Hmm Lightcone Commons uses some proportional distribution system I think -- though I didn't quite read it in enough detail to understand it -- which would be similar in spirit but sligtly different from what I meant.
Yeah I think the "ideal design" you mention afterwards is what I had in mind here. I've even read SFF grantmakers (I think in Zvi's reflection from 1-2 years ago) stating they are less likely to fund something with SFF money if they think CG will fund it. Which seems like a particularly bad outcome -- though understandable with the current mechanisms.
Awesome to see your write-up! (context: I'm the founder of Lens Academy)
> Instead of assigning dozens of disconnected papers, blog posts, podcasts, and videos, Lens Academy builds the programme around a single primary book:
If Anyone Builds It, Everyone Dies
Yup. So, our other courses do consist of dozens of disconnected recources, but we wanted to try out an intro course with IABIED and so far are pretty happy with it. The book isn't perfect, but provides a relatively compelling case for AI takeover risk at an introductory level.
We'll probably be expanding this intro course to contain some different resources too, especially to provide context about the AI Safety ecosystem and possible next steps people can take.
I'd love to see something that's somewhat similar to SFF's speculation grants, though a bit different: - Small, fast funders can fund as they do - if the receiving org then gets a grant from a larger funder, the smaller funder gets their money back.
E.g. get 50k from a small funder with a 1M yearly budget. Then get 500k from CG, 50k of which is used to pay back the small funder.
This seems to move the incentive from "don't fund a project if you think CG will fund them anyway" to the opposite: "fund a project especially if you think CG will fund them anyway"
This seems useful if (I guess) these two conditions are met: - dollar for dollar, the impact of smaller funders is higher - smaller funders are faster
(I think Matt Brooks recently mentioned something like this to me, though possibly in a different form)
Hmm Lightcone Commons uses some proportional distribution system I think -- though I didn't quite read it in enough detail to understand it -- which would be similar in spirit but sligtly different from what I meant.
Yeah I think the "ideal design" you mention afterwards is what I had in mind here. I've even read SFF grantmakers (I think in Zvi's reflection from 1-2 years ago) stating they are less likely to fund something with SFF money if they think CG will fund it. Which seems like a particularly bad outcome -- though understandable with the current mechanisms.
Awesome to see your write-up! (context: I'm the founder of Lens Academy)
> Instead of assigning dozens of disconnected papers, blog posts, podcasts, and videos, Lens Academy builds the programme around a single primary book:
Yup. So, our other courses do consist of dozens of disconnected recources, but we wanted to try out an intro course with IABIED and so far are pretty happy with it. The book isn't perfect, but provides a relatively compelling case for AI takeover risk at an introductory level.
We'll probably be expanding this intro course to contain some different resources too, especially to provide context about the AI Safety ecosystem and possible next steps people can take.
I'd love to see something that's somewhat similar to SFF's speculation grants, though a bit different:
- Small, fast funders can fund as they do
- if the receiving org then gets a grant from a larger funder, the smaller funder gets their money back.
E.g. get 50k from a small funder with a 1M yearly budget. Then get 500k from CG, 50k of which is used to pay back the small funder.
This seems to move the incentive from "don't fund a project if you think CG will fund them anyway" to the opposite: "fund a project especially if you think CG will fund them anyway"
This seems useful if (I guess) these two conditions are met:
- dollar for dollar, the impact of smaller funders is higher
- smaller funders are faster
(I think Matt Brooks recently mentioned something like this to me, though possibly in a different form)