Honestly if I knew how to do this I would do it. How do I convince a bunch of billionaires to hand me money to do very fundamental research? Dream would be to have it be very open too and hire people from multiple domains but keep a high focus on being nuance and detail oriented. Honestly, I don't think companies like OpenAI are going to end up being the first to AGI because they get caught in the profit stages and so just pivot to say that more data and more compute is all you need. Maybe it is true, but I doubt it. And most certainly there are other ways considering a 3 pound piece of meat does it using less than a kW per day.
> Honestly if I knew how to do this I would do it.
Do an MVP in your spare time.
I guess if you have a normal programming job in most countries (even in countries where programmers are not paid that well), you will likely be able to pay the bills for, say, server renting, domains, ...
For things that you will do likely badly by yourself like design drafts (if you are bad at visual design), ask some friends whether they would be willing to do it for you in exchange for something that you could do for them (so that you don't have to spend money).
Start with kinds of research that can be done with little money, but have a possible insanely high impact.
Iterate until your platform cannot be ignored anymore.
I think you gravely misread my comment if this is your first suggestion. I am essentially saying that what I would want to do is closer to starting something akin to DeepMind or what OpenAI looked like in the beginning. There is no clear minimal viable product because the goal is to create general intelligence. There are many demonstrations of narrow cases of generalizability and I even have work that demonstrates this, but the stronger sense and actually trying to build sentient machines.
I would make the argument that there is an advantage in what I'm proposing. Since the existing labs that are trying to perform a similar goal are all following similar paths. The chance of beating them using similar methods is quite low. But if you believe (as many researchers do) that the conventional approach is not correct or not efficient, then it reasons that you should take an unconventional approach. I'd also say that my preferred method can be great publicity for them if we do allow for a fully open research platform (open training/models/communication) and given the goals the computational costs are not quite what they are for OpenAI nor DeepMind. As I believe I could do a lot of work with a small team with only a handful of A100 or H100 nodes. That's pocket change to the Musk types and especially if split between multiple entities. It is a high risk high reward situation but I think also low cost.
> There is no clear minimal viable product because the goal is to create general intelligence. [...] I would make the argument that there is an advantage in what I'm proposing.
Then set the MVP to some project that partially solves some small aspect of what you general intelligence is and proves that your approach is so much better than what other labs are doing.
If you succeed on this, you can iterate.
> if we do allow for a fully open research platform (open training/models/communication) and given the goals the computational costs are not quite what they are for OpenAI nor DeepMind.
Then work on a way to decrease these computational costs by magnitudes. Or find a way how even volunteers with "skinny" GPUs in their computers can still support your computations via crowdcomputing.
Perhaps this is even the more important problem to solve ...
