Instead, what I've been doing for my personal projects has been about designing algorithms with TLA+. I write that by hand, model check and sometimes even prove the properties I think are interesting or useful, and then tell Claude to implement it based on the spec. Usually I still need to correct the generated code but it's a lot less.

Certainly in these cases, I take pride in what I did, because ultimately the "Invisible Hand" always means I'm outsourcing some of the work of a project no matter how involved I want to be. If I build a deck outside, I'm not forging the nails myself, I'm not growing and chopping down the trees for lumber, I'm not designing my own power tools, etc. There's always a level of outsourcing of the work that I don't find interesting or that I don't know how to do, or don't have the resources to do myself.

With software, the parts that I end up feeling proudest of, and this predates widespread use of LLMs, are the algorithms and tricks I come up with to build something. When I figure out a way to avoid a lock, or figure out a way to use less memory, or figure out a way to increase throughput, I'm less concerned with the actual "code", and much more concerned about the "how".

You can be proud of having commissioned a thing… be it landscaping or software… just not for making anything you didn’t.

And of course there are shades of grey in between, depending on how critical you were to the doing.

Sometimes I'm not interested in the inner workings of a technology, I just want a one-off tool. Ex: No interest in web tech. But, vibe-coded a tool to scrape and cross-reference a couple of online reports. Now I have the info I need for other projects.

Sometimes I know exactly what I want and it's quicker to hand-hold the AI through making it for me. Ex: Put together yet another SIMD math lib recently. SIMD intrisics are an obnoxious API. But, a non-SIMD implementation is easy to verify and many SIMD implementations are easy to validate vs. the non-SIMD reference. It's just a huge amount of obtuse code.

Sometimes I don't know exactly what I want and it's great to have a always-online partner to bounce questions off of, do rapid-prototyping for me, do code reviews for me. AIs aren't perfect. But, being instant, patient and often right makes them a huge improvement over online forums.

So like when I went to a restaurant last week and "made" a cheeseburger with no lettuce and extra onions, because it wouldn't exist if I hadn't created it.

To me, making something can mean you did the physical work of its creation even if there's no particular creative expression, like "I made a peanut butter and jelly sandwich."

It can also apply to tasks where you've offloaded a bunch of work that once would've been done by hand, like "I made this bench using a circular saw."

Where things get hairy even with physical objects is processes like CNC or 3D printers. If you download "Articulated dragon.stl" from a website and press the print button, did you make that? Personally, no. But if you designed your own flexy dragon and printed it, then yes. If you set up a relief carving on a CNC machine or do a perfectly fit inlay, are you making that? Operating these still takes quite a bit of knowledge and skill, but different knowledge and skill from hand woodworking. What if the tools improve and become very easy to operate?

I don't think there's a hard line you can draw. But if there is, it's somewhere between "did everything by hand with no assistance from recently invented tools because those are cheating" and "ordered a sandwich and waited for the finished result to show up."

This became more relevant to me when I began meeting producers as a software engineer in game development.

You didn't make it

If anything, I would say your conclusion should be flipped: you didn't really do the "making," but you can still be proud of what you _did_ do: provide direction and a vision for a solution to a problem.

I still believe it's vital to learn how to code, but now we have to front-load more of the SE skills that have historically been obtained on the job.

"What a time to be alive!" I admit that I wouldn't have minded if all this LLM stuff has waited another 15 years to happen. :)

In the code scenario, I most likely won't go over the final result: for various reasons, I often lack the interest and drive to understand the nuts and bolts of it. So I'll just try to assess whether it's sufficiently good for my goals, and then move on.

But in the circuit design scenario, I only use it as a "technical expert" to bounce ideas back and forth with, and to explore concepts I don't fully grasp. I use it as a sort of ELI5 machine, and never ever ask it to do something for me - just to explain.

The fundamental difference in outcome is that I remember next to nothing about the code, and have no feeling of "ownership" over it, nor do I feel proud or care about it at all.

Conversely, I am incredibly proud of the PCBs and circuits I design, even if technically some parts of it came straight out of AI recommendations. I really feel like I did the job, and AI was just a helper tool to get the tough concepts untangled quicker in my head.

Ultimately I think if you delegate 100% of the work to AI, and don't care enough to use its output to at least learn something, it's only natural that you'll never feel any connection to it, let alone any sense of ownership over its output. Being the prompt author is simply not enough. That's why I think you hit the nail on the head there: deeply understanding the output is the key, I just couldn't put my finger on it before, and now I can. Thanks for that.

In making something, you develop a tangible understanding of and opinions for how and why your creation is how it is.

I have a lot of vibecoded projects. I feel "ownership" in varying degrees proportional to the amount of care and depth I put into them.

