3 ms·
> Do an MVP in your spare time. I think you gravely misread my comment if this is your first suggestion. I am essentially saying that what I would want to do i
by godelski 3y ago
> Do an MVP in your spare time.
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.
- aleph_minus_one 3y ago> 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 ...