Evidently, building this sort of program is no longer “hard”. LLMs now enable a dedicated solo dev to build a raw editor! That’s really bad news for Adobe, but otherwise it’s more Open Source software for the rest of us.
I struggle a bit with your implied derision of “ai vibed raw editor”, which others have expressed as well:
So far, RapidRaw, Safelight, and FerrumPix seem alive and well-maintained, by motivated, technically proficient developers. They obviously put thought and care into their software. By all accounts, these are not vibed slop apps, but well-designed programs by capable software developers. The fact that they posted them here, on a dedicated Open Source forum, is a positive signal as well.
It is I think far too soon to cry wolf. Evidently, LLMs allow new people to tackle harder problems. We live in strange times. Values change extremely quickly. What counts today, more than anything, is proof of care. But as far as I can tell, these three programs clearly have shown that their developers do care!
This is Open Source software. Plenty of great software is maintained by solo devs, either by choice, or because it’s a small niche. Not every program needs a community. LLMs allow solo devs to scratch their own itch. I’ve done it myself, too! Both building my own experimental raw editor early last year, and LLM-extrude various bespoke pieces of software for my own use. Among all the terrible things AI companies have wrought upon this world, this stuff isn’t one of them, I reckon.
This stuff is from them, this stuff enables them, this stuff is indivisible from them.
Hey Anthropic, the supposedly “more moral” one, is building a dragnet precrime prediction dragnet to track their detractors. You should definitely keep buying tokens from them.
I address this in the OP. The difference between a solo project developed over years is that the developer had time to develop a strong mental model of his project. No developer no matter how good he is can fill his brain with 100k vibe coded lines in a few weeks, and hence he will always be dependant on LLM assistance, or at least be nerfed by the lack thereof.
Yes, perhaps I didn’t explain myself properly. It is definitely possible to draw a line. What I mean as not avoidable is that none will do it and if they do they get forked. Even projects who once resisted it are succumbing to it now. It’s only a matter of time where every large project accepts it in some way.
All these three AI vibed raw editor seem to be the same thing: a Lightroom alternative that is open source. They all offer AI features, but then frequently fail to properly to do basic things like provide proper lens correction.
And also, it seems there is more interest in starting something new from scratch every other week than helping improving the vibed editor that is already there. Like, it is nice people want to contribute to FOSS, but the vibe (hahaha) I get from them is that they just want to be the star behind from a new brilliant open source project instead actually contribute to developing good FOSS software.
IMHO, we need better FOSS software more than we need just more FOSS.
That is my negative feeling towards these projects.
But I could be wrong, as I am just feeling the vibes here.
I have no knowledge of coding of any sort but the idea of “let a thousand vibe-coded flowers bloom” seems like something we’ve seen before in the pre-social media internet. That is diversity, as beautiful as it may have been for some, leads to confusion and cost-of-discovery problems for the majority. The result might be a dominant player shaped hole waiting to be filled, whether that’s FB, Instagram, Apple, Google, Spotify or whoever. Maybe the network effects aren’t there this time but I’m sure clever people are trying to figure out how to build them.
Well, perhaps I am being a little too fatalist. I agree with you that there will always be examples of projects without LLMs, I mean, my very own image viewer is such a project, even if it is a one man show and I doubt even a single other person uses it.
I just believe (sadly) that eventually most of our software will be partly LLM written.
I guess new groups will form, much like today we have suckless, gnu-less linux systems, bsd, people trying to keep X alive, and so on, code purist groups will form.
Text editors are amazing, I have a suspicion that much like handwriting has an influence on how you think and perceive information, so do emacs, vim, etc
I tried a couple of them, only to realize how outstanding the existing ones (Darktable, Rawtherapee, ART) are.
Make no mistake: current AI (LLMs) can produce useful code, but still require a knowledgeable adult in the driver’s seat. Eg writing a raw editor would require at least the following skills:
understanding of color science,
understanding some areas of applied mathematics,
ability to write fast code, for CPU and GPU,
how to design a lasting code architecture (IMO this takes decades of wrong turns to acquire),
how do design a sidecar format one can grow into with future versions,
UI design,
UI implementation mechanics,
setting up comprehensive unit testing and benchmark frameworks.
