Not ready for prime time

[hwlentz]

I asked Apple’s new AI image creator for a trombone emoji…

IMG_4195.jpeg

[hyperbolica]

Yeah, I asked for a trombone in the grass once, and got this

[AtomicClock]

Reminds me of an old M*A*S*H episode. Charles gets his French horn back from the shop.

[KWL]

The Mumford & Sons European tour poster contains an interesting trombone.

[JTeagarden]

I asked Apple’s new AI image creator for a trombone emoji…

This is the latest BAC design.

[Posaunus]

The Mumford & Sons European tour poster contains an interesting trombone.

Is this a result of AI art? Hard to believe that a human being imagined this.

[harrisonreed]

Of course it is. It's still relatively easy to differentiate between real and fake. Probably won't be for much longer.

AI video upscaling is past the point of being able to tell, now, though. I have a software I use to make old shows that are 480P upscaled to 4K. It adds in details like material texture on clothes, individual strands of hair, the inner structure of the iris... It looks real.

I'm sure art will be there soon. Nvidia just announced AGI today. Not sure I believe that one, but.... I guess we'll see.

[Matt K]

Honestly, yeah the AI art is bad but I've seen some hilariously bad interpretations of trombones way before AI was a thing

[harrisonreed]

Yeah, was it Kimball's site that had all those old paintings? Nobody knew how the heck a trombone worked back then. They definitely still don't now.

[AtomicClock]

It adds in details like material texture on clothes, individual strands of hair, the inner structure of the iris...

If that's the case, then upscaling is the wrong word.

[harrisonreed]

It adds in details like material texture on clothes, individual strands of hair, the inner structure of the iris...

If that's the case, then upscaling is the wrong word.

It's called an "AI upscale". So the video goes from 480 to 4K, but each frame is also run through a model to infer detail (or add frames for a frame rate change). But it's not heavy handed or "fake". It's as if there was a 4K camera next to the original video cameras in the 90s or whenever the show was filmed, and they just recently found the 4K video in a dusty box.

Contrast with an upscale that just converts the blocky version of the old 480 video to be just as blocky but in 4K.

The AI version is absolutely nuts. And it doesn't really feel as anti-human as the generative diffusion type of AI making those weird trombone paintings. This is breathing life back into works made by humans that was confined by equipment that sucked. I think they trained the models by literally filming things on crap equipment right next to the best equipment, making two versions of the same video with drastically different quality levels, and having the model figure out how to make the crappy video look like the 4K video.

[robcat2075]

I asked Apple’s new AI image creator for a trombone emoji…

OT1H I'm surprised it is still that bad.

OTOH my sister can't discern between a trumpet or a trombone so maybe however is training the AI didn't know what a trombone was either.

[hwlentz]

OT1H I'm surprised it is still that bad.

OTOH my sister can't discern between a trumpet or a trombone so maybe however is training the AI didn't know what a trombone was either.

Played a euphonium solo in church a coupe of years ago. Woman who sits behind me (very intelligent, wonderful voice, former choir member, daughter is a semi-pro vocalist in Cincy), said “I just love hearing the French horn.” (FWIW, bulletin listed me as Bill Lentz, Euphonium.)

[Mamaposaune]

Not trombone related, but made me wonder where AI collects their "facts" from. I belong to several turtle groups on Facebook, this morning one showed up from a turtle group that I'm not a member of, AI showed up at the top.
The question was: "What do juvenile snapping turtles eat?"
Like most fb groups, anyone can jump in with an answer, whether or not they are knowledgeable. Additionally, the question did not specify whether it was for Snappers in the wild or captivity. If someone were to do a search on what young turtles eat, will all the answers submitted show up?

[Matt K]

The short answer is that language models are typically trained on significant portions of existing material ranging from literature, scientific papers to other things like news, code, and social media, sometimes of dubious ethical origins (https://www.arl.org/blog/training-gener ... -fair-use/, https://www.wired.com/story/new-documen ... i-lawsuit/).

There are weights put in place that heavily emphasize "proper" grammar and syntactic structure, which is why things like the em dash and certain words and phrases are now associated with LLM usage. (Much to my chagrin, em dashes and rocketship emojis constituted a non-insignificant part of my pre ChatGPT communication ).

Similarly, companies typically put "super prompts" and have routing so that if you ask a question about things like sea turtles, it can put the thumb on the scale of what gets returned, in layman's terms. I know very little about sea turtles, but if, for example, they can be one of two colors - one of the prompts that the answer might get routed through might say, "Any answer about the color of a sea turtle may include only blue and green." It doesn't guarantee the "correct" answer, or that the LLM won't return a hallucination (though the router can refuse answers and force another generation).

