It was also inefficient for a computer to play chess in 1980. Imagine using a hundred watts of energy and a machine that costed thousands of dollars and not being able to beat an average club player.
Now a phone will cream the world’s best in chess and even go
Give it twenty years to become good. It will certainly do more stuff with smaller more efficient models as it improves
That’s the hangup isn’t it? It produces nothing of value. Stolen art. Bad code. Even more frustrating phone experiences. Oh and millions of lost jobs and ruined lives.
It’s the most american way possible that they could have set trillions of dollars on fire short of carpet bombing poor brown people somewhere.
Not even remotely close to this scale… At most you could compare the energy usage to the miners in the crypto craze, but I’m pretty sure that even that is just a tiny fraction of what’s going on right now.
Despite improving AI energy efficiency, total energy consumption is likely to increase because of the massive increase in usage. A large portion of the increase in energy consumption between 2024 to 2023 is attributed to AI-related servers. Their usage grew from 2 TWh in 2017 to 40 TWh in 2023.
This is a big driver behind the projected scenarios for total US energy consumption, ranging from 325 to 580 TWh (6.7% to 12% of total electricity consumption) in the US by 2028.
(And likewise, the last graph of predictions for 2028)
From a quick read of that source, it is unclear to me if it factors in the electricity cost of training the models. It seems to me that it doesn’t.
Racks of servers hum along for months, ingesting training data, crunching numbers, and performing computations. This is a time-consuming and expensive process—it’s estimated that training OpenAI’s GPT-4 took over $100 million and consumed 50 gigawatt-hours of energy, enough to power San Francisco for three days.
So, I’m not sure if those numbers for 2023 paint the full picture. And adoption of AI-powered tools was definitely not as high in 2023 as it is nowadays. So I wouldn’t be surprised if those numbers were much higher than the reported 22.7% of the total server power usage in the US.
It probably would have if IBM decided that every household in the USA needed to have chess playing compute capacity and made everyone dial up to a singular facility in the middle of a desert where land and taxes were cheap so they could charge everyone a monthly fee for the privilege…
Not the same. The underlying tech of llm’s has mqssively diminishing returns. You can akready see it, could see it a year ago if you looked. Both in computibg power and required data, and we do jot have enough data, literally have nit created in all of history.
This is not “ai”, it’s a profoubsly wasteful capitalist party trick.
That’s the argument Paul Krugman used to justify his opinion that the internet peaked in 1998.
You still need to wait for AI to crash and a bunch of research to happen and for the next wave to come. You can’t judge the internet by the dot com crash, it became much more impactful later on
One of the major contributors to early versions. Then they did the math and figured out it was a dead end. Yes.
Also one of the other contributors (weizenbaum i think?) pointed out that not only was it stupid, it was dabgeroys and made people deranged fanatical devotees impervious to reason, who would discard their entire intellect and education to cult about this shit, in a madness no logic could breach. And that’s just from eliza.
Edit: The underlying math and method. Not alone, of course. The main difference between then and now is the data set and some tuning, not a fundamentally new metjod or kibd of thing.
It seems like you are implying that models will follow Moore’s law, but as someone working on “agents” I don’t see that happening. There is a limitation with how much can be encoded and still produce things that look like coherent responses. Where we would get reliable exponential amounts of training data is another issue. We may get “ai” but it isn’t going to be based on llms
You can’t predict how the next twenty years of research improves on the current techniques because we haven’t done the research.
Is it going to be specialized agents? Because you don’t need a lot of data to do one task well. Or maybe it’s a lot of data but you keep getting more of it (robot movement? stock market data?)
We do already know about model collapse though, genai is essentially eating its own training data. And we do know that you need a TON of data to do even one thing well. Even then it only does well on things strongly matching training data.
Most people throwing around the word agents have no idea what they mean vs what the people building and promoting them mean. Agents have been around for decades, but what most are building is just using genai for natural language processing to call scripted python flows. The only way to make them look coherent reliably is to remove as much responsibility from the llm as possible. Multi agent systems are just compounding the errors. The current best practice for building agents is “don’t use a llm, if you do don’t build multiple”. We will never get beyond the current techniques essentially being seeded random generators, because that’s what they are intended to be.
Current approaches are displaying exponential demands for more resources with barely noticable “improvements”, so new approaches will be needed.
Advances in electronics are getting ever more difficult with increasing drawbacks. In 1980 a processor would likely not even have a heatsink. Now the current edge of that Moore’s law essentially is datacenter only and frequently demands it to be hooked up to water for cooling. SDRAM has joined CPUs in needing more active cooling.
Unless you mean to include those 1980 computers, in which case stockfish won’t run on that… More than about 10 year old home computer would likely be unable to run it.
Only because they are not 32 bit so they won’t support enough RAM. But a processor from the 90s could, even though none of the programs of the time were superhuman on commodity hardware.
