These young people have seen less in their life, and a much larger percentage of their life was filled with the advertising of the AI companies. So no wonder that they are a little bit slower seeing the limitations of the AI models.
I don't think so. It's a part of general and wide-ranging technical ignorance.
They know how to navigate through the intricate settings of their favourite social app but debugging connection issues is too much (even on a very basic level, "can my browser access the same site?").
Stating "Software engineers are all-in on AI" because of an increase in github projects being created is hilarious. I didn't realise creating a github repo made someone a software engineer. If only I had known this I wouldn't have bothered learning all the other stuff!
> The report estimates that training the latest frontier large language models, such as xAI’s Grok 4, can generate over 72,000 tons of carbon-equivalent emissions.
That seems pretty trivial, relative to 38bn per year globally?
The training of one LLM requires as much emissions as 17,000 people over a year. Which, according to the article, is 8 times more than last year, and may be underestimated by a factor 2.
That does not cover the whole usage: the hardware, the bots that collect learning data, the prompts, etc. And there are now many models of this size, and thousands and thousands at smaller sizes. And some of this parameters are increasing.
AI is estimated to emit more than 80e6 tons of CO2-equivalent this year. Much more than whole countries like Austria or Israel. Is that trivial?
How much does that come out to per user or per request? And how does that compare to anything else those people do? Like drive 10 minutes. Or eat a burger.
These numbers keep being put up as large in absolute terms but that’s deceiving for the average person who doesn’t have a way to compare them to something relevant in their lives.
Another way to put it: if training a model cost 72,000 tons of carbon, and it then gets used by 100 million people (typical of major models), the cost per person is 0.00072 tons.
Per the article, the average human uses over 5 tons per year (Americans: 18). Adding 0.00072 to 5 is not really noticeable.
Yeah it's basically nothing despite the fact that xAI seemed to intentionally crank up the carbon intensity for no reason.
Also, hilarious to select 2 major models from 2025 and they're both Grok, almost certainly the least useful, least used, and least interesting of that year.
Of course they will. Tokens are valuable, you can always spend a finite budget on specialized tokens or fewer and higher quality tokens, size of user base and engagement gives you a flywheel moat that is difficult for newcomers to compete with. The market is complex and easy to oversimplify.
My new startup tokencoin will blah blah blah exchange rate, (something AI writes here), 3. profit (more AI), benefiting all human kind and helping our users scale up their productive intelligence!
It's hard and complex to enter any mature market. The vast majority of firms that attempt to enter a new market fail. LLM's have no more than this normal moat.
If every market has a moat, then saying that a particular market has a moat is a statement without meaning. The OP was probably not trying to make a meaningless statement, therefore the OP was probably saying that LLM's don't have an abnormally effective moat.
I agree, LLM's don't have an abnormally effective moat, just the standard moat most mature markets have due to market complexity. IOW, LLM's will likely end up with the standard oligopoly most modern western markets end up in, which have minor but relatively ineffective pricing power.
Isn’t capital and momentum a moat? Sure Chinese models use distillation but I don’t see them training models from scratch. At least not today. But maybe as chips get cheaper and they have Chinese made ones?
Apparently not much of one. There are, what, 5 or more companies with frontier models? And open weights models like MiniMax are snapping at their heels
There are many markets where open source has been nipping at heels for a long time.
Obviously product areas differ for reasons structural and happenstance. But there is definitely a pattern that occurs, where open source fast follows commercial advances, benefiting from having a clear target to develop for.
Which is of course, a great service. Even if it never unseats the commercial version, it forces the owners to reinvest more in improvements, by undermining their moats. As well as providing a much better value alternative version for many people.
And it does not even consider that e.g. the EU might one day decide that AI should be for everyone, thus releasing a heavily subsidized open source model.
Or that at some point AI is good enough, and so at that point any model will do.
I’m not technically familiar but I remember someone saying that models like MiniMax basically skip the cost of training by using distillation to basically “steal” the models from OpenAI or Anthropic, and that these companies now have various defenses against this. What happens when MiniMax has to do the full work themselves?
> The report estimates that carbon emissions from models with the least efficient inference are over 10 times as high as those with the most efficient inference. DeepSeek’s V3 models were estimated to consume around 23 watts when responding to a “medium-length” prompt, while Claude 4 Opus was estimated to consume about 5 watts.
This makes absolutely no sense. I suppose they meant watt hours, and that's a weird way to explain carbon emissions...
"On the other hand, Perrault noted that 'Epoch AI independently estimates Grok 4’s emissions to be significantly higher at approximately 140,000 tons of CO₂.'"
