4 ms·

> They're probably right that building AGI will require a ton of computational power, and that it will be very expensive. Why? This makes it seem like computer

> They're probably right that building AGI will require a ton of computational power, and that it will be very expensive. Why? This makes it seem like computers are way less efficient than humans. Maybe I'm naive on the matter, but I think it's possible for computers to match or surpass human efficiency.

Jensson 3y ago
Computers are still way less efficient than humans, a human brain has less power draw than a laptop and do some immense calculations to parse vision, hearing etc better than any known algorithm constantly. And the part of the human brain that governs our human intelligence and not just what animals do is much larger than, so unless we figure out a better algorithm than evolution did for intelligence it will require a massive amount of compute. The brain isn't fast, but it is ridiculously parallel with every cell being its own core so total throughput is immense.
FeepingCreature 3y ago
Computers are more efficient, dense and powerful than humans. But due to self-assembly, brains consist of many(!) orders of magnitude more volume. A human brain is more accurately compared with a data center than a chip.
- Jensson 3y ago
  > A human brain is more accurately compared with a data center than a chip. A typical chip requires more power than a human brain, so I'd say they are comparable. Efficiency isn't per volume but per power or per heat production. Human brains wins those two by far.
  - FeepingCreature 3y ago
    To be fair, we've locked ourselves into this to some extent with the focus on lithography and general processors. Because of the 10-1000W bounds of a consumer power supply, there's little point to building a chip that falls outside this range. Peak speed sells, power saving doesn't. Data center processors tend to be clocked a bit lower than desktops for just this reason - but not too much lower, because they share a software ecosystem. Could we build chips that draw microwatts and run at megahertz speeds? Sure, probably, but they wouldn't be very useful to the things that people actually do with chips. So imo the difficulty with matching the brain on efficiency isn't so much that we can't do it as that nobody wants it. (Yet!) edit: Another major contributing factor is that so far, chips are more bottlenecked on production than operation. Almost any female human can produce more humans using onboard technology. Comparatively, first-rate chips can be made in like three buildings in the entire world and they each cost billions to equip. If we wanted to build a brain with photolithography, we'd need to rent out TSMC for a lot longer than nine months. That results in a much bigger focus on peak performance. We have to go "high" because we cannot practically go "wide".
tsimionescu 3y ago
Perhaps the finalized AGI will be more efficient than a human brain. But training the AGI is not like running a human, it's like speed running evolution from cells to humans. The natural world stumbled on NGI in a few billion years. We are trying to do it in decades - it would not be surprising that it's going to take huge power.
hackerlight 3y ago
Scaling laws. Maybe they will figure out a new paradigm, but in the age of Transformers we are stuck with scaling laws.

Jensson 3y ago

Computers are still way less efficient than humans, a human brain has less power draw than a laptop and do some immense calculations to parse vision, hearing etc better than any known algorithm constantly. And the part of the human brain that governs our human intelligence and not just what animals do is much larger than, so unless we figure out a better algorithm than evolution did for intelligence it will require a massive amount of compute. The brain isn't fast, but it is ridiculously parallel with every cell being its own core so total throughput is immense.

FeepingCreature 3y ago

Computers are more efficient, dense and powerful than humans. But due to self-assembly, brains consist of many(!) orders of magnitude more volume. A human brain is more accurately compared with a data center than a chip.

Jensson 3y ago
> A human brain is more accurately compared with a data center than a chip. A typical chip requires more power than a human brain, so I'd say they are comparable. Efficiency isn't per volume but per power or per heat production. Human brains wins those two by far.
- FeepingCreature 3y ago
  To be fair, we've locked ourselves into this to some extent with the focus on lithography and general processors. Because of the 10-1000W bounds of a consumer power supply, there's little point to building a chip that falls outside this range. Peak speed sells, power saving doesn't. Data center processors tend to be clocked a bit lower than desktops for just this reason - but not too much lower, because they share a software ecosystem. Could we build chips that draw microwatts and run at megahertz speeds? Sure, probably, but they wouldn't be very useful to the things that people actually do with chips. So imo the difficulty with matching the brain on efficiency isn't so much that we can't do it as that nobody wants it. (Yet!) edit: Another major contributing factor is that so far, chips are more bottlenecked on production than operation. Almost any female human can produce more humans using onboard technology. Comparatively, first-rate chips can be made in like three buildings in the entire world and they each cost billions to equip. If we wanted to build a brain with photolithography, we'd need to rent out TSMC for a lot longer than nine months. That results in a much bigger focus on peak performance. We have to go "high" because we cannot practically go "wide".

tsimionescu 3y ago

Perhaps the finalized AGI will be more efficient than a human brain. But training the AGI is not like running a human, it's like speed running evolution from cells to humans. The natural world stumbled on NGI in a few billion years. We are trying to do it in decades - it would not be surprising that it's going to take huge power.

hackerlight 3y ago

Scaling laws. Maybe they will figure out a new paradigm, but in the age of Transformers we are stuck with scaling laws.