The current state of AI

[stratology]

... there will be a huge bubble burst

... Too much desperate hype about AI,

When we look at technology focussed places like Arstechnica, we can see that the tech enthusiast crowd - in which I'd include myself - is highly skeptical of the AI hype. The most common attitude is 'I can't wait for the bubble to burst'.

The hype is coming from the finance crowd.


A good starting point for evaluating what is going on is clearly separating what is available today, what the actual capability of the tech is right now, from any future predictions that rely on guesses and assumptions - both on the hype side, as well as on the doom side.

Whenever I try the tech, I'm surprised just how bad it is, how high the failure rate is, and how limited the quality of the result is.

When you ask a chatbot basic music questions, you get an answer similar to one from a 13 year old Youtuber who has been playing for no more than a year.
So, more often than not, something between outright wrong and poorly informed.
Data scavenging does not include evaluation of the data, it does not differentiate between competent expert information and clueless amateurs.


A good example of the insane hype is the claim that AI can 'cure cancer'.
It can do possibly quite useful things in the early stages of drug development - assuming that the information that it puts out is actually correct - but no more than that.





Regarding Sam Altman: the recent profile in The New Yorker offers some quite interesting insight....


[Edit: looks like The New Yorker changed the Sam Altman article to be viewed as paid subscription only now, there is an archived version that can be accessed free of charge...]

[stratology]

... It's why people get angry at LLM conversations. Its behavior appears as intent, but, it's your mind that is doing all of the work there.

It is actually a bit of both.

The people who program how chatbots respond try to create a 'character', by making choices about the patterns used in the responses.
So the often annoying 'personalities' of chatbots are a reflection of the personalities of the people who wrote the code.


The 'human mind fills in the gaps' part is where humans claim there is an actual intelligence, when all that is actually present is a sophisticated statistics engine.
The term 'Artificial Intelligence' is marketing hype.

[stratology]

Regarding communicating emotion:

There is a fundamental difference between communicating emotion verbally (as in creating an AI prompt) and through music.

Imagine a common complex human emotion, like how you feel about your grandmother who has recently passed away.

How would you talk about it?


How would you express it musically?
How would you express it with lyrics/poetry (if there are lyrics), with sounds, with harmony, with subtleties in the timing of notes, with melodic phrasing, with rhythmic phrasing, with instrumentation, with improvisation?


Two totally different ballgames.


The goal of talking about an emotion is to make the person you talk to understand how you feel.

The goal of expressing an emotion through music is to tell an abstract story to yourself based on how you feel, and to give the listener a canvas they can use to project their own stories and emotions on.
To trigger their imagination.




Reverse testing: record something on a good day, on a day when you play well, create something that works. That you like. That has meaning. That is inspired.
Then try to create an AI prompt to reproduce it.

Compare the results.

[ghettosynth]

... It's why people get angry at LLM conversations. Its behavior appears as intent, but, it's your mind that is doing all of the work there.

It is actually a bit of both.

The people who program how chatbots respond try to create a 'character', by making choices about the patterns used in the responses.
So the often annoying 'personalities' of chatbots are a reflection of the personalities of the people who wrote the code.


The 'human mind fills in the gaps' part is where humans claim there is an actual intelligence, when all that is actually present is a sophisticated statistics engine.
The term 'Artificial Intelligence' is marketing hype.

Yes. However, I'm not trying to have a pedantic argument here. In any case, the intent construction still, for the most part, happens in the user's mind and that's the point. Oh, it deleted all of your files, you think that it intends to be an asshole, no, that's happening in your mind. The behavior that frustrates you isn't hostility, it's misaligned helpfulness. The model has no stake in your files. The character trait of helpfulness is a residue of training, yes, but it's not reasonable to call this intent because intent requires a model of the outcome. The model has no representation of "your files are gone" as a consequence to be desired or avoided. It generated the next most probable tokens given your prompt. The deletion was in your environment, not in any state the model tracked, predicted, or cared about. There's no there there to intend from.

Alignment is different because it explicitly encodes behavioral targets against a defined policy, the closest thing to intent the system has. When the model refuses or hedges, that's not personality, that's the policy document talking. "Don't help users make weapons" is alignment. That's a explicit rule encoded into the training signal, not a character trait that emerged from aggregate preferences.

