Everyone, especially people working in this area, know that data and energy are the huge costs of current generation AI, so many people are working to make those things better. Neurosymbolic methods will give more AI bang for fewer energy bucks. Lots of work is being done to make all flavors of DNN more compute-efficient. It's a very good bet that those efforts will produce good results within 5-10 years. That's eternity in VC time, of course (so those guys can't even think about it), but comparable to the time to put, say, nuclear generation online. I suggest that you put some effort into following those efforts: (a) to get far more value from orders-of-magnitude less training data (like human learning does), and (b) to do functionally the same neural net computations with significantly less energy (maybe orders of magnitude, maybe not). This is stuff that will change the nature of the game. Everyone is currently playing one game, and soon it will become clear that there is a different game to be played. Get in front of it. Cheers, Ben
Very true. The bet is that these space data centers are still economically viable and efficient in 5-10 years which is when they will be come fully functional vs. nuclear fusion, nuclear fission, SMRs and other technologies that are improving.
Actually, my point is that AI research will progress to reduce the energy need, perhaps very greatly, so neither earth-based nor space-based huge increases in energy generation will be necessary. If that scenario plays out (and I will bet on it), then huge investments in energy generation capabilities will be largely wasted. Much like the huge amount of unnecessary fiber optic cable installed in the 1990s.
Everyone, especially people working in this area, know that data and energy are the huge costs of current generation AI, so many people are working to make those things better. Neurosymbolic methods will give more AI bang for fewer energy bucks. Lots of work is being done to make all flavors of DNN more compute-efficient. It's a very good bet that those efforts will produce good results within 5-10 years. That's eternity in VC time, of course (so those guys can't even think about it), but comparable to the time to put, say, nuclear generation online. I suggest that you put some effort into following those efforts: (a) to get far more value from orders-of-magnitude less training data (like human learning does), and (b) to do functionally the same neural net computations with significantly less energy (maybe orders of magnitude, maybe not). This is stuff that will change the nature of the game. Everyone is currently playing one game, and soon it will become clear that there is a different game to be played. Get in front of it. Cheers, Ben
Very true. The bet is that these space data centers are still economically viable and efficient in 5-10 years which is when they will be come fully functional vs. nuclear fusion, nuclear fission, SMRs and other technologies that are improving.
Actually, my point is that AI research will progress to reduce the energy need, perhaps very greatly, so neither earth-based nor space-based huge increases in energy generation will be necessary. If that scenario plays out (and I will bet on it), then huge investments in energy generation capabilities will be largely wasted. Much like the huge amount of unnecessary fiber optic cable installed in the 1990s.