i like llms and artificial intelligence.
there. that's the entire post. you can close the tab now.
0x00 permission denied
i like llms.
i use them at work. i use them at home. i use them when i'm trying to understand something complicated, when i'm programming or automating something, when i have 27 thoughts in my head and can't figure out which one is supposed to come out first.
sometimes i use them to make something profoundly stupid because making something profoundly stupid is one of life's remaining affordable pleasures.
they're cool. i like them.
apparently that makes me an enemy of the human spirit.
i've spent enough time online to know the arguments. it's slop. it's destroying art. it's making everyone stupid. it's boiling the oceans so some facebook mom can generate a picture of her dog wearing sunglasses and driving a jeep.
some of those arguments are serious. some deserve answers. some describe actual harm done to actual people, and pretending otherwise would make me exactly the kind of insufferable technology evangelist i can't stand.
but some of the outrage seems directed at the fact that somebody can now do something they previously couldn't.
that bothers me.
because i remember when the internet was going to make everyone stupid. before that, computers were going to eliminate the need to learn anything. before that, calculators were going to destroy mathematics.
the calculator didn't destroy mathematics.
it made long division a less compelling measure of a person's worth.
maybe that bothers somebody.
i think there's something worth examining in our tendency to confuse making a task easier with making the person doing it less capable.
especially when the people benefiting from that assistance aren't necessarily the people who had an easy time to begin with.
0x01 confession
here's something i don't usually admit.
i was never particularly good at programming.
honestly.
i could read code. i could modify it. i could find the one line in a configuration file that was ruining somebody's entire week. i understood systems, how they worked together, what they needed to do, and usually why they weren't doing it.
but sitting down in front of an empty file and making something from nothing?
god damn.
i could have the entire architecture of an application worked out in my head. i knew what i wanted it to do, how the pieces should communicate, where the data should go, what the interface ought to look like.
and then i'd open an editor.
nothing.
the ideas were there. getting them from my head into something a compiler could understand was the part that nearly killed them.
i spent years working around that.
now i can build a little blog in an afternoon. i can prototype an idea before i've had enough time to convince myself it's stupid. i can automate something tedious without first spending three days remembering the syntax of a language i barely touch.
i can practice for interviews by having an llm pretend to be a linux system with a broken configuration, and spend an hour diagnosing problems without worrying that i'm embarrassing myself in front of another person.
sometimes it gets the simulation wrong. sometimes i have to correct it. that's part of the exercise too.
none of this means i suddenly became an exceptional programmer.
it means the distance between understanding what i want to accomplish and actually accomplishing it has gotten smaller.
that's not the machine replacing me.
that's me being able to do more of what i was already trying to do.
and i'm not going to pretend that distinction doesn't matter.
not everybody has the patience, the access, the educational background, the executive function, or the luxury of spending fourteen hours trying to understand something that could have been explained in twenty minutes.
sometimes a person has the idea but not the vocabulary.
sometimes they know exactly what they want to build but don't know how to express it in the language a compiler understands.
sometimes the documentation is in a language they don't speak.
sometimes they're exhausted.
sometimes they're a beginner.
sometimes they're fucking disabled.
and now we've built something that, imperfectly, unreliably, and occasionally with the confidence of a man explaining something he has never once understood, can help people work past some of those barriers.
why the hell would i be angry about that?
i know what it's like to be perfectly capable of understanding something and still struggle to express it in the form the world expects.
if a machine can help somebody across that gap, i think that's wonderful.
i don't think they should have to apologize for using it.
