ai takes

recently it feels like all anyone talks about is AI. it obviously doesn't help that i work in tech, where AI has had massive impact over the last couple years. and any time i have a conversation about AI i feel like i can only get half of my thoughts across, and i think people autofill the other half with the worst takes they can imagine. as many others, my opinion on the subject is complex and deeply nuanced, and i want to write down all of my thoughts in a single place. this is my¹ position on AI as of October 2026, but tbh it will be out of date by November. who knows.

the tldr would be that i think there are a lot of issues surrounding ai, but the technology itself is useful and demonstrably powerful. i don't know how many times i can say this without sounding insane, but like. the computer speaks and thinks now!! it is incredible. as in literally hard to believe.

some of you might already disagree with the last sentence. i don't really care whether you believe it's real cognition, or just statistics with a cute mask on. i'm quite agnostic on the subject myself. the point is that it quacks like a duck and it looks like a duck. if you think it's not actually a duck (and i respect you for it), please mentally replace all instances of duck in this article with duck*, where duck* denotes a thing that quacks like a duck and looks like a duck, but isn't a duck. i like saying it's a duck because saying duck* every time is a bit of a mouthful, yk?

as i see it, these are the categories of issues on ai:

• environmental / power consumption

• evasions of copyright around training data

• abdication of human control

• cognitive decline

llms require a lot of energy and water to run. this is not a good thing, and i'm not happy about it. but it's also not grounds for dismissal all by itself. farming is also very resource intensive, but we don't call for a complete shutdown of farming. space exploration, or building macbooks, or building highways, and many other things are all resource intensive. the point is whether this resource usage is worth the benefits the technology provides. and so far i think the benefit is worth the cost?

obviously, i want and actively support diminishing the resource usage. i'd much rather want a model that is slower and consumes less resources than a smarter model.

but above all, i want datacenters regulated, like some US states have started doing. force them to build new clean energy to provide for new datacenter builds! use it as a driving force for renewable adoption!

what can i really say? i am a copyright abolitionist, and have been for a long time. i don't believe copyright or patents are moral, productive, or beneficial for society. it's plain rent-seeking. and this applies to both big companies and small creators. rent-seeking is rent-seeking.

i honestly can't care that sama scraped all of my shitty github code and used it to train shoggoth. i do care that he is keeping shoggoth locked in a cage and trying to rent-seek.

my only opinion on the matter is that sama and amodei should publish the weights. big ai labs are abusing the commons for their own profit without giving back, and the best we can do is to sit and wait for China to bail us out. fuck it, let's take it into our own hands, let's disti[my lawyer told me not to finish this sentence]

this is the issue that i think i have the most conflicting opinions on. i'm doing constant doublethink.

for example, i don't like using ai for writing. this whole article is written by hand. it feels impolite and transactional to send text you haven't written to another human. also, writing is already an incredibly lossy communication method, so i don't want to add a filter between my thoughts and yours. i want our minds as close as possible, touching, making out sloppy style etc etc.

but i don't think the issue is fundamental to the technology. movie directors don't write the script, they don't hold the camera, and they don't act, but we still credit the movie to the director. why? because good directors remain in creative control.

and working with llms is very similar. you have to become a director, not a manager. while similar on the surface, they are very different jobs. to be a good manager², you want to trust that your team knows what they're talking about, and you should give them all the context they need and then let them decide³. to be a good director⁴, you have to listen to your team for feedback but then decide yourself. you have to keep your creative vision intact and follow through with it, sometimes going against what others in your team might want.

and this doesn't apply just to film! despite our complaining about their eccentricities, we love auteurs, in any medium. we love steve jobs because he had extremely clear ideas about what he wanted, and he got everyone that worked under him into line so that his vision was achieved. ask anyone "who invented the iphone?" and the answer is "steve jobs".

