Comment by nikisil80
5 days ago
Yeah, I do tend to have a rather nihilistic view on things, so apologies.
I really think we're just cooked at this point. The amount of people (some great friends whom I respect) that have told me in casual conversation that if their LLM were taken from them tomorrow, they wouldn't know how to do their work (or some flavour of that statement) has made me realize how deep the problem is.
We could go on and on about this, but let's both agree to try and look inward more and attempt to keep our own things in order, while most other people get hooked on the absolute slop machine that is AI. Eventually, the LLM providers will need to start ramping up the costs of their subscriptions and maybe then will people start clicking that the shitty code that was generated for their pointless/useless app is not worth the actual cost of inference (which some conservative estimates put out to thousands of dollars per month on a subscription basis). For now, people are just putting their heads in the sand and assuming that physicists will somehow find a way to use quantum computers to speed up inference by a factor of 10^20 in the next years, while simultaneously slashing its costs (lol).
But hey, Opus 4.5 can cook up a functional app that goes into your emails and retrieves all outstanding orders - revolutionary. Definitely worth the many kWh and thousands of liters of water required, eh?
Cheers.
A couple of important points you should consider:
1. The AI water issue is fake: https://andymasley.substack.com/p/the-ai-water-issue-is-fake (This one goes into OCD-levels of detail with receipts to debunk that entire issue in all aspects.)
2. LLMs are far, far more efficient than humans in terms of resource consumption for a given task: https://www.nature.com/articles/s41598-024-76682-6 and https://cacm.acm.org/blogcacm/the-energy-footprint-of-humans...
The studies focus on a single representative task, but in a thread about coding entire apps in hours as opposed to weeks, you can imagine the multiples involved in terms of resource conservation.
The upshot is, generating and deploying a working app that automates a bespoke, boring email workflow will be way, way, wayyyyy more efficient than the human manually doing that workflow everytime.
Hope this makes you feel better!
> 2. LLMs are far, far more efficient than humans in terms of resource consumption for a given task: https://www.nature.com/articles/s41598-024-76682-6 and https://cacm.acm.org/blogcacm/the-energy-footprint-of-humans...
I want to push back on this argument, as it seems suspect given that none of these tools are creating profit, and so require funds / resources that are essentially coming from the combined efforts of much of the economy. I.e. the energy externalities here are monstrous and never factored into these things, even though these models could never have gotten off the ground if not for the massive energy expenditures that were (and continue to be) needed to sustain the funding for these things.
To simplify, LLMs haven't clearly created the value they have promised, but have eaten up massive amounts of capital / value produced by everyone else. But producing that capital had energy costs too. Whether or not all this AI stuff ends up being more energy efficient than people needs to be measured on whether AI actually delivers on its promises and recoups the investments.
EDIT: I.e. it is wildly unclear at this point that if we all pivot to AI that, economy-wide, we will produce value at a lower energy cost, and, even if we grant that this will eventually happen, it is not clear how long that will take. And sure, humans have these costs too, but humans have a sort of guaranteed potential future value, whereas the value of AI is speculative. So comparing energy costs of the two at this frozen moment in time just doesn't quite feel right to me.
These tools may not be turning a profit yet, but as many point out, this is simply due to deeply subsidized free usage to capture market share and discover new use cases.
However, their economic potential is undeniable. Just taking the examples in TFA and this sub-thread, the author was able to create economic value by automating rote aspects of his wife's business and stop paying for existing subscriptions to other apps. TFA doesn't mention what he paid for these tokens, but over the lifetime of his apps I'd bet he captures way more value than the tokens would have cost him.
As for the energy externalities, the ACM article puts some numbers on them. While acknowledging that this is an apples/oranges comparison, it points out that the training cost for GPT-3 (article is from mid-2024) is about 5x the cost of raising a human to adulthood.
Even if you 10x that for GPT-5, that is still only the cost of raising 50 humans to adulthood in exchange for a model that encapsulates a huge chunk of the world's knowledge, which can then be scaled out to an infinite number of tasks, each consuming a tiny fraction of the resources of a human equivalent.