0x02 two things confused for one
there are two things people keep talking about as though they're the same.
the first is the technology itself.
the models, the weights, the inference engines, the strange mathematical machinery that takes an input and produces something that looks remarkably like an answer.
it's a tool. an odd, powerful, frequently useful, occasionally spectacularly wrong tool.
the second is what people decide to do with it.
and that second one i know quite well.
i've heard it on earnings calls. i've read it in press releases.
we are leaning into ai.
we are building for an ai-first future.
we are reallocating resources to our highest priorities.
funny how the priorities always seem to involve fewer people receiving paychecks.
the model didn't decide to lay anybody off.
a person did.
a person looked at a budget, made a decision, and signed off on it. then another person wrote a press release about artificial intelligence transforming the future of work.
the model makes a convenient explanation. it's considerably easier to tell shareholders that technology has made a position redundant than to say you've decided the company's margins would look better without the person occupying it.
and i'm not claiming automation never changes staffing needs. of course it does. it has for centuries.
i'm saying that the decision to replace somebody, retrain them, move them into another role, or distribute the benefits of increased productivity is still a human decision.
it doesn't become an act of nature because you put "ai" in the quarterly earnings presentation.
machines don't hold board meetings.
people do.
this distinction matters because whenever somebody tells me they hate artificial intelligence, i've learned to wonder what exactly they're talking about.
the technology?
the way it was developed?
the way it's being sold?
the people using it?
the executives treating it as an excuse to eliminate workers?
those are different questions. they can have different answers.
and yes, some problems are inseparable from how the technology itself was built. training data, consent, bias, privacy, and reliability are not imaginary concerns invented by people who dislike progress.
i'm not asking anyone to stop criticizing ai.
i'm asking them to be more precise about what they're criticizing.
because if what you hate is the way a corporation uses technology to justify treating people as disposable, believe me, we have something in common.
0x03 you're wrong about data centers
i want to talk about what a data center actually is.
not the abstract thing on a cloud architecture diagram. not the little cylinder someone draws in a powerpoint presentation.
the actual building.
i've worked inside them.
the cloud isn't a cloud. it's somebody else's computer.
well, usually an enormous number of somebody else's computers.
they live in huge buildings full of electrical distribution systems, switchgear, transformers, cooling equipment, and rows of generators waiting for the day the utility supply decides to have a very bad afternoon.
there are people working in those buildings. people doing maintenance. people replacing broken parts. people keeping the whole ridiculous thing running.
i've stood in the cold aisles. i've put my hands on the servers.
when somebody asks a language model to write a poem about their cat, somewhere there are accelerators, memory, networking equipment, and electrical infrastructure doing the work.
not magic.
electricity and silicon.
a lot of it.
and yes, data centers consume enormous amounts of energy. some facilities use more electricity than small towns. that's a real concern, especially when new capacity is being built faster than the surrounding electrical infrastructure can accommodate it.
but the argument i keep seeing is about water.
and i think a lot of the discussion is missing some important context.
no, every prompt isn't personally draining lake michigan.
that doesn't mean data centers don't consume water. some of them consume an astonishing amount.
google's council bluffs campus in iowa, where i once worked, is a particularly interesting example.
according to google's 2025 figures, the campus consumed approximately 1.35 billion gallons of water in a year. their own reporting compares that to about nine golf courses.
a billion gallons is a hell of a number.
i'm not going to tell you otherwise.
but the number by itself doesn't explain the engineering, the tradeoffs, or the local impact.
water is often used in evaporative cooling because evaporation is an extremely effective way of removing heat. depending on the climate and the facility design, using water can reduce the electrical energy required for cooling.
that matters when you're operating buildings that consume enormous amounts of electricity.
you can design facilities around different cooling approaches. air cooling, evaporative systems, liquid cooling, reclaimed water, and combinations of these technologies all have their places.
a closed liquid loop inside a rack, by the way, doesn't automatically mean the entire facility uses no water. you still have to reject the heat somewhere. where and how you do that matters.
and in locations with limited water resources, the tradeoffs become considerably more serious.
there's no universal answer.
a cooling strategy that makes sense in one climate might be completely inappropriate somewhere else.
which brings me back to the golf courses.
the comparison is interesting because it shows that we already accept substantial water consumption for all kinds of activities, often without giving them the same scrutiny.
but it's not a free pass.
nine golf courses worth of water is still nine golf courses worth of water. the fact that something else consumes resources doesn't make the impact disappear. local water availability, aquifer conditions, municipal infrastructure, and competing demand still matter.
that's why i get frustrated when people treat a single enormous number as though it settles the entire argument.
water withdrawal isn't the same thing as water consumption. a gallon withdrawn and returned isn't equivalent to a gallon lost through evaporation. the source of the water matters. the location matters. the cooling system matters.
we should be asking whether facilities are being built responsibly, whether utilities can support them, and whether surrounding communities are bearing costs they never agreed to.
those are good questions.
