...I was really ready for the "contempt of court" charge that they were 100% going to end up with.
Joshua Browder, CEO of DoNotPay, made headlines for claiming an AI chatbot was due to defend a man in an upcoming court hearing, but has pulled out of the stunt. Browder runs a consumer rights startup that was originally built to help people appeal parking tickets more easily, and has since grown with the aim of building "the …
I wonder what's being said about AI in corporate media empires like Sony, Universal and so on. If they think it'll make them money, I'm sure they're in favour. I do wonder how annoyed they'd be if someone trained an AI just on (for example) Taylor Swift and then started releasing "new" Taylor Swift songs into the public domain. I can't imagine they'd be particularly happy about that.
I worked with someone last century whose research thesis involved turning code into music in the expectation that good code would sound good and bad code would be discordant.
He was also a 'top ten' trance music composer and had a side-line in lift music.
I can't think that AI music will be much better.
Icon: because that seemed to be everywhere at the time :-)
It's perhaps of some interest that Dirk Gently was published in 1987, and that same year David Cope published his first two papers on computer-assisted music composition. Clearly it was part of the zeitgeist.
Even in the late 1990s, Cope had systems that were creating music which fooled human judges. It's similar in some ways to the situation with Philip Parker and Icon Group, which has been very successful selling specialist computer-generated books. In well-defined niches, machine content generation has been working very well (at least in economic terms) for quite a few years.
There are also plenty of computer-generated novels, short stories, and poems in print – you can find any number of lists and articles about them online – but they've generally been less successful.
Back to the original topic: It's not difficult to imagine an architecture for deriving music from source code that would represent aspects of the code. Training even a now somewhat old-fashioned deep CNN stack (or a transformer, if we must) to extract salient features, then using them to prompt a unidirectional LLM or a stable-diffusion model that was trained on music rather than text or images ... seems pretty straightforward, really. By itself it's maybe a moderately interesting grad-student project.
The output looks good but you would be foolish to trust anything in it that resembles a fact.
Since this is also true of most humans, I'm not sure why you think it's an interesting thesis.
Like LeCun, I'm not particularly impressed by ChatGPT. I know enough to know it's not particularly exciting. In the realm of LLMs, I like the bidirectional BERT family more than the unidirectional GPTs, and even for the BERTs I feel a bit meh. Yeah, transformers are a bit more elegant than some of the previous work, but I want to see something more like the heterogeneous-competency architecture of EfficientZero in a language model. And something rather more resource-efficient. (Also, as always, what the public sees is already some way behind what's coming out at conferences and on arXiv...)
But I'm also tired of the ignorant hand-waving dismissals of LLMs and other big ML projects. Arguments like "it's just a statistical model" are irrelevant; a statistical model can be a non-deterministic UTM and so compute any computable function (assuming the classic CTT), so if you want that argument to count for something, show me the human CNS is strictly more powerful than a UTM. I'm not holding my breath. Or "it just regurgitates what it learned": show me that's not true for human beings.
I've studied and taught composition (writing) and rhetoric at the graduate level, and I've studied natural language processing and machine learning at the graduate level, and while I certainly don't know nearly as much about these areas as someone actively working in them (who could keep up otherwise?), I'm pretty confident that the vast majority of comments I see about LLMs and similar – whether fawning or dismissive – are just as much "shiny bullshit" as the output of those systems.
Assuming it is legal for a participant in court to be getting instructions via headphones (which I'm skeptical of) if they had let people assume the instructions were coming from a human, and the judges didn't cry foul at what was being said by the robot lawyer, it would be very difficult for them to claim contempt of court after the fact when the curtain was pulled back.
That would be a much better publicity stunt - if the judges didn't know it wasn't a real lawyer (even if they thought it was a bad lawyer) that would be a much bigger feather in this guy's cap than the stunt he planned and announced in advance.
Why headphones? Why not use a bluetooth enabled hearing aid and claim the person is hard of hearing? In fact why would you have to justify a hearing aid? Who the fuck would question it?
Wear an earpiece under some religious headgear and request that they respect your beliefs...use bone induction through the frames of some specs...
If you wanted to feed audio to someone in a courtroom, it would be a piece of piss.
There's a million different ways to pipe audio into someones skull that don't arouse suspicion and that wouldn't be questioned at all.
In fact, why even audio at all? What's preventing a lawyer using an AI on his laptop / tablet in court that is listening in?
I think being told in advance is probably the one and only time a court will ever get the ability to opt in to this with any sort of control over how it pans out and with the knowledge that it is happening, I think these people were being nice letting them know their intentions.
Now that everyone knows that a court won't allow it, the only way to do it is to be clandestine about it...the only situation that can arise now is that a court will be...ahem...caught with their pants down. Because if an AI fails, we'll likely never hear about it, but if it's successful, then a court is going to be made to look really daft.
This post has been deleted by its author
Code that takes a personal MP3 collection and builds a model.
The devil is in detail, can't have singing (you've seen stable diffusion fingers right?), analysis of waveforms will need normalising to a standard sampling frequency etc etc.
Concept easy enough though.
We've had successful (commercially successful, indistinguishable to human judges) music for a quarter of a century. Text-description-to-music might not have been done successfully before – not an area I pay close attention to – but machine composition of music is a solved problem.
Machine composition of music that interests a given audience could be tougher.
"Google builds music-making AI, but won't release it due to copyright
Researchers at Google have trained a new AI model, MusicLM, that can create audio samples based on text descriptions, according to a research paper on arXiv.
The first sample, for example, is generated with the prompt: "The main soundtrack of an arcade game. It is fast-paced and upbeat, with a catchy electric guitar riff. The music is repetitive and easy to remember, but with unexpected sounds, like cymbal crashes or drum rolls." You can listen to it here."
And because if they released it, YouTube would then probably flag everyone's song ever made as Googles AI copyright. That's how broken the detection system is on YouTube. For years my Disneyland Paris parade footage was being flagged by some Chinese account claiming copyright on the Disney sound tracks. Only recently have all those claims been removed.