AI More Energy Efficient than Humans, New Study Finds

I recently read that using artificial intelligence creates fewer carbon emissions than human labour. Really? The human brain doesn't have much going for it, but we know that it is much more energy efficient than computers. What's going on here? Let’s have a look.

🔗 Join this channel to get access to perks ➜
#science #sciencenews #artificialintelligence #technology #technews

[albertocontu]

Does the study take into consideration that for any "used" ai generated images many more are generated and discarded?

[antongromek]

I had a chess computer once, but it was much too good for me - so I ordered another one, now they play against each other, and I can do whatever I want.

[oofcloof]

I’ve never thought of a brain as an edible computer 😂

[CamAlert]

Prompt: "Minimize carbon emissions"
AI: Destroys humans

Let's just hope we get this alignment thing right...

[LB-vfhm]

I'm with you on this. I think it's ridiculous to take into account the carbon footprint necessary for living in a society as a human as part of the "cost" of art, but not take that human / society "cost" into account for AIs, considering they're tools made by humans within a society to fill the needs of other humans within a society, and will in the future both rely on, support, and be deeply entwined with the same supply chains and costs that keep that society running. Seeing as humans and industry are the creators and users of AI, the human upkeep costs remain.

[adonisengineering]

You perfectly described the dead internet dilemma: AI produced script for AI produced content which will be AI ingested to get context for AI produced comments and engagement.

[alexxx]

We'll produce less CO2 as paperclips, for sure!

[alieninmybeverage]

Politician brains are currently working on rebranding carbon footprint to carbon treadprint so they can blame someone else. It paints a picture very efficiently.

[dominic.h.]

Did they account for having to regenerate an output fifteen different times to get a result that isn't garbage? I can smash the keyboard randomly while holding my breath and claim I'm more carbon efficient than a New York Times Bestseller...

[utkua]

Video Game crash of 80s, Dotcom crash of 90s, has a similarity to today's AI hype, it is not like Web failed or games failed, but at the beginning they oversold it, hyped them way above the reality and in the end they did disappoint everybody. I think similar things will happen.

[gigaherz_]

I find the implication of that paper that you can remove the entire human from the equation if the human is not producing generative content. If we replaced all the "generative" jobs with AI, are they proposing to ... remove those humans permanently?
I would look into who wrote that paper. It may be they have already been taken over by AI.

[guard]

So I made a calculation like this myself back in February, where I estimated the energy cost per day per user, and compared that to a human's average energy cost per day (just looking at food intake). My intent was to say that a person + ChatGPT can produce work at a rate similar to 2 humans working, but at a cost less than 2 humans. To me, the point is that it is an effectiveness multiplier. I think what everyone keeps forgetting is that it can't completely take away work, and someone needs to babysit it, but that babysitting is more productive than just having more humans working together. Don't say that's it's more efficient so we should replace humans, say that it is efficient enough that we can have fewer people doing tedious work and devote more effort and time to things it can't do (or to leisure time, we deserve that).

It scares me to consider that some would use this to justify getting rid of people, especially when they think of it in terms of "it would be better to have fewer people".

[Deruzejaku]

" - We should use AI instead of humans, they use less energy for same task
- But what with humans then?
- Ohhh just shut them down so they don't Emit CO2"

Gota love today's world xD

[tomamberg]

Hi Sabine! I've been involved with a couple of recent huge AI-dedicated chips. Two points I can think of off the top of my head: 

1. the energy per inference is going down A LOT thanks to new AI chip architectures. They try as much as possible to not move data around (but rather mix compute circuits with memory circuits), and work to reduce the bits of resolution that need to be processed (use clever tricks to take an 8-bit problem and execute it in 4 bits).

2. Engineering these chips takes tons of energy! Each requires years of use of server farms comprised of nothing but the most powerful systems possible. For my work on one chip, I needed 6TB of RAM for my jobs to complete! I wonder if this study takes _that_ into account!

[crashfactory]

At 1:08, I think you meant that the image generation of the unicorn costs 3wh, not 3w. It's still a valid comparison if the assumption is that the human brain takes an hour to generate that image, but if the human brain takes 9 minutes to generate, then they would both consume 3wh of energy

[ZappyOh]

This calculation only works, if humans can be turned off, while the AI does their work.
Is that the plan?

[hosermandeusl]

Q: is it moi or has anyone else noticed a veritable cornucopia of AI gibberated YouTube videos? The symptom (to me) is the lilting prose & cartoonish imagery that copies already eggstablished QUALITY channels & their topics. Seriously, the last two months have been a daily blocking of these "whack-a-mole" copycats.

[josiah]

Never trust someone without a biology degree to be able to do these calculations accurately. I have a PhD in computational biology and I've been following AI progress for decades. One of the truly astonishing differences is how much more energy efficient biology is than computers.

As an example, silicon circuitry will use a stream of electrons (conducting wire) in order to represent a single bit. Whereas the same amount of electricity could be used in the brain for a single neuron's action potential, which carries 100+ bits of info. There was a study published in 2023 showing that to correctly model a single real neuron it would take 3 to 4 layers of a neural network.

It's true that brains and neural networks are doing similar things, But technologists almost always forget about the larger context when making these comparisons. You have to include the factory that makes machines all the people that repair the machines all the people that set them up all of the work in things like air conditioning and water cooling, then finally tear down. Which by the way every single atom in biology gets reused so a fair comparison would really require 100% recyclability. On top of that life uses the most abundant elements available so that you don't have supply chain issues around rare earth metals etc.

The ATP synthase power distribution inside of a single cell uses individual protons to turn a turbine. It's literally impossible to get more energy efficient than that. They're protons, they can't be downsized.

I think it may be possible to one day make a robot that's more energy intensive than humans, but only because they lack all the perks like self repair. Don't underestimate the cost of training and manufacture. I think in the case of this paper, the issue is that they're assuming equivalent quality in art while noting what everyone already knows which is that AI art is relentlessly prolific. Though each one of these models is only going to last a couple of years. They're not really talking sustainability. They're just playing around with big numbers in order to sell a pro-AI agenda.

[madcow]

Edible?!? This is why I watch your channel.

[eddys.]

"... created by AI, watched by AI... "

That cracked me up...