It was never about writing every line by hand, it was about solving a problem and making my life better. LLMs let me do that at a speed never seen before.

I was and am and avid Home Assistant fan but the number of automations/scripts I now have is higher by a magnitude than ever before. Few things would reach the level where it made sense to spend my time writing and debugging a workflow but now I can speak my automations into existence.

Yesterday I got a candle warmer, I wanted to set it up so when I turned it on it started a timer to turn off and I wanted a push notification to see if I wanted to extend the timer when it was close to finishing. All things I knew were possible but HA YAML is miserable (all YAML is) and the UI is not much better (it's a hard problem, I get it). I prompted Claude and had a working implementation in <10min. Additionally it asked if I wanted to auto-turn off if I left the house (yes!) and if it should auto-start the timer if ever detected the candle warmer was on without a timer set (power loss, HA restart, manual interaction with warmer).

My HA automations have never been more bulletproof and if they do something wrong it's as simple as telling Claude Code and having it fix the problem. For example, my motion activated lights (motion sensor -> HA -> smart switch) were working great but I realize that while in the shower my main bathroom lights would sometimes turn off. 1 prompt later and Claude grouped up my motion sensor in the shower to the one in the bathroom so if motion on either was detected the lights would stay on but the shower light would only turn on if its motion detector detected motion.

HA is only one small part of my LLM usage but it's representative of my overall point.

There's a line there somewhere that gets crossed.

Part of it has to do with direct one-to-one determinism. I pull that lever, the excavator bucket pulls up a ton of earth. There was no extra decision made in the middle. And it was completely reproducible.

We might even go so far as to say, "I moved a ton of earth." Even though it was the backhoe that did it.

Back to CNC, consider the manager who orders the machinist to use the CNC to make a widget. Does the manager qualify as a machinist?

I think most people would say, "No. Not unless he coincidentally was one."

There's another person in the middle in that case. That level of indirection is somehow making the manager not a machinist.

I speculate it has something to do with the fact that the machinist in the middle is an independent being that holds knowledge to make things go. The manager gives a very high-level instruction and doesn't instruct the machinist on how to do it.

And I think LLMs are starting to feel (to folks like me) a bit like that machinist. They're not deterministic like the backhoe and the compiler. And they seem human, convoluting the matter further still.

Instead of saying "create instructions for CNC to make me a _thing_, then instruct the CNC machine to actually create the _thing_" they went and created the instructions themselves.

To some degree, some folks esp. for "lights out" machining are just pushing buttons and loading/unloading pallets of stock/parts, but usually the job description for that is "machine operator".

A machinist is expected to understand feeds and speeds and how they interact with toolpaths and materials and tooling.

The big question of course will be how things fall out with the on-coming generation of 5-axis machines which are intended to be AI-driven:

https://toolguyd.com/new-ai-powered-desktop-cnc-machines-202...

Alongside understanding, you also loose taste and perception. You can still say whether some things is good or bad, especially if you interact with it, but you can no longer articulate why. And with perception loss, you can no longer preempt a bad situation as you lose your awareness of things.

So for you a car may be a vehicle that you drive around. But for a mechanics it's a much complex system. When something doesn't work, you will lack the vocabulary to express what's wrong. And even when everything works, you may not be able to explain how.

The thing is, LLM are very buggy translators. Agents in software development help because they provide a way for LLMs to verify their output. But correctness is very hard to express. OK for a simpler task, but not really for a complex systems. Just like fluency in a foreign language is not about knowing the rules of grammar and a lots of words, but also knowing the idioms and other figure of speech.

I think there's a lot of parallel with experienced programmers using LLMs and that idea about steroids that it just gets you there faster. I'm pretty sure I could eventually program most of the things I've used AI to write, and what it produces at the top end I/you can't really prove that I wouldn't have thought of it. But is that really true?

Of course, you may not care, and that's fine. It's important that people not be delusional about this though. A person with no medical knowledge and an AI will never be as good as someone with prodigious medical knowledge and an AI. The same is true for any profession, whether programming, music, writing, etc.

I think for tasks that are about decisions, having the LLM make decisions is what makes it feel like the LLM did something for me.

Consider mowing the lawn. Imagine I had a lawnmower robot that does the mowing all on its own. Despite perfect accuracy, I didn't mow the lawn; it did. If I sit on that lawnmower the whole time and start driving it instead of letting it go on its own, then I mowed the lawn. Even if I stand there and control it with a joystick, I still mowed the lawn. Ownership comes from the decisions about where to mow.

Some are ok with deciding if and when the process should be started and whether the outcome is good and bad. But often they lack the judgment to truly decide what is good and what is bad. They trust the makers to provide equally good options in terms of lower concerns (that matters to the makers) and only make the judgment according to some other concerns.