I realize these are still estimates, but when the other independent analysis nearly doubles the outcome I'm not left feeling optimistic. One could argue some numbers from others are underestimates... which of course just bums me out all the more!
I agree 100% if those are the only two options. I guess my point is that it's fair to assume that Elon's crew is doing the bare minimum in terms of efficiency and pollutant mitigation-- at least when compared to other data centers who do legally compliant business with real power companies.
The "China leads in robotics" seems to be unaffected by AI. The China line is basically on the same trajectory since 2012. The chart does no belong in the article.
> The capabilities of AI models have improved with incredible speed over the past decade, and as the graph above shows, progress seems to be accelerating.
errr… no? Every discipline is clearly hitting a plateau so far. Some started recently and hence haven’t yet (competition maths) but based on past graph, they will all plateau.
That's the lead in industrial robot installed. That lead is understandable because of manufacturing concentration in China. Here are 10 top robot makers, none of them are Chinese (*), and five are Japanese:
The graph says "new industrial robots installed", which is a bit misleading. For example the newest BYD factories are still stuffed with German/Japanese robots.
They also lead the world in EV production on paper, but in practice a large portion of those numbers might be driven by government pressure, not actual demand [1].
I’d personally take this data with a big grain of Goodhart’s law.
What's worse is that this the predictable result of a choice that America made decades ago and continues to make.
Outsourcing manufacturing capacity to China and letting domestic manufacturing skills atrophy and institutional knowledge die out was a choice that many people opposed but were ultimately helpless to stop because the people making the decisions ignored them and did it anyways for personal gain is how we got here.
You'd think that the supply chain shocks that we saw during COVID would be a wake up call that would have jolted people into action.
You'd think that Ukraine-Russia war would have been a wake up call that would have jolted people into action.
You'd think that the recent failures by the US military in Iran and the depletion of years of missile stockpiles would have been a wake up call that would have jolted people into action.
I'm at a loss to explain it. It's like the American oligarchs want to weaken America, or at least are willing to do so if it means that they have greater control over it. Maybe they don't care about manufacturing capacity because they know that America is ultimately a nuclear protected island and that even if things continue to decline they'll be safe to rule it like a king?
> It's like the American oligarchs want to weaken America, or at least are willing to do so if it means that they have greater control over it.
The capital holders want it under their control. The fact that it harms the state is a consequence they ignore, or worse, believe that other people will deal with. There is not thought given to how much harm will be caused, because the harm is seen as part of the process used to acquire that control. It's the sort of thinking that aligns with beating a dog to teach it not to bark and then ignoring the cataracts that form from the repeated blows.
Plus that graph is the first derivative of industrial robots. the actual # of new robots since 2012 is the area under the respective curves, so a very big lead.
Worth calling out AI sentiment among young people is not nearly so rosy: https://news.gallup.com/poll/708224/gen-adoption-steady-skep...
That's temporary. They will adapt and find ways to use it to its full potential - just like it happened with every new technological shift in history.
Would you mind asking your crystal ball some other questions - like what those ways of using it are exactly?
when exactly did that happen ?
cause up until now I have observed the exact opposite which is coherent with expectations: https://coding2learn.org/blog/2013/07/29/kids-cant-use-compu...
Why are you adding 's' for a HTTP-only server?
don't the young usually pick up new tech faster?
really no: https://coding2learn.org/blog/2013/07/29/kids-cant-use-compu...
My Firefox doesn't accept your https cert. Maybe check that out?
There isn't a cert.. https://www.ssllabs.com/ssltest/analyze.html?d=coding2learn....
archive.today suggests, there's never been (The only https returns 403 in 2015, the 2013 links are http) https://archive.is/https://coding2learn.org/
The domain has been mentioned on HN before (without TLS), this account seems to be just messing up the links (replace https with http to see the page)
Computers are old tech nowadays...
Only if they're interested.
I don't see that.
bullshit. you hear that you are not needed, you data is not yours. the AI lovers thinking: "humans also consume energy".
This seems to be a slow discovering of the inherent limitations.
These young people have seen less in their life, and a much larger percentage of their life was filled with the advertising of the AI companies. So no wonder that they are a little bit slower seeing the limitations of the AI models.
Or the realization that you will lose your job
I don't think so. It's a part of general and wide-ranging technical ignorance.
They know how to navigate through the intricate settings of their favourite social app but debugging connection issues is too much (even on a very basic level, "can my browser access the same site?").
