WCSTombs

> I love building startups, and they love writing code. For me, building projects is a way to make money. For them, programming is art.

At its very high points programming approaches art, but programming is primarily problem solving, which isn't fundamentally artistic because its main goal is correctness and fitness for purpose, not aesthetic value. I'm pretty sure for most of us the choice to use LLMs or not has nothing to do with it being art or not art. If I estimate a cost-benefit analysis, the true cost of using the LLM is much more than it would likely ever provide, so it makes little logical sense to use them except in very limited situations.

By the way, I have to say this is a really strange type of article, an explanation of the supposed opinions and attitudes of non-AI-users written by an AI user, seemingly without any input from the former group. It reads a bit as if it's building a strawman to justify the author's own views.

I think you're describing the the Diátaxis framework [1], which would further split your (1) into fully guided tutorials and discursive explanations.

[1]: https://diataxis.fr/

This came up before (same topic, different article): https://mostlyobvio.us/2026/02/responsive-code-formatting/

I'll also paraphrase what I said then: I like the idea in principle, but there's a fundamental problem that makes me think it would be really difficult to use on my website. The problem is that code autoformatters typically work great in like 90% of cases, but the last 10%, they make worse. In "real" code, the programmer can often opt out of autoformatting in problematic cases by adding special comments in the code, but that's a non-starter for code samples on my site because it would look tacky.

So what I actually do is manually format my code and break lines at about 64 characters. This doesn't completely solve horizontal scrolling, but hopefully it's good enough.

That's not to say I would never use this idea. There's a threshold of "good enough" where it could become worth it even if occasional workarounds are needed.

I agree that there can be a tendency to conflate the two things: how advanced the technology is, and how aesthetically pleasing what it produces is. It's also important to remember that we're not obligated to like something just because it's "pretty" or has high aesthetic value in some abstract sense.

In other words, the three questions of:

- is the technology good?

- does it produce aesthetically good outputs?

- do I like it and want to engage with it?

are all pretty independent.

A lot of the stuff you wrote about is not what most of us are talking about when we call it intellectual catfishing or say your intellectual fly is open. For example, I have few problems with people using LLMs in private for purely personal things. I still won't go ahead and recommend it, but it's a decision we can all make for ourselves. Also, writing an email to your city is something you could ask your assistant to do if you were super rich, and using the LLM to do it isn't that different, so I'm not going to judge you for that.

As a more nuanced case, someone could dictate a long rambling stream of thoughts full of contradictions to an LLM to transcribe and summarize, and then they could reflect on the summary and write the real thing themselves. That's like talking to someone about a subject and writing about it afterwards, and also not something I would really take issue with.

That's my partial answer to the cognitive prosthesis question: keep it off to the side as a tool to help you do the work. There are still many caveats here, and the more complex the demands you make of the LLM, the greater the risk of some kind of break down. For example, continuing the example in the previous paragraph, if I did have the rambling conversation with an actual person who helped me refine and clarify my thoughts, I have some kind of mental model for the dialogue partner that can help me correct for some biases. With an LLM, anything resembling a mental model I could have would be very far off the mark, so I would need to be aware that I can't treat its output the same way I would treat something written by a person.

(By the way, for myself I have a stricter rule: I would only use an LLM in a way that made me a better person independently of the LLM and not dependent on the LLM, and thus I don't use them at all. That's what I actually recommend, but I don't expect everyone to adopt that rule.)

My main problem with having the LLM just do your writing is that it's simply a misrepresentation. It's wrong to claim you wrote something unless you chose the words. I really empathize with the struggle of expressing oneself clearly, but there's just no getting around this. And to be clear this isn't something pedantic or just a technicality, since the act of formulating a thought in an actual human language with valid syntax does require a level of care and attention that simply is not there unless you do it yourself.

I agree with the sentiment with a small caveat, since IMO good writers don't write the way that they talk. The very simple reason is that you can revise your writing but not your talking. But yeah, I completely agree that using an LLM to write your words is intellectual catfishing, and that's a great way to phrase it.

I'm in the process of writing a pretty long essay that has taken a few months. I've written the first draft, and I'm almost done with the second draft, which incurred substantial revisions. At the end of the process, maybe no sentence will be something I would actually utter in person, but the writing is still something I created, and as such, it is a representation of who I am as a person. Had I used an LLM to do any of the writing, that would no longer be the case, and the writing would at best represent me as a person when the LLM is at my side, and at worst (and most likely) not really represent me at all.

