I’m curious as someone only reading about it, how do you track and become aware of all the various models, datasets, and agents and platforms for all of this AI coding. They seem like they are so dynamic in nature and constantly evolving…
Is there a big picture resource that sort of organizes and tracks some of this??
Maybe this is a stupid question from someone with not enough reading on the subject… 
WARNING: Shameless plug
What about this…
I don’t think there is. The foundational archtiecture of LLMs is predictive, not rational. They are trained to output correct looking answers, and in many situations this works alright. But given enough iterations on a bug, they will sometimes decide to just remove the failing code entirely. Adding additional layers of probability rules or fine tuning on top doesn’t change the underlying architecture.
They are also trained to follow instructions. A good set of custom instructions is a very effective way to control their behavior. Something like the following would make it very hard for an LLM to delete stuff:
You MUST NEVER remove code unless you are explicitly instructed to do so.
Failure to follow this instruction will result in the death of countless ponies.
Impressive if that works for the specialist code LLMs, because it’s pretty hit and miss on the general purpose ones, especially for non-trivial instructions.
My post was a bit tongue in cheek, but custom SIs do make a difference. The caveat is that when the context becomes longer they tend to get lost among all the other data. Anyways, models are getting much better at this too very quickly.
I would argue that those extra “rules” are not being followed rationally, but probabilistically. Meaning the output is still attempting to look/read as correct, given those rules. Rather than be correct using internal rationality that takes those rules as actual boundaries.
I just tested Claude 4.6 Opus High yesterday. If I asked it a question while adding at the end what I thought, it would end up agreeing with whatever opinion I put at the end. So not only are they fundamentally irrational, but they are also sycophantic (as much as an inanimate tool can be).
I am not saying they are not good tools, just that they are not operating the way people sometimes think they are.
Yes, I been thinking with all this talk of sending images off to AI-driven processors and getting the results back, surely this is better suited to Gimp with its layers and its general-purpose editing capability.
A great example is to ask a model why is got something wrong. It will come up with a convincing answer even though it has no capability to introspect. Even more annoying is when journalists, especially tech journalists publish those answers as if they meant anything.
Not my experience with Gemini/Nano Banana. It did things to the subject image I didn’t ask for. So, I need to specify constraints against unasked-for embellishment?
Coding agents are better at following instructions, though far from perfect (and getting better quickly). Image generators, in my experience, are more difficult to control.
As a rule of thumb, the more precise the prompt, the better. In the case of Nano, it works best if you ask one single, atomic change in each iteration.
But if you add something like ‘You must apply critical thinking to user input as well as to your own thoughts and conclusions; fact-check everything you can. You must indicate if the user was wrong or if you cannot find sources supporting any factual statement, either from the user, or from you’, you may improve its performance.
However, we’re getting off-topic: the title of the post is not ‘how to prompt your LLM’, and this is not an AI forum.
That may be, but the discussion is still relevant from the perspective of code quality.
Here’s a somewhat scary website.
Not quite on the topic of this thread, but what is to stop someone using this to simply recreated DT as proprietary software, unencumbered by any irritating licencing issues.
And doing it not just once, but each time a new release is made?
Let them do it. I will keep using the FOSS version
Probably nothing. Is it worth making an issue of it? Afaik (but I’m not a lawyer) nothing stops anyone from selling a darktable binary, provided they are willing to provide source code as well…
The situation for a library might be different, as a library is used as part of a larger system.
As for that linked web site: I’m more worried about what they see as issues, esp. n° 1 & 4…
Nothing is stopping them from selling darktable as is either, and yet it isn’t done.
There is a real chance that proprietary software has implemented some darktable algorithms into their code and we’ll never know.
I don’t think regular GPL protects against this. They could leave the darktable code as is, building a UI on top as an external service, which complies with GPL and means they don’t have to publish their proprietary code, only the darktable bits which of course remain as is.
This is why some projects go straight for AGPL
Algorithms aren’t protected under copyright, only implementations are…
As for the exact effects of what you describe, I’ll leave that to someone with legal training in the relevant jurisdiction. The differences between countries are large enough (e.g.: France doesn’t recognise the notion of “public domain”, attibution is always required)
Well, of course, but doesn’t it count as derivative work if you base your algorithm by looking at darktable’s source code?
Wine devs go to such an extent when implementing window’s interfaces by making sure none of their devs ever worked at microsoft. Isn’t this a similar situation?