It just so happens that many video codecs are based on image formats, so ffmpeg already has a lot of the complex machinery to do so available to also implement these image formats - internally it can just handle it as a single frame of video with specialized formats for that.

Imagemagick (and other tools) also work, but why use multiple pieces of software if what you already have is adequate? ImageMagick is also software, and can also have vurnabilities.

That’s fair. I actually don’t think we disagree that much - I just think I have trouble conveying what I am trying to say. Whenever someone talks about ‘shallow statistical predictions’, I think about older techniques like Statistical Machine Translation which even had trouble with things like word order, LLMs handle text on a higher level of abstraction (which I described as a form of textual understanding) - and hence handle things like word order better - but are still inherently statistical predictors. The model stores info about how words interact and relate to one another, but it does not ‘understand’ what the words actually (physically?) represent beyond these interactions nor does it ‘understand’ what it is doing. Albeit, those interactions are modeled well enough to give a convincing replica of doing so.

That is no more understanding the problem than my spelling or grammar checking understands the comment I’m writing.

My general point would be that even a grammar checking can have some form of understanding of the text, no matter how shallow. The checker probably has a rule for when a is used versus an, if this rule generalizes across new words that rule that were previously unseen is a form of ‘understanding’ of the language being used in my view, despite being overly simplistic, while rote memorization - having a list of words that are followed by an - may not be.

LLMs are a weird case, because their internal representations for many concepts generalize even across new settings / inputs - in that sense the model has a form of understanding of what is being given, while for many other concepts the patterns break down even in the simplest of cases. It may ‘know’ that the preceding text is the writing of an essay - and it should autocomplete accounting for that -as is defined in its weights, but this understanding is shallow - it does not know why it knows, or how it knows - it cannot self reflect as it does not see and understand its own internal workings, or account for that. Yet the internal representations represent a form of text understanding that can be useful nonetheless - it is a language model after all.

My comment was intended to show this duality, hence the duality between the two paragraphs.

And this is the crux of my beef, I think, because stochastic pattern matching is not understanding, it’s a mathematical representation of how the model processes your input tokens. The fact that it has to start over every time you provide it input, and uses the previous input/output tokens as context is why this is not ‘understanding’, it’s just fancy token prediction that gives a middling-to-passable facsimile to intelligence and understanding things.

The problems you note in your second paragraph fundamentally undermine the argument that there is any form of understanding to the AI, because those are basic mistakes that a trivial understanding of the problem would prevent.

I am not entirely grasping the point you are trying to make here. I am certainly not arguing that it is conscious, self-aware, or in any way not a mechanical procedure that is being performed (I would not argue for that!). My key point is that it is not a simple black or white it understands / it does not understand - it may have internal representations that relate many concepts together, allowing it to draw upon these links when generating text, giving it a certain semantic understanding of the language and text it is using, while simultaneously not having a bit of self-awareness.

These models do have some form of understanding though. There are features for bugs and typos, and general features that map descriptions and pieces of code. It understands the code in so far it helps with next token prediction.

The bigger problem is that these language models are inherently unreliable and stochastic in how they generate. You request a feature - and it destroys something else in the process - because a single incorrect prediction caused it to diverge and skip a portion of the original code. You request a small modification - and it decides to restart from scratch - because random sampling made a different way of doing something more likely rather than what was there already. Errors compound and the model has no way of fixing or correcting them.