This is meant as a criticism of the interview rather than a defense of the AI company, but you have to weigh the "brutal candidness" of the interview against the tendency of the LLM to agree with the prompter. A lot of the questions were asked in a pretty leading way, and reading the responses, it's apparent that the LLM caught onto that and gave the user what they wanted, which was a scathing self-critique. For example, one of the questions simply called the LLM the "Temu of AI" (not really a question I guess) and the LLM enthusiastically agreed with a detailed reply, which is a foreseeable outcome.

I don't personally use LLMs (and won't get into the reasons why), but for those who do, I strongly recommend not trusting any output from them that can't be independently verified. This LLM made a number of very specific claims about itself and Google that to my knowledge are not independently verifiable. Thus, while parts of this interview are somewhat entertaining (amid other parts that are extremely dull and repetitive), I don't know how informative it really is, and moreover, since I'm pretty sure the LLM is designed on some fundamental level to be entertaining, the whole exercise just doesn't seem that interesting at the end of the day.

Knuth didn't write SICP, it's by Abelson, Sussman, and Sussman.

The creator of Omarchy is a pretty controversial figure because of culture war stuff and also a prominent software industry entrepreneur, so it intersects those two other spheres to give it more prominence than it would have otherwise.

Personally I think it's a bit pointless, but you seem to be taking it personally somehow. Can you simply choose to ignore it and not get annoyed? It's just water under the bridge and won't matter at all in the big scheme of things.

Yeah, it's the speed-running homework and unchecked potential for misinformation that are really problematic. I agree that if it was only a more efficient and accessible repository of information, it wouldn't be nearly as bad, and that's partly what I was getting at in my top-comment (the "may be a version of AI that helps people learn" part).

If you overly rely on the computer assistance during the learning stage, there are fundamental principles you won't get a firm grasp on, and then you'll be at a disadvantage if you need it later. As someone who did use calculus after school, I'm glad I actually learned the subject.

You're right that most people don't need to know calculus in their lives, but to my knowledge calculus also isn't generally required to get through college/university for most people, so it's fair to assume that the people in the course are there because they need or want to learn it. When I was a TA for calculus, most of my students were pre-med, engineers, or in other sciences.

Here's another (non-paywalled) source for this news: https://abcnews.com/Technology/new-york-city-public-schools-...

The ban goes through eighth grade, and in high school apparently there's a specific AI literacy class. It's unclear from the article if that's the only AI in high school or it's just unregulated.

My opinion: it's a good policy and frankly just common sense.

- Students using AI risk offloading their cognition onto the AI and thereby skipping the actual learning process.

- Teachers using AI likewise do a disservice to the students.

In the middle of all that, there may actually be a version of AI that can help people learn without incurring the major downsides, but I've not seen any evidence of that in the current AI systems, to the point that an outright ban in the school setting makes sense. I'm heartened to see more education organizations think similarly [1].

[1]: https://www.law.berkeley.edu/academics/registrar/academic-ru...

I recall the Linux kernel uses "Assisted-by" lines or something like that, which sounds about right.

> In my view, adding an LLM as co-author gives you a way for you to shirk responsibility for mistakes – even if a little bit. You can say, oops sorry this happened because it did its thing and it was not careful. And that is exactly the behaviour I want to guard against. If a piece of work is under my name, I own it. Fully.

One of the few aspects of generative AI in coding that to my knowledge has never been controversial is that no matter what, the code's human author is fully responsible for it (even if an LLM actually wrote it). So adding anything about LLMs in a commit message has never been a real way to escape responsibility, even a little bit. Maybe this author's workplace is different, but I kind of doubt it, and even then they're not only talking about their coding at work.

The author goes on to say that they also do want to be transparent about their LLM use, and they suggest some ways to do so...great! On the receiving end of the code, that's all I need. There's nothing sacred about the commit message in my opinion, to make it the one true place to disclose LLM assistance. So ultimately I don't think I have a huge disagreement with the author, just that their justification makes no sense to me. The only time I hear that LLMs supposedly help you evade responsibility is from people trying to avoid disclosing LLM use.

> Its final speed didn’t just break the sound barrier. It soared well over the 767 mph threshold at a staggering 776 mph.

Was Popular Science always this bad?

Anyway, this looks like quite a feat of engineering. Hats off!

I'm a little confused. Is this independent research, an expository piece on the state of the art, or something else entirely? The title sounds like an expository piece, but the document itself looks more like "something else"...but what exactly, I'm not sure.

