I currently have something like 50% credence that the point of no return will happen by 2030. Moreover, it seems to me that there’s a wager for short timelines, i.e. you should act as if short timelines scenarios are more likely than they really are, because you have more influence over them. I think that I am currently taking short timelines scenarios much more seriously than most people, even most people in the AI safety community. I suppose this is mostly due to having higher credence in them, but maybe there are other factors as well.
Anyhow, I wonder if there are ways for me to usefully bet on this difference.
Money is only valuable to me prior to the point of no return, so the value to me of a bet that pays off after that point is reached is approximately zero. In fact it’s not just money that has this property. This means that no matter how good the odds are that you offer me, and even if you pay up front, I’m better off just taking out a low-interest loan instead. Besides, I don’t need money right now anyway, at least to continue my research activities. I’d only be able to achieve significant amounts of extra good if I had quite a lot more money.
What do I need right now? I guess I need knowledge and help. I’d love to have a better sense of what the world will be like and what needs to be done to save it. And I’d love to have more people doing what needs to be done.
Can I buy these things with money? I don’t think so… As the linked post argues, knowledge isn’t something you can buy, in general. On some topics it is, but not all, and in particular not on the topic of what needs to be done to save the world. As for help, I’ve heard from various other people that hiring is net-negative unless the person you hire is both really capable and really aligned with your goals. But IDK.
There are plenty of people who are really capable and really aligned with my goals. Some of them are already helping, i.e. already doing what needs to be done. But most aren’t. I think this is mostly because they disagree about what needs to be done, and I think that is largely because their timelines are longer than mine. So, maybe we can arrange some sort of bet… for example, maybe I could approach people who are capable and aligned but have longer timelines, and say: “If you agree to act as if my timelines are correct for the next five years, I’ll act as if yours are correct thereafter.”
Any suggestions?
If AGI happens soon, there’s a decent chance it happens at an existing industry leader.
So one naive answer would be to buy Google (owner of DeepMind, which may be a significant fraction of their company’s value). Maybe also Microsoft: it does AI research, accepts US gov contracts, and has interacted with OpenAI, including buying some rights to GPT.
Yes. I did that a while ago. But that just gets me more money, and not even a lot of money. I need knowledge and help much more than money.
I don’t even think there’s enough money in all of EA to even get a board seat.
I just meant to buy GOOG shares to gain money.
If you wanted to buy large AI companies, you wouldn’t buy all of Google or Facebook, you’d just try to acquire AI projects. You could ask whether you can spend $20B to get a 1% chance of $2T somehow (options? crypto schemes? A big startup?) but in practice I think if you’re hoping to buy a $2T company, you’re not targeting properly.
Re: actual betting: Use a basket of proxies and warning signs which you expect to happen before your relevant timeframe. For example, to bet for/against a 25% of AGI by 2025, bet whether:
For all years before 2025, a [trusted third forecaster or a trusted aggregate] will give less than 25% probability to AGI being reached in or before 2025.
Google and Microsoft remain less valuable than Apple, with the rationale that they seem to be investing more heavily in ML/AI than Apple.
Specific capabilities which one would expect to see in 2023 given AGI before 2025.
the other party changes their mind by 2024.
Compute used by the largest model will not be scaled beyond [x].
...
If you wanted to do this at scale, then the thing to do might be to create some markets for each of the potential warning signs on Augur/Gnosis next year, when the platforms will a little bit more mature. Gnosis in particular combines neatly with Kleros, a distributed jury system which might be useful to resolve some questions with a subjective component.
What do you think about setting up a bunch of research prizes? Raise some questions that seem relevant (I suppose those could go from vague deconfusion to specific empirical questions) and put enough money on them, and repeat. I’ve seen some good results of this approach on Lesswrong, with prizes on the order of a couple hundred Dollars.
I’m also really interested in this as I also expect AGI sooner than others (50% 2030 is also what I thought last time), and also am unsure what to do about it. My PhD in Cognitive Science seems only rather vaguely useful so far.
Hmm, yeah, I guess doing a prize is less costly than hiring someone in the event that it doesn’t work. So I might as well experiment with that for a bit. Thanks!
If you are looking for ideas for things to do, want to chat sometime?
I’ve also been looking at prizes (though for a different topic). This list of past prizes might be of inspiration.
Yes, I’d be really interested, send you a PM.
Earmark donations to AI safety orgs/grantmakers for short timeline work. There might be issues with counterfactuals/fungibility you’ll need to talk through with them.
Start a fund/grantmaking organization to pool funds with others to support short timeline projects? You might be able to get advice from current AI safety grantmakers. You might be able to fund work at the intersection of short timelines and the priorities of existing orgs, so there might be room for collaboration.
Imo, prizes/awards are less motivating, since they’re not predictable enough to use to cover costs of living, so only people with other sources of income, financial support or savings can work full-time on prize problems.
Robin Hanson would probably take a bet: https://twitter.com/robinhanson/status/1330552185738899460?s=21