As such, even accounting for training costs, these models are far more efficient than humans for the tasks they do.
2 replies →
> I want to push back on this argument, as it seems suspect given that none of these tools are creating profit, and so require funds / resources that are essentially coming from the combined efforts of much of the economy. I.e. the energy externalities here are monstrous and never factored into these things, even though these models could never have gotten off the ground if not for the massive energy expenditures that were (and continue to be) needed to sustain the funding for these things.
While it is absolutely possible, even plausible, that the economics of these models and providers is the next economic crash in waiting, somewhere between Enron (at worst, if they're knowingly cooking books) or Global Financial Crisis (if they're self-delusional rather than actively dishonest), we do have open-weights models that get hosted for money, that people play with locally if they're rich enough for the beefy machines, and that are not too far behind the SOTA as to suggest a difference in kind.
This all strongly suggests that the resource consumption per token by e.g. Claude Code would be reasonably close to the list price if they weren't all doing a Red Queen race[0], running as hard as they can just to retain relevant against each other's progress, in an all-pay auction[1] where only the best can ever hope to cash anything out and even that may never be enough to cover the spend.
Thing is, automation has basically always done this. It's more of a question of "what tasks can automation actually do well enough to bother with?" rather than "when it can, is it more energy efficient than a human?"
A Raspberry Pi Zero can do basic arithmetic faster than the sum total performance of all 8 billion living humans, even if all the humans had trained hard and reached the level of the current world record holder, for a tenth of the power consumption of just one of those human's brains, or 2% of their whole body. But that's just arithmetic. Stable Diffusion 1.5 had a similar thing, when it came out the energy cost to make a picture on my laptop was comparable with the calories consumed while typing in a prompt for it… but who cares, SD 1.5 had all that Cronenberg anatomy, what matters is when the AI is "good enough" for the tasks against which it is set.
To the extent that Claude Code can replace a human, and the speed at which it operates…
Well, my experiments just before Christmas (which are limited, and IMO flawed in a way likely to overstate the current quality of the AI) say the speed of the $20 plan is about 10 sprints per calendar month, while the quality is now at the level of a junior with 1-3 years experience who is just about to stop being a junior. This means the energy cost per unit of work done is comparable with the energy cost needed to have that developer keep a computer and monitor switched on long enough to do the same unit of work. The developer's own body adds another 100-120 watts to that from biology, even if they're a free-range hippie communist who doesn't believe in money, cooked food, lightbulbs, nor having a computer or refrigerator at home, and who commutes by foot from a yurt with neither AC nor heating, ditto the office.
Where the AI isn't good enough to replace a human, (playing Pokemon and managing businesses?) it's essentially infinitely more expensive (kWh or $) to use the AI.
Still, this does leave a similar argument as with aircraft: really efficient per passenger-kilometre, but they enable so many more passenger-kilometres than before as to still sum to a relevant problem.
[0] https://en.wikipedia.org/wiki/Red_Queen%27s_race
[1] https://en.wikipedia.org/wiki/All-pay_auction
> For now, people are just putting their heads in the sand and assuming that physicists will somehow find a way to use quantum computers to speed up inference by a factor of 10^20 in the next years, while simultaneously slashing its costs (lol).
GPT-3 Da Vinci cost $20/million tokens for both input and output.
GPT-5.2 is $1.75/million for input and $14/million for output
I'd call that pretty strong evidence that they've been able to dramatically increase quality while slashing costs, over just the past ~4 years.
Isn't that kind of related with the amount of money thrown at the field? If the economy gets worse for any reason, do you think that we can still expect these level of cutting costs in the future?
> But hey, Opus 4.5 can cook up a functional app that goes into your emails and retrieves all outstanding orders - revolutionary. Definitely worth the many kWh and thousands of liters of water required, eh?
The thing is in a vacuum this stuff is actually kinda cool. But hundreds of billions in debt-financed capex that will never seen a return, and this is the best we’ve got? Absolutely cooked indeed.