I run Claude Code most of my day. agent loops firing all day, one after another. the usage screen tells me I'm at 26% of my weekly limit. it doesn'...
For further actions, you may consider blocking this person and/or reporting abuse
I don't want to be a luddite, but what are really the pros and cons of this whole AI hype, can we have a nuanced (or maybe not so nuanced) discussion about that?
'Pros':
some people get filthy rich (is that really a pro? I've put it here as a "straw man's argument", lol)
we can spit out code 10 times faster with AI (does anyone actually need that much code? are we still able to understand how it works?)
managers at companies are getting so happy, coz they can fire people and cut costs (at least that's what they think ...)
'Cons':
the toll on the environment is insane - not to mention the "mental toll", which has been the topic of quite a few articles here on dev.to ...
computers, smartphones and game consoles have been getting a lot more expensive, because AI data centers gobble up all the RAM chips
widening the already huge gap between haves and have-nots (the point you're talking about mainly)
But I'm sure AI can be beneficial to mindkind - curing cancer, fundamental scientific breakthroughs, solving pressing societal issues ...
It can't be stopped really, but it all comes down to: what will we use it for, and how will we use it?
In fairness. Nothing is stopping you from creating 30 google accounts and switching between them to make the most out of AI Studio and Antigravity. So the have-not's still have, if they're willing to be smart about it.
As for the filthy rich part, yeah... But worth noting those people are the investors in these AI companies and you can always invest yourself. What it has done though, is allow startups to scale FAST and people from grass-roots build a functional business with freebies like Azure and Google Cloud startup trials ($1000 and $300 respectively in cloud credits).
It has opened doors for people, if you were an app dev, but knew nothing about cloud run, you can finally deploy to cloud.
Suddenly language specific tasks become specialty agnostic. You dont need to be able to write a F# function to understand it, it makes sense reading it, but to conform to the strict requirements, it's a massive undertaking to learn (it took me 5 years just to call myself 'proficient', nowhere near expert).
As for the environment, yes, it uses ALOT of power, it takes up alot of real estate and it uses alot of computers, kicking out alot of heat. But consider this... Data Centers use water to cool them, that water is vaporized and emitted into the air. Generally speaking, the piping used is copper for better heat dissipation which isnt all that bad, it's anti-microbial. That water vapor emitted raises humidity and in turn either fixes greenhouse gasses and drops to the earth (imo, marginally better than building up in the outer atmosphere), humidity also raises chance of rain. Right now (I live in a desert), the world needs more rain, so that aspect, it's actually the first time we're producing more heat, but the outcome is actually good?
The way I see it, is AI, is like computers. When they came, it created a divide, with 1 difference... When computers came out, you needed to fork over half a year's income just to buy 1. Whereas to access AI and use it, you dont need a cent, just a phone or a computer. When you look at how the industry is changing, existing companies are leaning down, while startups are scaling up. The shift in dynamics is essentially a goliath slaying event, where small companies scale fast enough that they can actually compete against giants. Take the 12 part series I did, it took me just about 5 days to scale from building an extension to building a full-fledged OS running it's own custom language at a kernel level... That's like the guy with the computer designing a fighter jet in a week, vs pen and paper took a year.
More code isnt necessarily better, but when you look at the displacement caused by it, people are actually doing unit tests now... People are actually securing their endpoints now... People are shifting how they design, because their fatal flaws were exposed by AI. That's global growth, accessible with a free account. The only real burden of it, is whether you choose to embrace it, or ignore it till you're obsolete? How many pen and paper employees rejected computers, till they became the industry standard in every single profession. That's how it's going with AI.
Water vapor is actually a pretty potent greenhouse gas. More water in the atmosphere contributes to runaway climate changes, which, in turn, makes weather phenomena like rain more extreme. Heat waves cause extreme droughts, leading to a sudden increase in water in the atmosphere (rivers and lakes evaporating), which leads to extreme rain fall, which can't be absorbed into the dry ground properly (because it's too much at once), leading to surface flow and - in turn - to inland floods, causing entire cities to be flooded.
Also, the energy consumption isn't free, either: Rapidly increased energy consumption is often met with burning more fossil fuels, because building new nuclear power plants takes ages (not to speak of the radioactive waste produced, the amount of energy necessary to build it and how we get the uranium out of the ground to begin with) and most countries are somewhat slow with implementing solar power and wind power because they don't like how it looks.
