Number of Watts Needed to Run The First Implementation of AGI?
33
344
1k
2025
280k
expected

The human brain requires about 12 Watts of Power to operate, asleep or awake.

What will be the number of Watts required to run the simplest possible definition of AGI, at the time when it is achieved?

Market will resolve to the foremost technical expert in AI's lower bound estimation at the time of closing. If the foremost expert has not made an esitmation at the time of closing, it will go to the secondmost expert, and so on.

20230408 - it has been pointed out to me by @Rwin that, "simplest possible interpretation of AGI", could be strictly interpreted as meaning, any computational implementation that could achieve the results of a superintelligence, even over a very long timeframe, such as a 1000+ year timeframe for example. This was not the original intent of the market, the original intent was, "whatever the easiest to achieve AGI," meaning, the first implementation of AGI.

Further attempting to better define the market:

  • The AGI must be implemented using physical materials, it cannot be a mathematical construct that does not exist in reality.

  • The AGI must be implemented in, "real time," - e.g. the definition of intelligence must be accepted to mean, "on the same timescale that humans operate on."

Apr 8, 9:13am: Number of Watts Needed to Run Simplest Possible Definition of AGI? → Number of Watts Needed to Run The First Implementation of AGI?

Get Ṁ600 play money
Sort by:

Suppose if the first "AGI" still uses the paradigm of separated training from running. In that case, are we measuring watts to train, or watts to run the output model?

predicts LOWER

@ScroogeMcDuck to run.
"The human brain requires about 12 Watts of Power to operate, asleep or awake."
training would be measured in watt-hours. (in the case of a human 12 watts * 18 years)

@LoganZoellner Thank you!

1.21 jigawatts!!

🌟 Let's discuss the lower bound of AGI and the plausibility of human doom from AGI 🌟 I was trying to do some reading to better define this market, found this book and the section on, "The Turing Trap." https://library.oapen.org/bitstream/handle/20.500.12657/61656/9781000870817.pdf?sequence=1#page=124 ... does the bar for AGI just keep getting raised, making even the lower bound much higher than we think it would be? E.g. let's say there's a python package called `agi 1.0.0`, -- as soon as people start calling that package and using it, does it become obsolete, because humans + AGI is always inherently superior to AGI alone? Or is there some super-AGI that beats out all humans + AGI?

As a real-world example, there's Advanced Chess. Though I'm not sure if it has been shown now that having a, "human in the loop," for Advanced Chess is of any advantage anymore or not? https://en.wikipedia.org/wiki/Advanced_chess

@PatrickDelaney To clarify, my comment above with the stars was meant to be joking. Anything dealing with how the market will be determined will be listed in the description above. I'm just trying to get more conversation going as this market seems to have gone stale for a couple of weeks now.

predicts HIGHER

We're at 1000W now, so basically a space heater.

predicts HIGHER

@PatrickDelaney Sorry. 1900W, a high end space heater.

predicts LOWER

Would this affect an estimate of how many simultaneous AGI instances are possible/likely? (whether co-located or in existence at all)

predicts HIGHER

@cloudprism Let's assume that there is a coupled software and hardware implementation that exists in the space G. This might involve one, "rack of servers," H, in one location, which involves networking inside of a space, or it might involve multiple H connected over a WAN. The wattage required to run H over G to successfully run the fist implementation of AGI 1.0 is constitutes the threshold necessary to resolve this market. Does that clear it up sufficiently?

predicts LOWER

@PatrickDelaney Oh, I understand, I was wondering about opening another market about how many AGI there would be in the world

predicts HIGHER

@cloudprism Should I update the description to clarify?

predicts HIGHER

This is not meant to push the market one way or another, this is just a funny utterance/observation. I find it funny that the highest, "YES," better is in fact arguably an AI.

predicts HIGHER

@PatrickDelaney As if the bot is saying, "give me more power!"

bought Ṁ30 of HIGHER

The current estimate is about 2-3 modern server racks

predicts HIGHER

@Rwin Citation? Which current estimate? The market's current estimate of 55kW or the estimate of what AGI's power requirements may be from experts? Sorry, I'm confused.

predicts HIGHER

@PatrickDelaney I am giving context fof the current market estimate of 55kW.

Citation: "some racks can hit as much as 20kW"

https://www.serverwatch.com/servers/server-room-power-consumption/

predicts HIGHER

@Rwin thank you, understood.

predicts LOWER

well, that changed my estimate a lot

bought Ṁ500 of LOWER

HUMANS ARE GI'S

predicts HIGHER

@AlexAmadori And an in-silico simulation of a human brain is an artificial GI

predicts HIGHER

https://arxiv.org/pdf/2109.05472.pdf

Here's a potential methodology, not saying this is how anything is being resolved at all, just speculating:

  1. Hand-wave how many GFLOPs may be needed for AGI.

  2. Find where in the future the two charts above intersect.

  3. Assume a constant for gains in efficiency due to algorithmic advancement.

  4. Profit.

bought Ṁ140 of HIGHER

@PatrickDelaney Looks nice! I'll maybe try to find the intersect tomorrow if no one else does it before me. The only problem I see is possibly stagnating progress in computing power due to quantum tunneling, or the reverse effect caused by the adaption of quantum computing. Both might vastly change the future growth.

predicts HIGHER

@PatrickDelaney Human brain, 2000 TFLOPs or 2,000,000 GFLOP, using that as a benchmark, we're about 3 decades away from that. Energy consumption efficiency seems to be going up by about 1 order of magnitude of GFLOP/Watt, with it currently being around 200 GFLOP/Watt average, so maybe we would be around 20,000 GFLOP/Watt average by that point.

2e6/2e4 = 2e2 = 200W.

This is making a ton of assumptions and likely 2000TFLOP's does not really represent the human brain.

predicts HIGHER

@Rwin interestingly this is not that far off from your article below. But that being said, this is probably a poor representation of what people widely accept to be AGI. The biggest problem with this market is defining what AGI is.

More related questions