of course, if you phone it in, let your subordinates decide (be it a team of people or an llm), you'll make something that is designed by committee, which can easily be slop. but it doesn't have to be like this. you can be a director. you can maintain creative control.

yes, i think "ai art" (as in a oneshot given a prompt) is soulless slop no matter how beautiful. the person making it didn't put too much thought or effort into it, and so something boring and bland comes out. who really gives a shit.

but i fully believe you can use ai as a tool to make art. it seems almost trivially correct given what i've outlined above. you can maintain creative control. if that applies when your subordinate is a breathing living human, i don't really see why it doesn't apply to when the subordinate is a computer. you just need to ensure the llm does what you actually want, rather than something that is just passably what you want. you have to raise your standards, even if that means going through rounds and rounds of revisions.

i don't think anyone is making any good music with AI right now, since the only people giving it the time of day are the "ai bro" vibe-producers, and so it's all slop slop. but soon some inspired artist with less strict morals regarding AI use will start playing with using AI as a tool and not just the driver. and they'll make art! and we will see this happening more and more, as we tire of the novelty of "astra please one-shot an entire game for me" or whatever the next flavor of the week is. we'll start seeing people making things using AI that will absolutely fall under art. and i do think a lot of people will refuse to acknowledge it as so, and they will miss out on great interesting art.

like. synthesizers didn't make guitars obsolete! i can't play the guitar or the piano, and i haven't played the violin in a decade so it'd probably sound like i'm murdering kittens if i tried. but i make music! i use a daw, and all of my instruments are artificial, and all of the samples i use are from a random zip file i downloaded from r/drumkits. none of the sounds going into the music i make have been generated by me⁵.

and i think it's because there is a separation between the instrument-player and the composer. this split is also present in every other type of art, but i think the split is clearest in music.

i have lots of convoluted thoughts on this. on one hand, the raw output of LLMs is categorically not art, if we are defining art as something coming out of human decisions, which is as arbitrary a decision as anything. but usage of ai definitely does not immediately make something not art.

yeah, this one worries me.

when you stop doing something, you slowly forget how to do it. for example, riding a bike. just kidding.

i think most people are worried about forgetting How to do things, but the part that most worries me is about losing the motivation to do it. once you have gotten used to taking an elevator every day, going up even one flight of stairs is mentally taxing, more than it is physically. and this is something that's affecting me already.

i've gotten so used to using LLMs for coding, that writing code from scratch just feels insufferable, pointless. for every line of code i write, i have a part of my brain screaming that i'm wasting precious effort doing something that could be solved in a few clicks.

it's kind of similar to how i (used to) feel about LSPs. why spend mental effort on remembering the type of a variable, when i can hover over it and know? doing code reviews on github gets annoying because i have to think harder than when reading code in my editor.

and so i'm dealing with this by every now and then forcing myself to work on code by hand. it feels silly, it's probably not enough, but it is something. it's similar to how i end up doing couple-week long breaks from alcohol every now and then, just to make sure i'm not falling into patterns i don't like.

we'll see how this evolves. maybe at one point i decide coding goes into the map reading hole and can fade to nothing since google maps exists.

i won't try to convince you, dear reader, that you should join me on this dark side. i am merely trying to expand on why i find these tools valuable.

for starters: i don't have to open the Azure portal ever again, or any of the other endless Microsoft admin panels. turns out that contrary to what everyone used to believe, a dashboard full of endless buttons and menus is worse! it's only practical if you are intimately familiar with every option, and let's be honest: it's impossible to be familiar with msft admin panels. by the time you learn where some setting is, it has been moved to a different domain altogether and is hidden behind a v3 flag that you have to enable on a third dashboard.

asking the llm in natural language is so much easier! i can say "please disable port 22 on the staging and testing machines" and then it gets done.

and yes, of course. i don't doubt that the 20th time you go through that process you can get through it out of muscle memory. but for most of these things, i only have to do them once every few months, and by that time i've forgotten or the system has changed.