The heat generated is the least problem here. It's how we deal with the massively increased energy consumption. The fact that everybody and their mother are building server farms like crazy to cover the immense token consumption doesn't exactly help the climate situation.
Pascal, the backfill point is the one that matters. New demand doesn't wait for clean supply, it gets met with whatever's fastest to bring online, and that's still coal and gas in most grids.
I studied geophysics before any of this . the water cycle math you're describing isn't new to me, just usually not the frame people reach for when they're impressed by a chatbot. Nigeria's 40 TWh a year isn't landing on any of these plants either way.
About this climate issue — I feel like I've been hearing it since I was a kid. Back then the news was flooded with the Antarctic ozone hole, global warming, greenhouse effect. How come you barely hear about them anymore?
Pascal. well said, very well explained - I noticed the remarks about water vapor, and already thought it didn't make sense, but forgot to comment on it ...
I don't doubt the potential benefits of AI, but when are we, as humanity, going to prioritize saving the environment on our precious little planet, our only home?
First things first - AI is great, but right now its environmental fallout is horrible - I wouldn't mind at all if we decide on a moratorium on the AI datacenter craze, and first focus on saving this planet - well I know it ain't gonna happen, but one can dream ...
P.S. gotta disagree with the suggestion that climate change is a fad, just because some disasters which were predicted haven't happened yet - I think we're still underestimating the scale of what's coming ...
The real tragedy here, in my opinion, is that many of us (myself included at times) aren't using the most appropriate tool for the job. We're flying a Boeing 777 200 miles to deliver a single passenger, when a Cessna 172 or a Honda Civic would be equally as effective.
The cloud-default design that has emerged routes surprisingly simple tasks through multiple network hops before they arrive at huge computer power for processing. Things like reading files, making small edits, validating data, running basic code analysis are often workloads that could be handled much closer to where the data already exists.
It makes me even more convinced that "local-first" is not just a privacy and security concern. It's becoming a thermodynamic necessity. A sovereign system escalates computation only when the complexity of the task justifies the cost of the escalation.
If we can run SLMs locally, get good results, and use a fraction of the energy of the API call, isn't that a win? I'm seeing a future in which we aren't cloud-only or local-only but a more thoughtful matching of the physical compute power needed for the task at hand.
Great post, Daniel. Thanks for sharing it.
Ken, the escalation point is the mechanism worth sitting with . most tasks don't need the 777, they get put on it because routing everything through one API is simpler to build than deciding per-task.
That's not a local vs. cloud question, it's a "who pays to make the decision" question and right now nobody's paying for it, so the 777 flies every route....
Another reason why I think using AI for every little task is simply wrong. It's just the same as Taylor Swift going on her plane to Paris to get a dress for a concert. In the end, it all comes down to your own sense of morality and what you can resign from for the bigger sake. Like resigning from fast fashion or makeup to not support sweatshops and child labor. I think AI brought a new level of moral dilemmas.
This trend I also cannot understand.
"You just don't see the meter" is the line. But a lot of that 41 Wh isn't the work, it's the loop being sloppy, re-reading the same file, retrying, dragging full context every step. Token and watt are the same meter, and most agent runs I instrument burn 2 to 4x what the task actually needed. Won't fix the Port Harcourt grid from my keyboard, but that waste is the one thing the person firing the loop controls.
Vinicius, the 2-4x is the part I haven't instrumented and should. But there's a ceiling on how much of this is fixable at the loop level .
you can cut re-reads and retries and still be running a 70B model for a task a 7B could've handled, still routing through us-east-1 because that's where the API lives. Waste inside the loop is yours to fix. The floor under the loop isn't.
Fair. Region I concede flat, you use the API where it lives. Model size is the one I'd still call half yours, right-sizing which step runs on which model is a caller call most loops skip. But the floor is real, and the honest version of my point is just: measure it so the floor stops being the alibi for the loop.
Vinicius, fair enough. model size is the one I've actually skipped not just underused. Most of what I run defaults to whatever's already warm in the session rather than what the step needs.
"The floor isn't the alibi" is the sentence I'd want in front of me next time I reach for the bigger model out of habit instead of requirement.
What if I told you I just used an LLM to translate your article? Would you believe me?