Stating "Software engineers are all-in on AI" because of an increase in github projects being created is hilarious. I didn't realise creating a github repo made someone a software engineer. If only I had known this I wouldn't have bothered learning all the other stuff!
I agree with you on that metric being not great - I would have definitely swapped it for this:
"Claude Code GitHub Commits Over Time" https://newsletter.semianalysis.com/p/claude-code-is-the-inf...
Sure - also an imperfect metric. But less imperfect? And more indicative of... something? Not nothing?
> The report estimates that training the latest frontier large language models, such as xAI’s Grok 4, can generate over 72,000 tons of carbon-equivalent emissions.
That seems pretty trivial, relative to 38bn per year globally?
The training of one LLM requires as much emissions as 17,000 people over a year. Which, according to the article, is 8 times more than last year, and may be underestimated by a factor 2.
That does not cover the whole usage: the hardware, the bots that collect learning data, the prompts, etc. And there are now many models of this size, and thousands and thousands at smaller sizes. And some of this parameters are increasing.
AI is estimated to emit more than 80e6 tons of CO2-equivalent this year. Much more than whole countries like Austria or Israel. Is that trivial?
How much does that come out to per user or per request? And how does that compare to anything else those people do? Like drive 10 minutes. Or eat a burger.
These numbers keep being put up as large in absolute terms but that’s deceiving for the average person who doesn’t have a way to compare them to something relevant in their lives.
Another way to put it: if training a model cost 72,000 tons of carbon, and it then gets used by 100 million people (typical of major models), the cost per person is 0.00072 tons.
Per the article, the average human uses over 5 tons per year (Americans: 18). Adding 0.00072 to 5 is not really noticeable.
(There is also the cost of inference, of course.)
Yeah it's basically nothing despite the fact that xAI seemed to intentionally crank up the carbon intensity for no reason.
Also, hilarious to select 2 major models from 2025 and they're both Grok, almost certainly the least useful, least used, and least interesting of that year.
Also nobody will ever have a moat, so the graph of investor stupidity is going through the roof.
Of course they will. Tokens are valuable, you can always spend a finite budget on specialized tokens or fewer and higher quality tokens, size of user base and engagement gives you a flywheel moat that is difficult for newcomers to compete with. The market is complex and easy to oversimplify.
My new startup tokencoin will blah blah blah exchange rate, (something AI writes here), 3. profit (more AI), benefiting all human kind and helping our users scale up their productive intelligence!
It's hard and complex to enter any mature market. The vast majority of firms that attempt to enter a new market fail. LLM's have no more than this normal moat.
Well yes that’s my point: AI does not suddenly do away with the market.
If every market has a moat, then saying that a particular market has a moat is a statement without meaning. The OP was probably not trying to make a meaningless statement, therefore the OP was probably saying that LLM's don't have an abnormally effective moat.
I agree, LLM's don't have an abnormally effective moat, just the standard moat most mature markets have due to market complexity. IOW, LLM's will likely end up with the standard oligopoly most modern western markets end up in, which have minor but relatively ineffective pricing power.
Isn’t capital and momentum a moat? Sure Chinese models use distillation but I don’t see them training models from scratch. At least not today. But maybe as chips get cheaper and they have Chinese made ones?
> Isn’t capital and momentum a moat?
Apparently not much of one. There are, what, 5 or more companies with frontier models? And open weights models like MiniMax are snapping at their heels
There are many markets where open source has been nipping at heels for a long time.
Obviously product areas differ for reasons structural and happenstance. But there is definitely a pattern that occurs, where open source fast follows commercial advances, benefiting from having a clear target to develop for.
Which is of course, a great service. Even if it never unseats the commercial version, it forces the owners to reinvest more in improvements, by undermining their moats. As well as providing a much better value alternative version for many people.
And it does not even consider that e.g. the EU might one day decide that AI should be for everyone, thus releasing a heavily subsidized open source model.
Or that at some point AI is good enough, and so at that point any model will do.
I’m not technically familiar but I remember someone saying that models like MiniMax basically skip the cost of training by using distillation to basically “steal” the models from OpenAI or Anthropic, and that these companies now have various defenses against this. What happens when MiniMax has to do the full work themselves?
Why would they have to do it themselves?
>Chinese models use distillation but I don’t see them training models from scratch
Maybe because they don't have to. If someone is doing the heavy work and they can take output of that, it's a win for them.
[dupe]
https://news.ycombinator.com/item?id=47758028
Source: https://hai.stanford.edu/ai-index/2026-ai-index-report
> The report estimates that carbon emissions from models with the least efficient inference are over 10 times as high as those with the most efficient inference. DeepSeek’s V3 models were estimated to consume around 23 watts when responding to a “medium-length” prompt, while Claude 4 Opus was estimated to consume about 5 watts.