While I have a distaste for most things AI-generated, the proposition that "AI-assisted [writing, art, ...] is its own medium" is probably something I can accept, in that I think it's just factually correct. However, that doesn't make it more palatable, or something that I think should be embraced in day-to-day life.

I don't think this author and I will ever agree on the extent to which it's acceptable to use AI to generate parts of a creative work, but let me propose a middle ground that hopefully a lot of us can accept: direct use of AI in a creative capacity must be disclosed. To be clear I'm talking about using the AI to generate literal text for an essay or artistic assets, and not simply to give feedback or ideas (which I don't fully condone either, but that's not the middle ground I'm proposing). Since this author detailed exactly how he used the AI for this article, let's acknowledge that had he given that same work to a human underling, I think almost everyone would conclude that the other person should be credited as a coauthor. I think that's a reasonable standard: if the AI is doing creative work that a human would be expected to be credited for, then its use should similarly be disclosed.

The author said he's proud to have used the AI to assist his writing. I appreciate that candor, so I'm saying let's just make it official and make disclosure the rule going forward. If using the AI is fine, then what's wrong with acknowledging that you used it?

Let's cover a few potential objections.

1) "It's unenforceable." But I'm primarily talking about cultural norms, not enforcement. There are also other forms of "academic dishonesty" (for lack of a better term) that are difficult to detect, but that doesn't mean society just embraces them.

2) "I'm not ashamed of using AI, but I have a right to privacy." I value your right to privacy, but we're talking about something you attached your name to and put out into the public. Your published work does not have its own right to privacy.

> they are one of the many genAI companies who basically view people with a kind of contempt.

It's absolutely this, and they're not even hiding it. It's everywhere in their public presence.

- Their now-retired marketing slogan is, of course, openly contemptuous towards people.

- Although I don't know all the marketing jargon, if I got it right, their pitch is to have LLMs churn out marketing emails - again, contemptuous toward whomever's on the other end.

- This very release is itself AI slop - total contempt toward us, the readers.

They're saying the opposite, no?

Good writing is rewriting. That said, in my experience the hardest part is writing the first draft, i.e., just getting something down at all. So do whatever you need to do to get your own thoughts on the page (except Claude, obviously), and then proofread, correct, and revise. If you do substantial revisions, then you'll have to do more editing passes. When you do a pass and only make minor changes, that's a sign that you're probably done.

If you think you're just a poor writer...well, you won't get better without doing it a lot, will you? Also consider that AI assistance will actively hamper your growth, and that's considered a feature, not a bug (would the AI labs prefer you to depend on their products, or to outgrow them?).

On spelling and grammar, you can't get flagged as AI for using spellcheck. I don't use grammar checking, but I read somewhere that grammar checking algorithms have gotten good now, so that could be something else to look into (as long as it just highlights potential problems and doesn't rewrite your words).

This analogy makes no sense to me. Being a good manager requires strong interpersonal skills. "Managing" your team of agents does not make you a leader.

Interesting essay! Yeah, I think there's some truth to the premise. As it becomes cheaper and cheaper to exploit predictable patterns, a person's value-add needs to include unpredictability. That's a necessary and not a sufficient condition, though, since you can be unpredictable and still mostly useless.

> Obviously, there is nuance. You can take this argument to the limit and say that ultimately the world is made of only two things: Chaos (a.k.a. the vast seas of Wolfram's computational irreducibility, a.k.a. pure noise, a.k.a. quantum randomness) and Logos (the small islands of compressible patterns).

Within computational irreducibility, there's pure noise as well as universal computation, which is neither meaningfully predictable nor random. Since you mentioned Wolfram, in terms of elementary cellular automata, it's the difference between rule 30 and rule 110. The latter seems to me to be on the boundary between Chaos and Logos, and where I would expect to find the human intellect. (I think this is also largely the same distinction I made above where unpredictability can be useful or not.)

Since you also mentioned the economy and the possibility of UBI...I also don't know the answer, but I think it's worth thinking about Maslow's hierarchy of needs, which to me is pretty tightly coupled to weirdness. Moreover, I think societies similarly have their own hierarchy of needs, but that's something I don't understand very well. In any case, weirdness seems to be coupled to both of these, in that it best flourishes at the higher levels, when the more fundamental needs are mostly taken care of. At the lower levels, there's too much pressure to reduce social friction to cooperate with others for the sake of survival, and you just can't really afford to be too weird. I don't know if UBI is the answer, but there may be some benefit to just solving the lower levels of the needs hierarchy so we can move past this.

