Glyph
RE: https://eigenmagic.net/@vampiress/116859241032699772
“I poisoned my own mind.”
More and more programmers are writing retrospectives like this — and a huge THANK YOU to Elissa for doing so — where they discover that Programming Is Not Special.
It is particularly interesting to read this from a solo game developer who is both a programmer and an artist, who began by thinking the programing isn’t art, and discovered by an emotionally torturous process that it actually is.
Thomas 🔭🕹️
@glyph @vampiress Programming has (like many other human activities) aspects of both artistry and craft mastery and boring grunt work.
I think LLMs can be a useful tool for some parts of it, but you do need to be prepared to become a manager to successfully use it—that is its own skillset and oddly is more like a “people skill”.
Even then, it’s questionable at best how much LLMs will really help you long-term with software that you have to maintain. Many parts of the grunt-work literally lay the foundation for the whole and lack of deep knowledge will come back to bite you later.
Rich Felker
@glyph @chroma Indeed, there is no acceptable level of usage that's not harmful. The only endeavors they actually work for are ones that are inherently harmful/antisocial. Things like propaganda/disinformation at scale.
Glyph
@chroma My position is actually more maximalist than yours, apparently; I don't even think they should be used as autocomplete, certainly not as much as 10 lines at a time; that just creates an opportunity to gloss over 1-character bugs without even really boosting productivity much. I am not even sure that search autocomplete is safe https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4535163&__cf_chl_f_tk=jNrJBWpgluZQNLfU_zryzI22167PyWPRvLUGILwdQMA-1783179018-1.0.1.1-7Cazxa1x8rp3flYrUDPfOao56Hj5cZyOxqFV5zWTgn0
Glyph
@chroma "Maybe I can go back to them in future, ***starting from scratch*** and doing them the right way" (emphasis mine)
The author is suggesting that the ideas are now ruined by LLM residue, and that she needs to shelve them until she can forget the dispiriting sludge that they produced, doing them again by herself ***from scratch***, i.e. *NOT* using an LLM, but with the same *idea*.
명준
@glyph My bad, I misunderstood that paragraph’s idea by reading diagonally and fast.
But I disagree, the only viable use of LLMs is multiline (5-10 max) autocomplete after extensive training with a specific repo tbh.
They’re just glorified autocomplete, any other use out of this is pure brain rot that isn’t far from Markov chains garbage.
Glyph
@chroma wait why do you say this is not a great ending? what door is still open?
명준
@glyph read the quoted paragraph, I will quote it again, with emphasis: “back to them in the future (…) doing them the right way”.
She suggests there’s a “right way” of doing LLMs.
This kind of person doesn’t learn anything, they still think it’s “a tool”. They still jerk off to the false and despicable idea of compressing days of work in seconds regarding creative endeavors.
명준
@glyph “I poisoned my own mind. Maybe I can go back to them in future, starting from scratch and doing them the right way, but for now, they’re shelved.”
The door is still open, not a great ending for the post. I don’t get why people are so tempted by laziness and cognitive decay.
Rich Felker
@bonkers @glyph No, somebody else you're plagiarizing from did the heavy work.
The slopbot did the copy and paste, the search for examples to smooth over the seams, the stripping of attribution.
Bonkers
@glyph I wouldn't have started those projects, as they need a lot of work. The clanker did the heavy lifting, so now we have the tools I wanted to create one day.
Glyph
@bonkers we don’t though. you just said you don’t need to
Bonkers
@glyph besides, Starlinks are the only reliable means of communication at the front lines in Ukraine. Even that they sponsor Musk, they're saving lives.
The world is fucked, and we have to use a tool that works right now.
Glyph
@bonkers even if you don’t come around to write such a post yourself in a year, you should still probably refrain from using them for a whole host of reasons, not least of which is that giving money to claude is giving money to elon musk
Bonkers
@glyph I agree it's a dilemma, and I hope that the bubble bursts soon. Also, if I had to pay the full price of LLM, I'd refrain from using it at all.
In my defense, these are nonprofit projects for the public good. They can easily be continued without the slop engine.
Bonkers
@glyph (a software architect with 30 years in the industry here) I have also made a few prototypes that would take me forever to code by hand. Claude just made them in a few hours. Honestly, I wouldn't even have started most of them.
My outcome is different from what Elissa says. It's been fun at first, as I didn't use LLM before, and it's still fun now, as one of the projects is production ready, and I'm actually using it. The others aren't ready, but only because the engineering task is difficult and I haven't figured out the solution yet. But it's still fun. Of course there are boring parts, like telling the clanker that it forgot to update the documentation, or other things, and please update your bloody agents.md so that you remember next time.
So, overall, it's been a useful tool for me and it saved me tons of time. Awesome result for side projects that will never be commercialized.
Rich Felker
@bonkers @glyph No, it is not. You've just bought into the delusion.
Bonkers
@dalias @glyph nope, it synthesizes a lot, but it's a new piece of software. The clankers have advanced pretty much in the past couple of years.
Glyph
@bonkers @dalias I consider asking me to look at LLM output to be an aggressive, antisocial act, which I have detailed in <https://blog.glyph.im/2026/06/adversarial-communication.html>. I am not going to block you immediately because you may not have been aware of this but please do not do so again.
Bonkers
@dalias @glyph here, you can try proving your claim yourself:
https://github.com/clackups/draftling
Rich Felker
@glyph @bonkers I consider it the same, and I can explain why.
Asking me to read the output is asking me to subject myself to the illusion and evaluate what the machine did in a way that's impossible to determine from seeing the output; it would require me to just make assumptions based on my perceptions. A process which, by design, shapes your mind to buy into it.
The only reasonable way to make a conclusion here is from looking at the mechanisms or the high level information-theoretic invariants underlying them.
Rich Felker
@bonkers No, you made a bold claim. I made the obvious default claim.
Bonkers
@dalias so, you made a bold claim, but you refuse to prove it on a rather limited code base. It just means that our discussion doesn't make any sense in the first place.
Rich Felker
@bonkers The kind of comments in LLM codeslop are the opposite of useful. They do not document motivations why something was done the way it was, connections to external contracts, etc. They're imprecise restatement of what's in the code. The same bad types of comments novice human codemonkeys write when told to comment their code better.
Bonkers
@dalias there's one significant difference from human written code: the amount of comments. Humans don't document as much as the machine. It helps the machine in adding more code and it helps a human understand the logic.
So, my bald claim is that it's a useful tool that's not going away any time soon, so it's alright to utilize it if you know what you're doing.
Bonkers
@dalias I made a useful product that I'd never make by myself, as it would take months to code. I reviewed the code and steered the clanker from its mistakes. It is a new piece of software. If anyone finds a plagiarized piece of code, I'll be glad to add the credits.
I don't even need to defend myself, the code is free and there are users already.
Rich Felker
@bonkers Of course I didn't.
Bonkers
@dalias I assume you didn't open my link
Rich Felker
@bonkers No, it's not a "bold claim". There is no model of reality or how anything works in these systems, only pattern regurgitation in ways that correlate/align with context. If you get any comments that appear to explain motivations, they're copies of someone else's ideas from a similar context, not something reasoned to apply to your code.
The way you support or refute what I'm saying is through examination & analysis of the mechanisms, not examination of the output and whether it looks like what you would expect from human reasoning. "Looks like human reasoning" is the deception working as intended, and if you're trying to assess it from that standpoint, you're falling for the brain rotting program.
Bonkers
@dalias so, another bold claim without any ground.
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.