At the end of the day, the LLMs are token-based. It basically is fed something between a character and a whole word, and sometimes a concept, and it returns the "most likely next token." What exactly the "next most likely token" is depends on what inputs the model has been trained on and what weights and routing are applied to it.

There is often a bias to say SOMETHING, which is where "hallucinations" come from. Without proper routing, most LLMs will spit out something that is absolutely grammatically perfect and total nonsense.

[robcat2075]

... If someone were to do a search on what young turtles eat, will all the answers submitted show up?

Sort of.

The Googlebot practice is to present a "overview" of what it knows about the topic (which may possibly be correct) but to also accompany that with links to (hopefully) non-AI-generated articles you may peruse to "evaluate the quality" of the overview.

There will still be thousands of search results after the few highlighted articles.

Current Google disclaimer

About the source

This overview was generated with the help of AI. It’s supported by info from across the web and Google’s Knowledge Graph, a collection of info about people, places, and things. Generative AI is a work in progress and info quality may vary. For help evaluating content, you can visit the provided links. Learn more about how AI Overviews work and how data helps Google develop AI in Search.

[AtomicClock]

There are weights put in place that heavily emphasize "proper" grammar and syntactic structure, which is why things like the em dash and certain words and phrases are now associated with LLM usage.

Off-topic, I know...
When I grew up, we learned to write with pen & pencil. Or a typewriter if we wanted to get fancy. Em dashes were only known about by people in the printing industry. Sure, computers game everyone access to them. But I don't know where the expectation came from that we should all follow the advanced typesetting rules.
Harrumph.

[robcat2075]

Played a euphonium solo in church a coupe of years ago. Woman who sits behind me (very intelligent, wonderful voice, former choir member, daughter is a semi-pro vocalist in Cincy), said “I just love hearing the French horn.”

That's a scene from an old movie...

Groucho: I just love hearing the French horn.

Countess: I was playing A EUPHONIUM!

Groucho: Yes, I just love hearing the French horn.

[Kbiggs]

My personal gripe(s) about AI:

1. The decisions we’re presented with (papers, translations, graphics, whatever) are still the result of binary decision trees. Yes—no, 1–0, black—white, blue-green. The decision trees are very, very long, and they’re quite sophisticated and complicated, but they’re still decision trees based on logic. Faulty reasoning is a form of reasoning, after all.

2. It’s difficult to think that CG-AI will eventually emulate human thinking and reasoning. Perhaps that’s not the ultimate goal or desire, but I can’t imagine it (which is another way to say that my imagination has boundaries). How can a program copy/emulate the awe-some, creative and fallible phenomenon of human thought?

I have of late, (but wherefore I know not) lost all my mirth, forgone all custom of exercises; and indeed, it goes so heavily with my disposition; that this goodly frame the earth, seems to me a sterile promontory; this most excellent canopy the air, look you, this brave o'er hanging firmament, this majestical roof, fretted with golden fire: why, it appeareth no other thing to me, than a foul and pestilent congregation of vapours. What a piece of work is a man, How noble in reason, how infinite in faculty, In form and moving how express and admirable, In action how like an Angel, In apprehension how like a god, The beauty of the world, The paragon of animals. And yet to me, what is this quintessence of dust? Man delights not me; no, nor Woman neither; though by your smiling you seem to say so.

—Shakespeare, Hamlet

3. How will we know when we’re no longer thinking for ourselves but letting AI (GIGO cr@p) make decisions for us ? Granted, AI has many uses that can help humankind—science models for climate change, medical models for pathogenesis, etc., but consider that we’re already allowing AI to control huge sections of the banking industry and monetary policy.

There’s a lot of AI-generated “art” that is laughable (alongside Medieval and Renaissance pictures of people playing instruments, including the sacbut), think of how many people see some of these images and either don’t know they’re not representative of an actual trombone, or believe they really are a picture of a trombone?

At what point will humans be deluded by AI?

# # # # #

Rant over… maybe I should stop reading dystopian novels… or watching Black Mirror…

[robcat2075]

2. It’s difficult to think that CG-AI will eventually emulate human thinking and reasoning. Perhaps that’s not the ultimate goal or desire, but I can’t imagine it (which is another way to say that my imagination has boundaries). How can a program copy/emulate the awe-some, creative and fallible phenomenon of human thought?

On that point, it won't need to do that to be economically devastating. Most of what we pay humans to do doesn't require daring creativity, just expert knowledge.

AI will just need to be as accurate and/or as fast as Joe the accountant or Sue the x-ray reader to make them an unnecessary cost. Joe and Sue follow many rules to do their work, which until now required substantial human intelligence to master, but who wants a "creative" reading of their x-ray?

Most of our trombone playing is not creative, we are recreating moments and ideas we've learned from others.

In another thread I asked why certain pieces were on an orchestra's audition list. Answer... because that's what everyone already knows and plays.