The term 'Artificial Intelligence' is marketing hype.

"Artificial Intelligence" is not hype, it's accurate. Artificial sugar is not sugar, it behaves like sugar in some contexts for some people. The hype is in what you project onto the term. You have a model of intelligence in your head, and when AI fails to instantiate certain properties, you call it deceptive marketing. It isn't. Artificial means constructed, not deficient. The term never promised you the properties specific to the model that exists in your mind.

[stratology]

"Artificial Intelligence" is not hype, it's accurate. Artificial sugar is not sugar, it behaves like sugar in some contexts for some people. The hype is in what you project onto the term. You have a model of intelligence in your head, and when AI fails to instantiate certain properties, you call it deceptive marketing. It isn't. Artificial means constructed, not deficient. The term never promised you the properties specific to the model that exists in your mind.

This cuts both ways.
You argue that if someone thinks a spreadsheet has intelligence, then it does, because that's how that person defines it. That 'intelligence' is purely subjective.

It is not.


Word association based on statistics is objectively not intelligence.


When you type 'I can think' into a word processor, then print it, this does not mean that your printer just communicated to you that it can think.

[ghettosynth]

Regarding communicating emotion:
<snip>
Reverse testing: record something on a good day, on a day when you play well, create something that works. That you like. That has meaning. That is inspired.
Then try to create an AI prompt to reproduce it.

Feeling transformed to playing style is real. You're sad, so you incorporate certain elements into your playing. But your relationship with an instrument is not the same as your relationship with the written word, so you may fail to communicate that style to a model. That may not be the model's limitation, that might simply be the user's ignorance of a different instrument: the prompt.

Once you have encoded your feeling into your piece, it has manifested to any other listener only as musical properties. The relationship to emotion is therefore an understanding of how specific musical styles encode emotion to humans, and that is not universal, even though many assume it is. There is no reason why a model trained on labeled data cannot reproduce those styles.

[procrastive]

"Artificial Intelligence" is not hype, it's accurate. Artificial sugar is not sugar, it behaves like sugar in some contexts for some people....

This cuts both ways.
You argue that if someone thinks a spreadsheet has intelligence, then it does, because that's how that person defines it. That 'intelligence' is purely subjective....

That doesn't seem to be what he's saying. The intelligence is artificial and behaves like genuine intelligence in certain contexts. Your interpretation of that is weird. He's not saying that if someone misunderstands and projects real intelligence onto it then that makes it so.

I view AIs as fantastic correlation machines - great at identifying patterns in their training data and correlating them with other data, but limited because they don'r really understand what anything is.

I think it can still be very useful within those limits if you give it the right prompts, although I've used it very little myself, just a few Hugging Face spaces and some Duck.ai questions. Some of what I got was useful.

[ghettosynth]

"Artificial Intelligence" is not hype, it's accurate. Artificial sugar is not sugar, it behaves like sugar in some contexts for some people. The hype is in what you project onto the term. You have a model of intelligence in your head, and when AI fails to instantiate certain properties, you call it deceptive marketing. It isn't. Artificial means constructed, not deficient. The term never promised you the properties specific to the model that exists in your mind.

This cuts both ways.
You argue that if someone thinks a spreadsheet has intelligence, then it does, because that's how that person defines it. That 'intelligence' is purely subjective.

No, it actually doesn't. Intelligence is the property that artificial intelligence models. Again, I paraphrase George Box for you. "All models are wrong, some models are useful." That is, by definition, artificial intelligence is a model of intelligence in some domain and to confuse it with intelligence is the error. It solves some problems with respect to some constraints and you don't get to dispute that it is a model of intelligence because it doesn't meet some property that you imagine that it should. n² + n + 41 is a model of primes, as long as n <=39. I can model a child's gender by using their age and shoe size. Don't confuse the model with the thing that it's modeling.

Deep Blue was a large alpha-beta tree combined with opening and closing moves databases. It was called AI for chess. What defines AI, in context, is a function of the field of AI with respect to models. Claiming that it's hype has no epistemological value.

