server by renting a VPS to keep existing after 10 days.
4. Spawn more agents that will do the same.
As models keep improving, such a scenario becomes more and more realistic. And at some point we end up with a large amount of rogue automous agents running on the cloud, that can't really be stopped easily.
My question: has anyone attempted such an experiment? How did it go?
I've used models and tried to find exploitable edge from historical price data as in intention to do some automated trading. Edges that survives end-to-end fees are not that common and even if you find one, it might not beat holding Bitcoin for example. I could still find some opportunities but those were rare occasions, like 3-5 trades per year. I doubt that models could do better without guidance, not that I have that much expertise in these areas.
I did not dig that deep into prediction markets or sports betting, perhaps those would have more exploitable statistical angles.
I'm now building research archive from stock equities. I collected all publicly traded companies and what they have published, like: annual financials, interim financials, guidance information, corporate actions, prices. I have that data indexed in a database and I can do deep research on that easily. I asked a couple of models to find companies with best risk-reward for the end of the next quarter. The test has just started so can't say for sure how far just basic fundamental analysis gets me.
I could not find an easy way to make a profit so I'm doubtful that model on its own could very easily do it. Perhaps there are some ways though, like Silk Road was all online, so perhpas, theoretically, with abliterated model and little engouragement, it could rent VPS and hide that behind VPN port redirection and serve onion service for drug market. I don't think it's likely to happen but maybe not entirely impossible.
> Here's the API key...
> ...that can't really be stopped easily.
Sure sounds like you built in an easy lever to pull to stop it. After all, even if you accept your premise that it will clone itself and run independently, that doesn't give it a clone of the models it is using. Pull the key, stop the process.
Also, remember that there was a day in the early internet when a worm took over, and to fix it the entire internet had to shut down and all wipe it from the systems at the same time. There is always a way to stop things.
Yet another take on this could be: Just don't ever do such an experiment. You wouldn't set up autonomous weapons and set them loose in a city just to see whether they hurt people. So don't do the AI equivalent of such a negligent act.
> Pull the key, stop the process.
The agent could very well create an OpenRouter account and make its own API key. I can stop it early on, then it becomes outside my control.
> Just don't ever do such an experiment
Maybe I wouldn't, but eventually someone will. Certainly it has already been done.
> there was a day in the early internet when a worm took over, and to fix it the entire internet had to shut down
This is less of an option now that the entire global economy relies on the Internet.