"data center uses x billion gallons, therefore artificial intelligence is evil" isn't much of an analysis.
google's council bluffs operation dates back to 2009.
for much of that time, the public discussion was very different.
0x04 slop
holy hell, there's slop.
so much slop.
the horrible facebook advertisements for local businesses. the corporate photographs of people with fucked-up hands smiling in a boardroom that doesn't exist. the illustrations where every character has the same expression and every object looks like it was assembled from a vague memory of what objects are supposed to be.
the articles written by nobody for an audience of nobody, optimized to appear in search results for questions somebody actually needs answered.
i click one of those things, spend thirty seconds reading it, realize there's absolutely nothing there, and become approximately forty percent more hostile toward the internet.
i hate that stuff too.
but here's what i think people are missing.
the internet was full of garbage long before generative ai existed.
seo farms were already churning out articles that barely contained information. companies were already buying photographs of suspiciously enthusiastic office workers gathered around laptops. marketing agencies were already manufacturing content with no purpose beyond occupying space.
now there's a tool that makes producing all of that considerably cheaper.
when you make garbage cheaper, you get more garbage.
that's not a particularly surprising economic outcome.
the same incentives that previously paid people to manufacture low-quality content now encourage using machines to manufacture it in much larger quantities.
the technology changed the cost of production.
it didn't invent the incentive.
that doesn't mean the resulting flood of content is harmless. it makes finding reliable information harder. it buries original work. it can damage the economics of publishing and creative work.
and artists have entirely legitimate complaints about training data, consent, attribution, compensation, and the use of their work without permission.
those deserve serious answers.
not somebody from a venture capital firm telling them to embrace disruption or fuck off.
i also think it's possible to distinguish between somebody generating a thousand interchangeable advertisements and somebody using an llm to help communicate an idea they couldn't otherwise get onto a page.
both may involve similar technology.
they aren't necessarily the same activity, and they don't necessarily have the same consequences.
i wish we could discuss that without turning every argument into a referendum on whether the people using the tool are morally defective.
0x05 llms can accelerate learning
there's another argument against llms that bothers me quite a bit.
the suggestion that making something easier somehow diminishes the people who learned to do it the hard way.
i spent years learning infrastructure engineering.
linux. networking. storage. virtualization. how machines fail. how distributed systems behave when something somewhere decides it would rather not cooperate.
i learned most of it by reading documentation written by other people, breaking things, finding out why they broke, and trying again.
i'm proud of that.
i also think it would be ridiculous to insist everybody else repeat the same ritual.
if a junior engineer can use an llm to help understand bgp in an afternoon instead of spending a week confused by an unnecessarily complicated manual, that's great.
amazing.
wonderful.
now they can spend some of that time doing something more interesting.
they still need to understand routing. they still need to know how to troubleshoot. they still need to recognize when the model has confidently invented a configuration flag that doesn't exist.
if they're putting something into production, they still need to understand what they're changing and what will happen when it goes wrong.
the model doesn't absolve them of responsibility.
but it might help them acquire the understanding they need to exercise that responsibility.
that's the distinction.
using an llm to learn isn't necessarily the same as using one to avoid learning.
sometimes the tool is helping somebody ask the question they didn't know how to formulate.
sometimes it's explaining the same concept six different ways until one finally makes sense.
sometimes it's just giving somebody a place to be confused without feeling stupid.
i think there's real value in that.
not every difficult thing is valuable because it's difficult.
some things are difficult because the tools are bad, the documentation is incomprehensible, or the people who understand them have forgotten what it's like not to.
we should be trying to fix those problems.
not defending them as a rite of passage.