The chess programs improved so much that even running with 1000 times slower hardware they are still hilariously stronger than humans
There will inevitably be a crash in AI and people still forget about it. Then some people will work on innovative techniques and make breakthroughs without fanfare
It was also inefficient for a computer to play chess in 1980. Imagine using a hundred watts of energy and a machine that costed thousands of dollars and not being able to beat an average club player.
Now a phone will cream the world’s best in chess and even go
Give it twenty years to become good. It will certainly do more stuff with smaller more efficient models as it improves
Show me the chess machine that caused rolling brown outs and polluted the air and water of a whole city.
I’ll wait.
Servers have been eating up a significant portion of electricity for years before AI. It’s whether we get something useful out of it that matters
That’s the hangup isn’t it? It produces nothing of value. Stolen art. Bad code. Even more frustrating phone experiences. Oh and millions of lost jobs and ruined lives.
It’s the most american way possible that they could have set trillions of dollars on fire short of carpet bombing poor brown people somewhere.
Not even remotely close to this scale… At most you could compare the energy usage to the miners in the crypto craze, but I’m pretty sure that even that is just a tiny fraction of what’s going on right now.
Crypto miners wish they could be this inefficient. No literally they do. They’re the “rolling coal” mfers of the internet.
Very wrong
In 2023 AI used 40 TWh of energy in the US out of a total 176 TWh used by data centers
https://davidmytton.blog/how-much-energy-do-data-centers-use/
From the blog you quoted yourself:
(And likewise, the last graph of predictions for 2028)
From a quick read of that source, it is unclear to me if it factors in the electricity cost of training the models. It seems to me that it doesn’t.
I found more information here: https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/
So, I’m not sure if those numbers for 2023 paint the full picture. And adoption of AI-powered tools was definitely not as high in 2023 as it is nowadays. So I wouldn’t be surprised if those numbers were much higher than the reported 22.7% of the total server power usage in the US.
It probably would have if IBM decided that every household in the USA needed to have chess playing compute capacity and made everyone dial up to a singular facility in the middle of a desert where land and taxes were cheap so they could charge everyone a monthly fee for the privilege…
Not the same. The underlying tech of llm’s has mqssively diminishing returns. You can akready see it, could see it a year ago if you looked. Both in computibg power and required data, and we do jot have enough data, literally have nit created in all of history.
This is not “ai”, it’s a profoubsly wasteful capitalist party trick.
Please get off the slop and re-build your brain.
That’s the argument Paul Krugman used to justify his opinion that the internet peaked in 1998.
You still need to wait for AI to crash and a bunch of research to happen and for the next wave to come. You can’t judge the internet by the dot com crash, it became much more impactful later on
No. No i don’t. I trust alan Turing.
NB: Alan Turing famously invented ChatGPT
One of the major contributors to early versions. Then they did the math and figured out it was a dead end. Yes.
Also one of the other contributors (weizenbaum i think?) pointed out that not only was it stupid, it was dabgeroys and made people deranged fanatical devotees impervious to reason, who would discard their entire intellect and education to cult about this shit, in a madness no logic could breach. And that’s just from eliza.
We’re talking about 80 years ago
~1948-52, yeah
Edit: The underlying math and method. Not alone, of course. The main difference between then and now is the data set and some tuning, not a fundamentally new metjod or kibd of thing.
It seems like you are implying that models will follow Moore’s law, but as someone working on “agents” I don’t see that happening. There is a limitation with how much can be encoded and still produce things that look like coherent responses. Where we would get reliable exponential amounts of training data is another issue. We may get “ai” but it isn’t going to be based on llms
You can’t predict how the next twenty years of research improves on the current techniques because we haven’t done the research.
Is it going to be specialized agents? Because you don’t need a lot of data to do one task well. Or maybe it’s a lot of data but you keep getting more of it (robot movement? stock market data?)
We do already know about model collapse though, genai is essentially eating its own training data. And we do know that you need a TON of data to do even one thing well. Even then it only does well on things strongly matching training data.
Most people throwing around the word agents have no idea what they mean vs what the people building and promoting them mean. Agents have been around for decades, but what most are building is just using genai for natural language processing to call scripted python flows. The only way to make them look coherent reliably is to remove as much responsibility from the llm as possible. Multi agent systems are just compounding the errors. The current best practice for building agents is “don’t use a llm, if you do don’t build multiple”. We will never get beyond the current techniques essentially being seeded random generators, because that’s what they are intended to be.
It might, but:
Stockfish on ancient hardware will still mop up any human GM
Umm… ok, but that’s a bit beside the point?
Unless you mean to include those 1980 computers, in which case stockfish won’t run on that… More than about 10 year old home computer would likely be unable to run it.
Only because they are not 32 bit so they won’t support enough RAM. But a processor from the 90s could, even though none of the programs of the time were superhuman on commodity hardware.
The chess programs improved so much that even running with 1000 times slower hardware they are still hilariously stronger than humans
Twenty years is a very long time, also “good” is relative. I give it about 2-3 years until we can run a model as powerful as Opus 4.1 on a laptop.
There will inevitably be a crash in AI and people still forget about it. Then some people will work on innovative techniques and make breakthroughs without fanfare