Former Google chief Eric Schmidt warns electricity shortages, not semiconductors, pose the biggest threat to AI superintelligence development. Here's what that means.
“AI’s natural limit is electricity, not chips,” Schmidt said, cutting through the industry’s semiconductor obsession with characteristic bluntness.
I mean, maybe in the very long term that’s a fundamental limit, and you face things like Dyson spheres.
But right now, I’m personally running one human-level AGI on roughly 100W of power, so I’m just gonna say that as things stand, the prominent limitation is software not being good enough. You’re, like, a software guy.
Ultimately AI is an optimization problem, and if we don’t know how to solve the software problems fully yet, then, yeah, we can be inefficient and dump some of the heavy lifting on the hardware guys to get a small edge.
But I’m pretty sure that the real breakthrough that needs to happen isn’t on the hardware side. Like, my existing PC and GPU already are more capable than my brain from a hardware standpoint. The hardware guys have already done their side and then some compared to human biology. It’s that we haven’t figured out the software to run on them to make them do what we want.
The military or whoever needs AI applications can ask for more hardware money to get an edge relative to competitors. But if you’re the (well, ex-) head of Google, you’re where a lot of those software and computer science guys who need to make the requisite software breakthroughs probably are, or could be. Probably the last people who should be saying “the hardware guys need to solve this”.
It’s going to be some more profound changes to what we’re doing in software today than just tweaking the parameters on some LLM model, too. There’s probably some hard research work that has to be done. It’s not “we need immense resources dumped into manufacturing more datacenters and powerplants, and chips”. It’s translating money into having some nerdy-looking humans bang away in some office somewhere and figure out the required changes to what needs to be done in software to get us there. Once that happens, then okay, sure, one needs hardware to make use of that software. But in July 2025, we don’t have the software to run on that hardware, not yet.
Give yourself some credit! Your human-level intelligence is only using about 20W. The other 80 is for the meat robot it has to pilot to get stuff done.
I mean, maybe in the very long term that’s a fundamental limit, and you face things like Dyson spheres.
But right now, I’m personally running one human-level AGI on roughly 100W of power, so I’m just gonna say that as things stand, the prominent limitation is software not being good enough. You’re, like, a software guy.
Ultimately AI is an optimization problem, and if we don’t know how to solve the software problems fully yet, then, yeah, we can be inefficient and dump some of the heavy lifting on the hardware guys to get a small edge.
But I’m pretty sure that the real breakthrough that needs to happen isn’t on the hardware side. Like, my existing PC and GPU already are more capable than my brain from a hardware standpoint. The hardware guys have already done their side and then some compared to human biology. It’s that we haven’t figured out the software to run on them to make them do what we want.
The military or whoever needs AI applications can ask for more hardware money to get an edge relative to competitors. But if you’re the (well, ex-) head of Google, you’re where a lot of those software and computer science guys who need to make the requisite software breakthroughs probably are, or could be. Probably the last people who should be saying “the hardware guys need to solve this”.
It’s going to be some more profound changes to what we’re doing in software today than just tweaking the parameters on some LLM model, too. There’s probably some hard research work that has to be done. It’s not “we need immense resources dumped into manufacturing more datacenters and powerplants, and chips”. It’s translating money into having some nerdy-looking humans bang away in some office somewhere and figure out the required changes to what needs to be done in software to get us there. Once that happens, then okay, sure, one needs hardware to make use of that software. But in July 2025, we don’t have the software to run on that hardware, not yet.
Give yourself some credit! Your human-level intelligence is only using about 20W. The other 80 is for the meat robot it has to pilot to get stuff done.
Tower, moving into refrigerator area
Hands, prepare for food extraction
Stomach stand by
Stomach stand by
Eyes show doors within range
Extending arm, prepare handle grasp
Handle secured
Contract arm for open
Door is opening
*alarm siren*
Unauthorized food ejection ALERT
Eyes reporting the raspberries fell out again
*launch expletives*
Stomach, stand by
All arms, hands defend against the animals
Emergency cleanup all hands
*launch expletives, threaten animals*
Head warning, open refrigerator door - repeat: Head warning
Stomach, stand by
*auxilliary cursing: ON*
Yeah and you can survive outside of a data center and walk through a room with a full cup of coffee without spilling it.