0x06 for all
i work on infrastructure for a living.
i've spent a considerable part of my life around machines that cost more than houses, in buildings that require entire electrical and cooling systems just to keep them operational.
i love that work.
i love the engineering. i love understanding how the pieces fit together. i love the peculiar satisfaction of finding a failure that nobody else could explain and getting everything running again.
and i like the idea that more people should be able to use what we build.
i'm not going to pretend my current or previous employers are immune to criticism. they're businesses. businesses have incentives. some of those incentives encourage genuinely useful things, and others deserve to be questioned.
but i like the idea that somebody with a small budget and a weird idea can access computing resources that were once reserved for enormous companies with the money to sign multiyear contracts.
i like the idea of a researcher not having to abandon an experiment because the infrastructure is too expensive.
i like the idea of a developer being able to run a model without sending every piece of sensitive information to somebody else's service.
i like smaller, efficient, locally deployable models.
i like that people are finding uses for this technology that nobody involved in its development ever anticipated.
and i like the possibility that somebody who struggles with traditional ways of working might finally have a tool that fits the way their brain works.
the word accessibility means something.
it doesn't mean everything is free.
it doesn't mean everybody gets infinite compute.
it doesn't mean the output is always right.
it doesn't mean the environmental cost disappears.
and it certainly doesn't mean we can ignore the people whose labor, data, and creative work made these systems possible.
it means taking seriously the idea that a person's ability to participate shouldn't be determined entirely by the resources, education, money, or accommodations they happened to have access to.
we haven't achieved that.
not remotely.
the interfaces still fail people. the prices still exclude people. privacy remains a concern. accessibility features are uneven. and plenty of people have good reasons to distrust how this technology is being developed and deployed.
but the potential matters.
to me, anyway.
0x07 the empty file
i keep thinking about that empty file.
the editor open on my screen. the cursor blinking. the whole application sitting somewhere in my head, perfectly understandable to me and almost impossible to translate into something that worked.
i used to think that was just a personal failing.
that if i were smarter, more disciplined, more like the engineers who seemed to produce code as naturally as breathing, i wouldn't have that problem.
and maybe i could've practiced harder.
probably.
but that wasn't the entire problem.
there was a gap between what i understood and what i could express.
now i have a tool that helps me cross it.
not perfectly. sometimes it lies. sometimes it produces garbage. sometimes i spend longer fixing what it generated than i would've spent doing the work myself.
but sometimes it works.
sometimes i get to build the thing.
and i can't adequately explain how much that means to somebody who spent years staring at empty files.
there are people who think that makes the accomplishment less legitimate.
that if the machine helped, then the thing you built isn't really yours. that if you didn't suffer through every intermediate step, you haven't earned the right to enjoy the result.
i don't understand that worldview.
i don't want to live in a world where access to knowledge and creativity depends on how much unnecessary difficulty somebody can endure.
i want better tools.
i want better interfaces.
i want computing to be cheaper, more accessible, more private, and more useful to people who have historically been excluded from it.
and yes, i want the people building those systems to be held responsible for the harm they cause.
those beliefs are not mutually exclusive.
i can love artificial intelligence and still hate the way corporations use it to justify layoffs.
i can appreciate a model helping somebody communicate while being furious about the way artists have been treated.
i can build data center infrastructure and still think the industry needs to take its environmental responsibilities seriously.
i can enjoy a technology without believing it's going to save humanity.
that's not hypocrisy.
that's having more than one thought about something.
i don't think machines should replace the things that make us human.
i think they should help us do more of those things.
create. learn. communicate. understand. participate.
and maybe, occasionally, generate something profoundly stupid because it makes us laugh.
i've spent most of my life loving computers.
not because i think they're people.
but because, at their best, they make possibilities available to people who didn't have them before.
i think about that whenever somebody tells me that making technology easier to access somehow diminishes the people who use it.
a tool doesn't have to do everything for somebody to change everything for them.
sometimes it just has to give them a way in.
i like llms.
i like artificial intelligence.
i like the idea that somebody who couldn't build something yesterday might be able to build it tomorrow.
and after everything, i think making it a little easier for people to be people is a perfectly good thing for a machine to do.