The main difference when using a compiler is that the compiler will (99.999...%) produce correct binary code for my programs and will fail loudly if it cannot. But the compiler is not concerned about correctness of the algorithm written in the programming language. Responsibilities are clearly outlined and the compiler does the job perfectly, while I try to do mine as well. So when I do a good job, I don't have to worry about the compiler part and I can assume the result will be good as well.

LLM usage has no such clear line. Assuming the user can judge whether the outcome is good or bad according to the prompt, the LLM tool is not reliable enough to produce a good result for everything that was not specified in the prompt. Which is different when collaborating with humans. If you say to a professional artist "I want a batman drawing", you needn't check for missing toes and the numbers of arms.

Yes. The CNC doesn't act on its own.

If I ask the CNC "make me a widget", am I a machinist?

Tools are tools. Same as Photoshop, Ableton, they get you to where you want to go.

Where is the line drawn?

"Make" gets confusing for AI-assisted projects because it carries a connotation of craft. It seems fully justified for a code project that is 90% you, 10% AI (you'd also say you made something that was 90% you, 10% human help) and misleading for a project that is 90% AI, 10% you (unless the meaning of the word shifts).

Here is a relevant comment I wrote in a thread about AI art (https://news.ycombinator.com/item?id=46705952#46708186):

>> If you do it daily, in-house, for your own products... you might just have the title "Art Director."

> "Art director" seems accurate for what a skillful user of art generators with a specific vision does.

> I have also thought that since people find "director" lofty (thanks to auteur theory?) and therefore pretentious to assume, one could borrow "producer" from Vocaloid: https://vocaloid.fandom.com/wiki/Producer (alternative front end: https://antifandom.com/vocaloid/wiki/Producer).

I ask 10 actors for 10 scenes and I combine them into a movie I wanted.

I ask 10 agents for 10 functions and I combine them into a program I wanted.

What's the difference and why is the guy who used agents not credited the same way as a director? Maybe all that's missing is a long list of credits for everyone involved -- so transparency in who did what?

In your analogy, what I’m talking about is more like talking to a chef with the idea of a novel recipe and describing the ingredients and how you would like it to taste (of course also accepting suggestions).

I would take more pride in that meal than in a standard burger I completely cooked myself - maybe because I know I am able to prepare a great tasting burger and already did so many times. OTOH if this would be the first burger I made in my life I would be immensely proud of it.

To be fair he didn't place a single stone, and la Sagrada Familia was finished long after his death.

So yes, I believe you can make products using LLMs.

If there is a line, it's definitely between "I designed the architecture, and decided on what each class should do, before letting AI do a single line" and "I told AI to clone twitter as is".

Although the latter could be akin to kids colouring books. "I made AI clone twitter, but painted it red and called it rettiw." As long as there was some decision or a modicum of effort it's "ok". We can let them have their victories and try again.

I feel the same way if the agent moves too quickly. If I ask for a feature in high-level language the LLM will make many guesses and assumptions to carry that out. I don't feel like I'm involved in the creation of that software at all. It feels more like running a find and replace on someone else's code to make it to what I want.

That's why I instruct the LLM at a low level, that is at the level of the code itself. I'll ask for refactors, I'll ask for a new class etc. This cuts down on time spent typing while I still remain on top of the code.

At least that's the idea. But there's still a problem. LLMs are trained to want to write working code. If what you asked for won't work, it'll just fill in the blanks and guess to make it work. And then you don't feel like you made it again.

Basically, we kid ourselves into thinking it's just a supercharged text editor that we control, but it's actually an LLM that just doesn't work that way.

This is analogous to the compiler example: It generally doesn’t matter which exact variation of machine instructions the compiler chooses to generate from the source code. It doesn’t significantly impact the ability to predict the resulting program’s behavior in detail based on the source code.

This predictability relation is what allows the creator to reason about the creation, to reason about which exact change to apply to the input to effect a precise modification in the output.

In that sense, the robot lawnmower is like a compiler, if it yields the state of lawn that I specified. I largely don’t care if it first moved left or right, the same as I largely don’t care if a compiler translates a multiplication by a power of two with a shift instruction or a multiplication instruction.

However, we use LLMs precisely for their heuristics, for their property to actually make material choices. Therefore LLMs inherently conflict with the ability to reason about which exact input will yield which desired output.

Currently trying to figure out the right kind of caffeine intake, seems hot chocolate might be my jam vs. coffee. I'm lucky we have a gym at work so I work out first 30 minutes of the day helps and do walks.

I have to grind on the side though to get out as I have $100K in debt. That six fig day job still means I'm check to check.