After reading your piece, all I can say is — everyone really is standing on very different ground. There's an old saying where I'm from: Trees die when moved; people thrive when they move. We still need to invest more in infrastructure. If you want wealth, build the roads first.
the translation makes the point better than the article did . you used the tool to close distance, I used it to notice how far the grid still is from doing the same. I'd flip the order tho: roads don't get built because people migrate toward them, they get built because someone pays for the wire before anyone needs it yet.
The title caught my attention.
Nice analogy BTW!
Worldwide, over 51,000 of us have mitigated that waste by up to 20x, keeping over 50 tons of CO2 out of the atmosphere and avoiding the burn of over a third of a **trillion **tokens. And that's only since March.
I won't turn this reply into an infomercial. I'll just offer this: It doesn't HAVE to be this way...
-J. Gravelle
GSF Certified Green Software Practitioner - jMunch, LLC
The part about not apologizing for the sessions but being done pretending they cost nothing — that's the only honest position to take. The discomfort is the point.
I've been building agent tooling and thinking about efficiency in terms of latency, token cost, and accuracy. Energy is almost never in the design constraints. Reading this made me realize that 'cost optimization' in agent design has a dimension almost nobody is measuring at the individual session level.
The 130x multiplier between a chat reply and an agent session is the number that sticks. It suggests most of the energy isn't in generating tokens — it's in the reasoning loops, the retries, the tool calls that go nowhere. If that's where the waste lives, then the efficiency lever isn't smaller models, it's fewer unnecessary iterations. Tighter pre-condition checks, better task decomposition so the agent doesn't wander, knowing when to stop before spending another 300 Wh on a dead end.
This really puts AI usage into perspective. Agent workflows are incredibly useful, but they also remind us that efficiency matters just as much as capability. As developers, we should build tools that solve real problems without wasting unnecessary compute.
At CodeCan.net, we're also thinking about practical development—helping developers ship faster with ready-to-use scripts and applications while encouraging smarter, more efficient workflows. The future of AI shouldn't just be more powerful; it should also be more responsible.
This is a conversation the AI industry needs more of. We often measure AI by intelligence or speed, but resource efficiency deserves equal attention. Smarter workflows, better model routing, and reducing unnecessary agent loops can improve both cost and sustainability. Thanks for bringing attention to the hidden side of AI.
We often think about AI in terms of speed and productivity, but rarely about the energy behind every agent loop. The contrast between always-on infrastructure and places where electricity isn't guaranteed is a reminder that technology doesn't exist in the same reality for everyone. Efficiency should be part of the conversation as AI adoption grows.
Your post makes a point most of us rarely think about: AI has a real physical cost, not just a token cost. At the same time, context matters, those energy estimates vary a lot by model, hardware, and workload, so it's worth being careful with broad comparisons. The takeaway for me isn't "don't use AI" it's to use it intentionally instead of treating unlimited compute like it's free.
This really hit home for me. I've been working on making my own tools
as lightweight as possible because of this exact reason. It started
with realizing how much bloat most software carries, and now I obsess
over keeping things under 30MB of RAM.
The "feels free" trap is so real. We click a button and have no idea
what it cost in energy, water, or someone else's grid somewhere else.
Thanks for putting numbers on something most of us feel but can't measure.
This gave me something to think about, Daniel.
We often talk about the monetary cost of AI, but rarely about the energy cost behind every prompt and agent loop. Regardless of the exact numbers, it's a good reminder that "feels free" doesn't mean "is free." Thanks for sharing this perspective.
AI’s hidden costs deserve more attention. The goal shouldn’t be using more compute, but using the right amount of intelligence for the right problem.
This raises some important points. In practice, I've found that the key is balancing theoretical best practices with pragmatic trade-offs — what works in a blog post doesn't always survive contact with a legacy codebase.
It’s just like during the Industrial Revolution, when a huge number of factories and steam engines replaced workers, and air pollution lingered for a long time. It’s hard for us to change the world.
Dope Stuff. Interested to learn
Different grid, same planet； AI token freedom
surprise
If you had written this post yourself instead of using AI, you would have saved a lot of energy to send to those in need!
hhhhhhhh
Who can do without LLMs these days?
The "marketable fluff" paradox is a brutal reality check. The 99% slop isn't a byproduct; it's the actual product sold to simulate adoption while burning global grids.
hhhhhhhhhh
Great work!