In December 2002, I had studied programming for one semester. I was experienced enough to have read and mostly understood Dijkstra's original essay (and probably had), but not to fully grasp its import. That's all to say, there's probably a lot of context I don't have regarding what this essay was reacting to. So I'm going to respond to this essay from 2026, ignoring whatever may have prompted it. Its basic thesis is that "considered harmful" essays are needlessly inflammatory and we should instead engage in more balanced critiques. My response is:

1) of course balanced critiques are necessary, but

2) sometimes you can come to a conclusion after considering all the pros and cons, and in those cases, you should definitely say so and not equivocate.

IMO that's obviously what Dijkstra did. I don't think he just had some axe to grind against `goto` for his own personal reasons. During my career I've encountered my own share of practices that I think are overused and misunderstood, and I don't need to say what they are (and thus hopefully I'll avoid a flame war), but if you've been at it for more than a few years, then I'm guessing you have some similar opinions.

I actually have nothing against "considered harmful" essays as a whole, and I often learn something from them. That said, of course you shouldn't overstate your position, which would be intellectually dishonest. For most of the problematic practices I've encountered during my career, there are some valid use cases, and it's more that they're misunderstood and misapplied that would warrant calling them out.

Anyway, I hope I've made my point.

Keep at it!

Yeah, I've been thinking this as well. I have some optimism, but not with high confidence. The exponentially increasing power of computing hardware up until this point is often cited as the reason performance optimization has been sidelined in the software industry. Now that there's a definite hiccup in that trend, I'm hoping programmers will remember that software actually can be fast and memory-efficient, and that poor design choices that lead to bad performance are exactly that, a choice.

The author of the article definitely seems to think LLMs are what enables this to happen, but I personally am much more skeptical of that. I think what it really needs is bringing engineering back into software, not just throwing LLMs at it and calling it a day.

Sorry but this is pretty bogus. (-1)^x is only well defined when x is an integer. This is generally the case for r^x whenever r isn't a positive real number. For example, when x = 0.5, r has two distinct square roots. Sure, you can choose one of them arbitrarily and declare it to be the value of r^0.5 (and math libraries typically do this), but there's unfortunately no good way to make this arbitrary choice consistently for all values of r simultaneously.

You do in fact need the complex exponential to define this correctly because the function a^x for nonintegers x is only unambiguously defined when a is a positive real number. For example, your function could be either e^(pi i x) or e^(-pi i x), which trace the circle in opposite directions as x varies over the reals. (They happen to agree when x is an integer.)

While I generally agree with what the author is trying to say, I have to point out that it's largely AI-generated, by the author's own admission. TBH it's pretty ironic given rule #1, and why limit "write in your own voice" only to the people "you love" anyway? Just write in your own voice, period. If it's not worth your time and effort to write, it's not worth anyone else's time to read.

I've used Taylor series in numerical optimization. A function we were implementing needed to be differentiable (for automatic differentiation), but its definition had a special case, so we used a couple terms of the Taylor series in the special case.

edit: Sorry, to clarify, this was a function involving trigonometry but not simply vanilla sine or cosine. However, angular values being represented in radians did help in the same way I described in the parent post.

I think I cautiously agree with this notion to some extent, but IMHO the real answer is that it's application-dependent, and if you're writing a low-level trig library and you have to pick one or the other, it really isn't clear to me that turns should win over radians.

I expect many systems that use trigonometry would sometimes use small-angle approximations either for efficiency or to bootstrap to the general case. It'd be natural to use Taylor series here, i.e.:

    cos(x) = 1 - x^2/2 + ...
    sin(x) = x - x^3/6 + ...

If you've committed to representing all trigonometry in "turn" units, then you instead need to use:

    cos(2 pi t) = 1 - (2 pi t)^2/2 + ...
    sin(2 pi t) = (2 pi t) - (2 pi t)^3/6 + ...

In this case it would be less accurate and efficient to force everything into turns if you ever need to work with radians.

Closely related to this, if you ever need the derivative of a function that does trig (e.g., in numerical optimization), you may as well use radians because if you don't, any extra factors you apply will appear in the expressions for the derivatives and you'll have to deal with them there anyway.

Basically for that reason, it's pretty clear that trigonometry in terms of radians is the "correct" convention mathematically speaking (away from computers), since derivatives of the radian-based trig functions are so easy to express. Given that, if we have to pick one convention...isn't it less confusing to use the same thing everywhere? That said, there are interfaces that provide both versions, and since as the article points out there are cases where the turn-based versions can be more efficient, that's probably the right way to go.