This makes absolutely no sense. I suppose they meant watt hours, and that's a weird way to explain carbon emissions...
Besides the lead in robotics for China, those Grok emissions charts are the thing that most leap off the page.
"These estimates should be interpreted with caution. In the case of Grok, they rely heavily on inferred inputs drawn from public reporting"
That chart doesn't really pass the sniff test.
The rest of the quote you began continues:
"On the other hand, Perrault noted that 'Epoch AI independently estimates Grok 4’s emissions to be significantly higher at approximately 140,000 tons of CO₂.'"
I realize these are still estimates, but when the other independent analysis nearly doubles the outcome I'm not left feeling optimistic. One could argue some numbers from others are underestimates... which of course just bums me out all the more!
I don't know if I would want to do too much sniffing around the Methane power they are using over at xAI.
https://www.theguardian.com/us-news/2025/jul/03/elon-musk-xa...
That's definitely a very visible use of carbon generating fuel, but I'd choose methane over coal power plants all day.
I agree 100% if those are the only two options. I guess my point is that it's fair to assume that Elon's crew is doing the bare minimum in terms of efficiency and pollutant mitigation-- at least when compared to other data centers who do legally compliant business with real power companies.
The "China leads in robotics" seems to be unaffected by AI. The China line is basically on the same trajectory since 2012. The chart does no belong in the article.
While chatGPT was not out then, the ML that drives robotics was acting by then very much.
> Training AI models can generate enormous carbon emissions
Sure, but what I'd really like to see is a graph for how much carbon is generated serving these models globally.
Profits generated by AI: <not graphed>
The absence speaks volumes.
There hasn't been one dollar of profit from any company, it's more a battle of how low you can keep your losses
I still don't understand the State of AI in 2026.
> The capabilities of AI models have improved with incredible speed over the past decade, and as the graph above shows, progress seems to be accelerating.
errr… no? Every discipline is clearly hitting a plateau so far. Some started recently and hence haven’t yet (competition maths) but based on past graph, they will all plateau.
China’s robotics lead holy cow.
That's the lead in industrial robot installed. That lead is understandable because of manufacturing concentration in China. Here are 10 top robot makers, none of them are Chinese (*), and five are Japanese:
https://manufacturingdigital.com/top10/top-10-industrial-rob...
(*) Kuka was a top German maker who got acquired by Chinese company Midea recently
The graph says "new industrial robots installed", which is a bit misleading. For example the newest BYD factories are still stuffed with German/Japanese robots.
They also lead the world in EV production on paper, but in practice a large portion of those numbers might be driven by government pressure, not actual demand [1].
I’d personally take this data with a big grain of Goodhart’s law.
[1]: https://www.bloomberg.com/features/2023-china-ev-graveyards/
What's worse is that this the predictable result of a choice that America made decades ago and continues to make.
Outsourcing manufacturing capacity to China and letting domestic manufacturing skills atrophy and institutional knowledge die out was a choice that many people opposed but were ultimately helpless to stop because the people making the decisions ignored them and did it anyways for personal gain is how we got here.
You'd think that the supply chain shocks that we saw during COVID would be a wake up call that would have jolted people into action.
You'd think that Ukraine-Russia war would have been a wake up call that would have jolted people into action.
You'd think that the recent failures by the US military in Iran and the depletion of years of missile stockpiles would have been a wake up call that would have jolted people into action.
I'm at a loss to explain it. It's like the American oligarchs want to weaken America, or at least are willing to do so if it means that they have greater control over it. Maybe they don't care about manufacturing capacity because they know that America is ultimately a nuclear protected island and that even if things continue to decline they'll be safe to rule it like a king?
> It's like the American oligarchs want to weaken America, or at least are willing to do so if it means that they have greater control over it.
The capital holders want it under their control. The fact that it harms the state is a consequence they ignore, or worse, believe that other people will deal with. There is not thought given to how much harm will be caused, because the harm is seen as part of the process used to acquire that control. It's the sort of thinking that aligns with beating a dog to teach it not to bark and then ignoring the cataracts that form from the repeated blows.
Plus that graph is the first derivative of industrial robots. the actual # of new robots since 2012 is the area under the respective curves, so a very big lead.
China’s manufacturing lead in a graph
Don't they have ten times more people than the next highest country (Japan) though?
It striking, but says nothing about AI.
[dead]