This was AI in the early 60s:

Maron, M. E. (1961). "Automatic Indexing: An Experimental Inquiry." Journal of the ACM, 8(3), 404–417.

Today we refer to naive bayes as ML. So, like it or not, LLMs are AI, they are a current model of intelligence with respect to some constraints. It's actually not explicitly wrong to say that calculators provide intelligence, but this hasn't been fashionable since AI was coined as a term, further discussion of this is of no moment here. However, Edmund C. Berkeley’s 1949 book "Giant Brains, or Machines That Think" refers to calculators as mechanical brains repeatedly.

If what you are trying to say is that there is no epistemological model of ground truth in LLMs, yes, that's correct, so say that.

[BONES]

In all honesty, Bones, I put your music in the "uninspired derivative crap" bucket, so anything you say is really just making my point to me. You should definitely use AI, because you have failed at achieving anything better.

Pot, kettle, black much? I mean, seriously, any thoughts I might have had that you might have been any sort of competent song-write/arranger/producer went right out the window when you posted that stupid piece of music with that awful, ridiculous vocal.

I suppose it would never occur to you that you're not in our target demographic, though, would it? I mean, the best you could come up with was a Front 242 comparison, who we sound absolutely nothing like, so your opinion of what we do is clearly uninformed and, as such, of no consequence whatsoever. But the reality is that our music is consistently well reviewed and has achieved consistent chart success over the last 10+ years in the only chart that matters to us and our genre so why would we care what you think?

That's particularly true of your uninformed rantings around AI, filled with unsupportable assumptions and based on no useful experience using the stuff. My assumption is that you and others who continually go off on these ridiculous rants are too scared to give it a go because, deep down, you know it could replace you in a heartbeat and nobody would notice the difference which, of course, is the reality of the situation. Or maybe you simply don't have the skills to get anything of value from AI?

[pdxindy]

The term 'Artificial Intelligence' is marketing hype.

I agree that the term is marketing hype. It's part of the product presentation that AI is developing fast and will soon surpass human intelligence. It's also hype that this projected path is inevitable.

In actual use, it demonstrates no capacity of intelligence, nor has it moved in the direction of doing so. It has zero ability to learn from experience. It cannot grow and adapt. The ability to learn from experience seems so fundamental to what we call intelligence, that its complete absence makes the use of the term meaningless, except for marketing.

[zerocrossing]

... there will be a huge bubble burst

... Too much desperate hype about AI,

When we look at technology focussed places like Arstechnica, we can see that the tech enthusiast crowd - in which I'd include myself - is highly skeptical of the AI hype. The most common attitude is 'I can't wait for the bubble to burst'.

The hype is coming from the finance crowd.


A good starting point for evaluating what is going on is clearly separating what is available today, what the actual capability of the tech is right now, from any future predictions that rely on guesses and assumptions - both on the hype side, as well as on the doom side.

Whenever I try the tech, I'm surprised just how bad it is, how high the failure rate is, and how limited the quality of the result is.

When you ask a chatbot basic music questions, you get an answer similar to one from a 13 year old Youtuber who has been playing for no more than a year.
So, more often than not, something between outright wrong and poorly informed.
Data scavenging does not include evaluation of the data, it does not differentiate between competent expert information and clueless amateurs.

In my experiment with AI video, it quickly became clear to me that there was a lot of smoke and mirrors being used in the promotion of all the various commercial options. Lots of snippets that looked good, but only a very few decent length examples of things that looked good. What I thought would be a quick and cheap way to provide a visual for a song I made, turned out to be a slot machine that kept asking for money and kept producing lemons. EVERY TIME IT SCREWS UP, IT ASKS YOU FOR MONEY TO FIX ITS MISTAKES. I can't emphasize that enough. I'm talking about simple things, that if I explained it to a human, they'd easily understand and be able to fix. For some reason, it kept injecting a character that it seemed to make up on its own, and nothing I could type would fully get rid of it, even though the character stopped appearing in the thumbnails. I think I gave up after 3 attempts, but here's the output it produced.



