If you dislike AI please stop saying that AI can code now.
-
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko well it generates code. I wrote deterministic code generators for EMF-based models. LLMs generate code in a more - hm - probabilistic way.
Anyway, coding itself requires the understanding of a subject and then formulating code that reflects it. And most LLM users also need to get the understanding part. That's the hard part.
-
@Tereya @eniko If it make you feel better, I've had access to existing tools that could do the decompiling and relabeling people are excited about since around 2016.
Today's pleasant answer machines expand access to tools most folks might not otherwise discover - which itself is interesting and some times quite useful.
But "more people notice computers can already do X!" doesn't make much of a headline, so we get sensationalism, instead.
@EdTheDev @eniko The main catches I’m aware of is it only really works with older (>15y/o) software and is still expensive. It also seems like one of the more plausible use cases for LLMs since the output is computationally verifiable.
I know some of the “decomps” like for Photoshop weren’t really decomps, but it seems there’s enough actual utility to be concerned with while LLMs are still being subsidized by capital.
-
@Tereya @eniko I guess I'm not even sure what would be "concerning" regardless, but generally the results are really bad, far worse than decent non-"AI" tooling for reverse engineering. If there are cases where it's spitting out something that looks like the real source, then almost surely the real source (or other sources that the real source plagiarized) are in the training corpus. That would explain your perception that it "works on older software".
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko today I encountered someone advocating for replacing "vibe coding" with "cuck coding" because the users like to sit in a chair and watch
-
@tuban_muzuru @eniko developing that spec is what programming is
... anyone would think it was about generating the code, the way these maniacs run around and complain about AI Gonna Take Yer Jerbs
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko when I make it generate code, it's usually snippets of 10 lines, and there are always 1 to 2 obvious errors in it. The thing cannot reason so gives the most probable code, but that's not the right code.
It's about the same process than to search stack exchange and copy paste the code from there. -
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko Yep, 100% agree with this.
-
@Tereya @eniko I guess I'm not even sure what would be "concerning" regardless, but generally the results are really bad, far worse than decent non-"AI" tooling for reverse engineering. If there are cases where it's spitting out something that looks like the real source, then almost surely the real source (or other sources that the real source plagiarized) are in the training corpus. That would explain your perception that it "works on older software".
@dalias @eniko My main concern is that copyright law could get even more draconian if/when LLM decomping gets to the point where it can work on modern software. I'm not sure how the law would shake out because it seems it would require legally declaring that LLMs violate "clean room" practices which would hit every big tech company in the crossfire.
https://momo5502.com/posts/2026-10-09-game-decompilation/
I can't find some of the more impressive stuff I saw before (bad sign for my memory, better sign for state of things), but decomping seems like it's in the sweet spot of what LLMs should theoretically be able to do. Even if it takes a lot of tokens/iterations at the moment (700 billion tokens for this article's project), because the output can be programmatically verified as correct, there seems to be a path to viability here that isn't just smoke and mirrors.
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko I generally say I'm not ready to consider and discuss capabilities or lack thereof of the tool before we discuss all the stuff around how those tools come into being.
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
I can write a tight, neatly-coded set of instructions for a specific task which is very specific in what inputs it accepts and in what outputs it returns.
I can write these instructions such that they are not single-purpose, and can then be used for more than one fixed set of inputs and one set of outputs. This is what is called 'object-oriented programming'. The set of instructions can then be used for multiple purposes within the code and called repeatedly in different contexts.
I can further optimize any sets of code such that the end result of my work is not merely 'input x -> output y', but that it operates efficiently and well.
To call the output of AI equivalent to what I do in these instances is not merely insulting, it is murderous.
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko Thank you for saying this as firmly and as loudly.
Perhaps slightly a tangent, but it really makes me think about how much people need to remember that the slop side of the fight are, in general, lying liars who lie... even when something seems impressive at first glance, it's just all so much slight of hard.
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko AI generates code based on hundreds of thousands of attempts by human developers to solve similar problems that match the prompt the AI was given
Some of those attempts were successful, many weren't & were posted as "pls tell me what I'm doing wrong", many are shitposts & some are so wrong they barely qualify as code
The AI can't discriminate the good code from the bad
The AI hoovers them all up & extrudes a statistical average which has more chance of being bollocks than being right
-
@Tereya @eniko If it make you feel better, I've had access to existing tools that could do the decompiling and relabeling people are excited about since around 2016.
Today's pleasant answer machines expand access to tools most folks might not otherwise discover - which itself is interesting and some times quite useful.
But "more people notice computers can already do X!" doesn't make much of a headline, so we get sensationalism, instead.
@EdTheDev @Tereya @eniko In fact "people notice computers can already do XYZ" is the whole media hype today (X=coding, Y=reverse engineering, Z=etc). And I'm beginning to doubt whether the "computers cannot code" stance is still valid. Sure, computers still need a lot of good guidance, but... that's hardly a new thing in dev teams.
-
@EdTheDev @eniko The main catches I’m aware of is it only really works with older (>15y/o) software and is still expensive. It also seems like one of the more plausible use cases for LLMs since the output is computationally verifiable.
I know some of the “decomps” like for Photoshop weren’t really decomps, but it seems there’s enough actual utility to be concerned with while LLMs are still being subsidized by capital.
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko GitHub is absolutely full of abandonware projects that someone AI-generated in an afternoon and then couldn't figure out how to maintain
-
@Tereya @eniko If it make you feel better, I've had access to existing tools that could do the decompiling and relabeling people are excited about since around 2016.
Today's pleasant answer machines expand access to tools most folks might not otherwise discover - which itself is interesting and some times quite useful.
But "more people notice computers can already do X!" doesn't make much of a headline, so we get sensationalism, instead.
-
If you dislike AI please stop saying that AI can code now. AI can't fucking code. Just because it can generate code that works for some very loose definition of the word works doesn't mean it can code because coding is so much more than that. Please, I'm begging you, you do not, under any circumstances, gotta hand it to them
@eniko If code that works isn't code, then what is code?
-
@EdTheDev @eniko The main catches I’m aware of is it only really works with older (>15y/o) software and is still expensive. It also seems like one of the more plausible use cases for LLMs since the output is computationally verifiable.
I know some of the “decomps” like for Photoshop weren’t really decomps, but it seems there’s enough actual utility to be concerned with while LLMs are still being subsidized by capital.
@Tereya @eniko I can see that.
To me, the exciting impact is software preservation for history. I suspect that we lose some little piece of interesting source code every day due to a combination of licensing hell and forgetfulness.
I welcome better tools to restore lost code from whatever binaries remain.
I know that the current unsustainable subsidies will end, but with the right experts in place, the same technique should be possible to do slower on reasonably priced hardware.
-
@EdTheDev @Tereya @eniko In fact "people notice computers can already do XYZ" is the whole media hype today (X=coding, Y=reverse engineering, Z=etc). And I'm beginning to doubt whether the "computers cannot code" stance is still valid. Sure, computers still need a lot of good guidance, but... that's hardly a new thing in dev teams.
I had my computer writing all of my boiler-plate code back in 2001.
Computer-assisted coding isn't particularly new.
Folks tell me how incredibly different it is this time, which is valid. But, of course, most of the folks telling me that don't know how impressive my code assist tooling was in 2001.