The challenge is that it is very rare for a single person to know all of these things. Now it may be possible to use AI for guided learning (like a mentor you can ask questions, and get references to read), but IMO it would still take about a year of person-hours to get to the point where one can supervise an AI authoring code. The very least.
I don’t think it really does that, on average. Yes there might be some edge cases, but I think LLMs just shift the goalpost to producing the first 80% really quickly, and ignoring the last 20%. And people will come to expect that, and it’s even seemingly becoming an amusing hobby of people here to test these Raw developers and tinker with them, rather than actually focusing on the long-term. LLMs inevitably condition that sort of behaviour. They condition people to “open the new package without learning in depth” because why learn in depth, which is less fun, when you can just open a new package every other week?
In my opinion, whether or not AI is used, very rapid software development and innovation past a certain speed is a mistake and leads to this. Same with YT. Anyone can make a video and maybe yeah, some are helpful and interesting but it has shifted the way people interact with media and it and such platforms have made people more scatterbrained.
Even if AI didn’t exist, I would say that any endeavor or desire to improve things shouldn’t go too fast, because beyond just the development, it modifies our thought processes.
Most people miss these more diffuse negative consequences of AI because they’re only focused solving specific problems, a mental model that’s been foisted upon us through the emergent phenomenon of rapid technological development. But we should be concerned about them. And as AI is the very embodiment of this foolish direction, it’s more than enough reason not to use it.
Of course, no one wants to even think about that because hey, there will be ten new updates for superficial amusement next week and what’s wrong with that?
But it’s just like a drug now, and it’s become something tangible beyond just software.
If I was a developer using AI, the nightmare… Look deep into the code and find a picture of a revolting round-faced chubby child, with curly golden hair, podgy pink cheeks, too-big eyes, exaggerated eyelashes. Dripping with fake cute.
Otherwise, though, as long as it is properly supervised I’m sure it’s fine!
(inspired by a poisonous experience on Instagram!)
My “AI-coding-assist” experience has been mixed. Just yesterday, had Gemini present a solution to cross-platform pipe code, actually found it in the cross-platform fork code I was using (reproc) to run exiftool from my program. I was about to engineer something myself, glad I asked first.
But last week, had Copilot of all tools tell me libtiff couldn’t load/save the ‘extra’ EXIF tags, fnumber, exposurespeed, etc., when I knew good and well that it did. And then, a few months ago the non-trig debacle of parallelogram divinity for the CAD modeling I was doing, code was incomprehensible to me, results looked right from afar but turned out to be wrong on close inspection.
So, small, specific questions, and a good bit of evaluation. Ain’t gonna let it write my next project,node-based mesh/image/polygon gonkulator, wish me luck…
I would argue with this. It is specifically in software where LLMs seem to have found their most lucrative application. This just happened in a few years, and without it the whole sector would be nonviable. Kind of a last minute save, though it is dubious whether it will keep the big players afloat once open/small models mature.
Yes, indeed for some people they are like a drug. Just like computers. But we learned to live with those too.
Contrary to its title, it’s about the possible futures of programming, the history of other technological revolutions, and the curious lack of ethics in our field.
I would argue, less powerful a drug. We learned to live with many levels of technology, but we can also buy many weapons without a license - up to a point. After some level of power, the power becomes too great. Just like one can buy marijuana in Canada legally, but not heroin.
It’s really low-hanging fruit, given that most of the corpus of software knowledge lies in public formal specifications of programming languages, and documented libraries. Other subjects, less-straightforward and very context-dependent.
Precisely! That’s the direction the whole thing is going. Retrospectively:
“We are going to create God” (before 2020)
“We are going to replace human workers” (before 2024)
“We are giving you a tool for automating the repetitive parts of programming” (now)
(the progression reminds me of a classic joke in Hungarian, with the punchline “toothpicks”).
Also, there two side quest, at the moment:
“We are making a Doomsday Machine. Dear government, please regulate entry to the sector so that we make it slower. Incidentally, do you want to buy it?”