Both of the examples the parent gave include some kind of planning and ideating to get to a final product, regardless of the fact that they employed tools to do a bulk of the coding work.

The only thing being "redefined" is corporate agendas to push the bar of "good enough" lower to the point where it lies below cheap, and equals wasteful.

The difference with Movie director at least is in details. IT director most likely does not care about Python, bash, js, mermaid, performance characteristics and so many other things that AI agent might have used to solve the problem. The user of these AI innovation are not having much say. They are not going to be voting with their wallet or feet.

For movie director the details of process matters a lot. Unless of course I am shooting 2 min video on my Motorola phone, while my kid is cutting her birthday cake, and call myself a movie director while doing it.

You prompted, the LLM made. The output of the LLM is not eligible for copyright.

You can perhaps copyright the prompt. That you did make.

Unironically, this is some of the software being shared. Maybe a couple more back and fourths, but not much more. I know I "made" a simple discord bot that periodically checked a webpage for updates and sends me a message when there are changes. It was basically just "I want a discord bot that checks this webpage and messages me when X, Y, or Z". I only looked at the code slightly to know that the entire thing passes around loosely structured strings everywhere and checks for substrings over and over again for logic flow.

This is what communication is. Words are representations for whatever it is that we're talking about. There is no other way.

Agree with you on what we're talking about, disagree on the stance. It's a discussion of value yes. vibe-coded software has less "I made this" value but still holds "does this have utility" value. They're different. Making things with AI is its own kind of making, but it is not the same kind of making as baking a cake from scratch, building a doghouse, or planting your own garden.

In more complex cases you might specify every nuance of the way the coffee should be prepared and a skilled barista carries that out.

I read a book about David Tudor a few years ago and something one of his teachers said to him has stuck with me since: there is a difference between thinking about things and doing things.

I like that because it doesn't (in my view) put those two things at odds, it just differentiates them. Thinking and doing are different.

They are both valuable domains but interrogate different experiences and outcomes. They both seem valuable.

Yet,

Christopher Nolan made "The Odyssey"

Shigeru Miyamoto and Eiji Aonuma made "Ocarina of Time"

Mark Zuckerberg made "Meta"

I feel just as proud of everything I'm making with LLM, image, and video models as I do the code I wrote by hand for years and the photons I shot on film to make movies.

This is just a new higher level thing. I'm still the "director".

It's my design and my taste.

Author is right about this though,

> "Didn't take long. Only about 50x longer than it would have taken Claude to do it."

I'm never writing code by hand again. I stopped in February.

So I think it's playing both sides. Nolan didn't make The Odyssey. He directed and produced it. Large projects require many many people; with this in mind I realize this whole discussion is just one's ability to make something of their own accord: a cake, a pottery bowl, a wood shed. The whole LLM thing blurs the lines but really it turns the conversation into AI rallying.

It's as simple as recognizing the difference between "making" The Odyssey and "making" breakfast from scratch for your family. (and no you didn't have to have laid the eggs to claim "making" so it's really not so rigid)

Hitchcock made "Psycho".

Miyazaki made "Princess Mononoke".

While you can say Roger Deakins shot "Blade Runner 2049" or Hans Zimmer scored "Interstellar", even these masters of their craft paired with directors to make their films. The film is still typically the work of the auteur director, and the director (usually) exercises control over all of these pieces.

You can tell when that doesn't happen. You can tell when the producers exercise outsized control. Or when the studio controls the editing table. And it's not like that's ever kept a secret - people talk about this openly and loudly when it happens.

Musk built his empire, and even if he bought his way in, he still used his unique methodology to turn it into what it is today. "Founder Mode", if you will. Same thing with Steve Jobs.

All of this hand-wringing about who painted the shed in "The Evil Dead" doesn't matter. It's Sam Raimi's. Everyone else on set dances to the beat of his drum.

Software can (and will) be like that too.

Game design, company formation, product development, you name it.

The fairy path of human language is going to adopt "creation" as assuming someone could have prompted it. It's the path of least resistance, and it's something that will only increase in normality in our daily lives and experience.

Nobody in the future is going to have these paragraphs-long diatribes about what does and does not constitute creation. People are just going to say, "I made this."

And that's that. We'll judge based on the significance and novelty of the output or work.

    Canon mountaintop photography : smartphone selfies :: Quality AI art and software : ChatGPT Miyazaki cartoons

Nobody will bicker about "created", either.

If someone in the future wants to spend a year making a stop-motion, claymation-animated feature, people will still think that's cool. It'll be something they "created". But it won't stop the world when most people are using new tools, nor will it stop us from calling their outputs "creations" and something they "made".

I'm making something with Claude Code right now. And to be very deliberate about it - I am making it.