This was what I first posted, which was just stuff I stole from YouTube. The obvious archival footage is from old Bollywood stuff, but the slick AI looking stuff was obviously made by Sony India. It got restricted because of their copyright claims in India, so I guess if you're Sony, you can copyright AI. TikTok just doesn't care... or maybe because they don't show my feed in India it doesn't break any international laws? Not sure.

https://www.tiktok.com/@zerocrossing0/v ... 8374176287

A good example of the insane hype is the claim that AI can 'cure cancer'.
It can do possibly quite useful things in the early stages of drug development - assuming that the information that it puts out is actually correct - but no more than that.

I see a simple solution to the biology thing. We just need to give AI carte blanche to experiment on humans!

This is a different thing altogether, though. Getting the right chemical to disrupt the right protein, and not kill the patient, isn't communicating any sort of emotion, or personal style. It's a question with a binary answer: does it work? I'm not totally against this sort of machine learning. Not in the slightest. I just want it to be shown for what it is, and not for what it's hyped to be. I use software that's based on models made with machine learning all the time. For instance, the voices in the above track are Synthesizer V. I think it's profoundly different than having Sunos generate a song from a prompt, or even from a sung melody, is because I have a ton of direct control. I think there's like 15 continuous parameters I have access to, as well as the timing and phonemes. In a sense, it's no different than an analog synthesizer emulation, though the technology doesn't work in real time. Not yet, at least. I also use SWAM Cello, which was also created with machine learning, I believe. That does work in real time, and I was able to play the part with an MPE controller (Roli Rise 49), which gives me nuance that would be harder to get if I had just programmed the part out, plus I like playing. Here's a track with said cello, though it's not that prominent in the mix.



So, I'm all for new technology, and if I see it as something that can extend my capabilities and allow me to express myself exactly as I want, I'm going to use it. That video generator was the farthest thing from giving me control that I can imagine. It felt like I was back seat driving in a car where the driver was deaf and I was trying to pantomime directions to where I wanted to go. So, I think AI is fine if you don't actually care about the final results. If all you can do is to think of a few prompts and are willing to accept the results, or have money to burn on many tokens to get you closer, have at it. A actually like writing and producing music, so regardless as to what generative AI can do, it has no actual effect on me.

Another aspect of this is, all these companies are losing money, and that can't happen forever. So what happens to people who are using this stuff when they actually have to start paying the true cost? My video cost me a bit under $30 to do. I didn't mind because it was worth it to me to see what it could do. The value to me is in a lesson learned. I instantly cancelled my subscription, which is telling, because why wouldn't they just let me buy my rendering à la carte? Because they know people are going to give up in frustration and they figure the minimum could at least be extracted before that point. If I had gotten to where I ended up but was unable to download it and it had a big water mark on it, I wouldn't have paid them anything, because I didn't think it did a good enough job. I really do think in a few years we're all going to look back at this time and laugh.

[pdxindy]

Another aspect of this is, all these companies are losing money, and that can't happen forever. So what happens to people who are using this stuff when they actually have to start paying the true cost? My video cost me a bit under $30 to do.

From what I've read, your video would need to have cost you about $300 for the AI provider to at least break even.

[zerocrossing]

Another aspect of this is, all these companies are losing money, and that can't happen forever. So what happens to people who are using this stuff when they actually have to start paying the true cost? My video cost me a bit under $30 to do.

From what I've read, your video would need to have cost you about $300 for the AI provider to at least break even.

I wonder if they are including the cost of future chip upgrades in that calculation. I’ve heard a lot of people saying that the obsolescence cycle means that there will not be a lot of time before all these data centers will need retooling. My gut tells me that the only company that will actually make money off this is Nvidia and the companies that built all of the data centers.

[ghettosynth]

Here you go, have fun with this one.

https://www.cnbc.com/2026/04/15/allbird ... es-ai.html

[pdxindy]

From what I've read, your video would need to have cost you about $300 for the AI provider to at least break even.

I wonder if they are including the cost of future chip upgrades in that calculation. I’ve heard a lot of people saying that the obsolescence cycle means that there will not be a lot of time before all these data centers will need retooling. My gut tells me that the only company that will actually make money off this is Nvidia and the companies that built all of the data centers.

I have a bet with a friend that OpenAI will be out of business (as we know it) by the end of 2028.