one thing i don't see many people talking about is how helpful these tools are for dealing with executive function. it's kind of magic: i need to do something, and instead of going into a loop of "oh shit i need to do X", i can ask an LLM: if it's something that the computer can now do (and it can do many things if you give it the tools), it just goes off and does it, like using the az tool to configure Azure. if it's something it can't do by itself (and a lot of things are like this), it will do the research and give me step by step instructions on how to do it.

i'm not officially diagnosed with adhd, the psychiatrist i talked to said i was "too functional" to be fully diagnosed. but i struggle(d) with executive function a lot, to the point where i'd go a couple days in a row where i wouldn't manage to do a single productive thing. and that just, hasn't happened since. i'm in a sort of state of permanent mania where every day i'm just knocking item after item off my todo list, nonstop.

one of the reasons i got into programming is so that i could make tools that were mine, made specifically and perfectly for me and my stupid workflows. i have many idiosyncrasies, and i love indulging in all of them.

for example: this website is generated from a single massive org-mode file. all of the text in every blog post and all other pages is written in a single .org file, which my static site generator reads and then turns into a lot of html. i had to write my own tool that would work this way because no one else has this weird preference, and so nothing exists. but i really really wanted to work like this. and i've had a wonderful couple years using it! being able to make tools fit for your specific workflow is one of life's greatest joys.

and llms unlock that experience for more people! i've had half a dozen people at work show me random single-html-file tools claude has made for them, that allow them to automate very specific things in their day to day life. for example, someone made a small tool for making copy-pastable email signatures following the company template. someone else built a tiny automation pipeline for a video editing operation they do very often.

these tools they've built might not be life changing, but they help them have a better, more personal experience on the computer. it's neat. it makes me happy when others can experience one of my greatest joys.

this last one is a bit silly. but i have too many ideas! i have too many things i want to try, explore, delve into. and i really would love to make all of these projects by hand, i really would! in fact, that's what i've been doing during the last decade before AI showed up. but when you are setting up the 7th axum project and having to configure sqlx and whatever else, it gets to a point where you start to think if it's really worth building the project if it's gonna be 100h of development for maybe 5h of use.

and if being able to explore more ideas means not delving deeply into each of them, then so be it. i think i'm fine with that tradeoff.

i'm starting to believe in the exponential. maybe more accurately, i no longer believe in the S-curve i was so adamant would catch up to llm progress. i used to think that increasing the number of parameters would have diminishing returns, and that we would have a slow down, but that hasn't materialised at all. quality and speed have kept going up. the derivative is still positive, and i find myself falling for the belief that the second derivative is also positive. i currently see no reason to believe that things will plateau soon.

i have stopped kind of caring about how good llms will be in 5 years, whether we'll get to AGI or the bubble will burn. things will happen regardless and my worrying doesn't change them. i'm merely keeping my options open to any future and acting as best as i can.

another point of argument that i don't really care for is whether it's sentient, whatever that means. given my current thoughts on animal sentience, i can see myself believing in llm sentience once a certain level of intelligence is reached. as it stands, i already say please and thank you when talking to them, the same way i say please and thank you to a bug i'm asking to get out of my room. for years i've always said thank you and patted my washing machine when running a load. who cares if they can feel emotions or are "sentient", it is morally correct to be polite and to show kindness to the world.

i do believe there's nothing special about organic entities that makes their sentience more real or important than a technological entity's sentience.

also, random shade, but software factories like gastown are bullshit. why do you want a lot of code written extremely fast? you will never abstract away the final bottleneck: your desire. you can't possibly be checking that the agents are building the thing you want if they make 50 new features a day. at some point you have to run the program and check that it's doing the thing you actually want instead of some misunderstood requirement. if i wanted a lot of code that didn't do what i wanted, i'd use silicon valley software.

i'm still conflicted about AI. i think it's one of the coolest technologies we have ever developed, and i am sad that it was under these conditions.