Coding, Data Science, A.I. catch-All

[gregh1]

Interesting copy-write ruling.

" This marks the first time that the courts have given credence to AI companies’ claim that fair use doctrine can absolve AI companies from fault when they use copyrighted materials to train large language models (LLMs)."

Tl;dr one of AI companies allowed itself to train on pirated data - novels... trying to build a library of all books. And judge seems okay with this. As AGI gets better, without guardrails, it will be able to sniff out pirated data... I can envision a future where i ask some "media AI" to just find me some content - a book or tv series or whatever and it will generate something pretty dang close, with fees going to the AI company instead of the original author.

A federal judge sides with Anthropic in lawsuit over training AI on books without authors' permission | TechCrunch

The ruling isn't a guarantee for how similar cases will proceed, but it lays the foundations for a precedent that would side with tech companies over creatives.

techcrunch.com

 

[gregh1]

They are mostly a black box. Even the companies that created them don't know exactly how they work. They'll openly admit it.

They do know why it gets slightly different answers. It's actually by design. When llms are creating an answer to your query, they're actually creating a lot of answers. And then it statistically ranks the answers it thinks are correct so maybe the first answer is 95% chance of being correct and second answer is a 92% chance and so on. Then the designers of the model will decide how often they want to give the best answer, second best answer, etc.

If you customize your own, you can actually set it to give the best answer every time which occasionally is completely wrong because the model is wrong. Or you can decide to have the model give the second and third and so on best answers some percentage of the time.

Interesting!

 

[nrsair]

Every once in awhile, I will give ChatGPT an offbeat prompt. Today, I tasked it with "create something unexpected"

Here is what it wrote:

A Conversation Between a Forgotten Sock and a Microwave


Sock:
"Do you ever feel abandoned?
Like one moment you’re part of a pair,
And the next—
You’re stuck under a bed, collecting dust bunnies and existential dread?"


Microwave:
"Try being expected to reheat existence
One soggy burrito at a time.
No one thanks me when I radiate time itself.
They just slam the door and curse when cheese explodes."


Sock:
"At least you get attention.
I was worn on a rainy Tuesday in 2017,
Dropped during laundry, and left to contemplate mortality
beneath a radiator."


Microwave:
"...I envy your stillness."


Sock:
"...I envy your heat."

I’ve asked it to write lyrics to a song about a given subject in the style of a well-known artist. For example, I asked it to write a Drake diss in the style of Morrisey, back when the Kendrick/Drake feud was blowing up. It was shockingly good.

Edit: asked it to do it again for giggles- this is different but still pretty good.

“Aubrey of the Empty Mirror”
(a lament in six verses)

Verse 1
Oh, the boy who weeps in velvet halls
Counting streams like sacred psalms
But fame is fickle, truth is tall
And no one’s moved by your false calm

Verse 2
You boast of woes in marble pools
With rented grief and diamond rules
But sadness sung from leather seats
Is sorrow worn by kings as cheats

Chorus
And oh, Aubrey—your tears are gold
But your soul is bankrupt, bleak, and cold
I’d offer a hand, were it not so clear
You'd only use it to fix your veneer

Verse 3
You trade in hearts like currency
And call it art, or therapy
But I’ve seen mannequins with more depth
Than your verses gasping for their breath

Bridge
Yes, I know your fans will cry
“He’s misunderstood, just shy”
But shyness doesn’t rent a jet
To sigh alone in private debt

Verse 4
A gilded owl, a champagne sage
A child who mistakes clout for age
The real ones left, they know your name—
A chorus boy in Caesar’s game

Final Chorus
So sing your ache, your faux malaise
Your scented pain in greyscale haze
But when they carve your marble bust
It will read: “Here lies a man no soul could trust.”

 

[superrific]

I’ve asked it to write lyrics to a song about a given subject in the style of a well-known artist. For example, I asked it to write a Drake diss in the style of Morrisey, back when the Kendrick/Drake feud was blowing up. It was shockingly good.

I've seen that also. I initially thought it was shockingly good, but now I'm not so sure. Who is the easiest person in Hollywood to impersonate? Al Pacino, right? Everyone does a Pacino. And that's because he's so distinctive that the impersonator doesn't have to be good to create believability. He can just punctuate sentences with "hu-ah" and randomly shout things etc. and it's so recognizable that it's successul.

I wonder if the AI is basically doing just that. It finds the parts of the impression that are most visible and leans on them, making the impression recognizable but kitschy. Sort of like when Dali became a parody of himself later in life.

For instance, I've found that it is good at impersonating TS Eliot. He's incredibly easy to parody because his style is so unique and recognizable. It is dreadfully awful at Plath, even though she's not all that difficult to parody either. It's pretty good at Borges (if you disregard the substance, which is an unfair request), and even better at Faulkner, but terrible at Pynchon.

 

[superrific]

I asked it to write a short story about a used car dealership outside of Jackson, MS, in the style of Faulkner. This is what it produced. I can't really tell if it's good Faulkner, but it's obviously Faulkner because the AI can take Faulkner's signature manipulations of time and run with it.

They called it Lucky’s, though no one remembered a Lucky, not even Ed Baldridge who ran the place and had since ’86 when he bought it from a man who may have been named Leland or Raymond but who vanished sometime between Reagan and Bush Sr., leaving behind a ledger, a pair of ostrich-skin boots, and a Thunderbird with a roach clip still dangling from the rearview. The dealership sat off the old highway, not the new bypass, in that no-man’s strip between Ridgeland and the woods that once held the prison work camp where Jesse Polk cracked limestone until his lungs collapsed one August and he died on a Tuesday, same day his granddaughter was born in Yazoo City and named Jessalyn, who twenty years later would buy a teal Elantra from Lucky’s that would throw a rod outside Meridian the day her boyfriend was supposed to propose.

Ed’s office had a mounted catfish and a Polaroid of a snowstorm that wasn’t from Mississippi but Ohio, from his brief time with a woman named Carleen who believed in forgiveness and time travel and who left one morning claiming both had run out. The gravel lot still bore tire marks from the night in ’94 when three teenage boys joyrode a Buick straight through the chain-link fence and into the billboard, the one that once advertised carpet but now just said JESUS IS LORD in faded red block letters.

Cars came and went, their titles scribbled in spiral notebooks, each one a history of something—divorces, deaths, salvation—while Ed watched them all, sat in the sun and let the dust gather, waiting maybe for Lucky, or at least for someone who’d remember the name.

 

[TarSpiel]

I’ve asked it to write lyrics to a song about a given subject in the style of a well-known artist. For example, I asked it to write a Drake diss in the style of Morrisey, back when the Kendrick/Drake feud was blowing up. It was shockingly good.

Edit: asked it to do it again for giggles- this is different but still pretty good.

“Aubrey of the Empty Mirror”
(a lament in six verses)

Verse 1
Oh, the boy who weeps in velvet halls
Counting streams like sacred psalms
But fame is fickle, truth is tall
And no one’s moved by your false calm

Verse 2
You boast of woes in marble pools
With rented grief and diamond rules
But sadness sung from leather seats
Is sorrow worn by kings as cheats

Chorus
And oh, Aubrey—your tears are gold
But your soul is bankrupt, bleak, and cold
I’d offer a hand, were it not so clear
You'd only use it to fix your veneer

Verse 3
You trade in hearts like currency
And call it art, or therapy
But I’ve seen mannequins with more depth
Than your verses gasping for their breath

Bridge
Yes, I know your fans will cry
“He’s misunderstood, just shy”
But shyness doesn’t rent a jet
To sigh alone in private debt

Verse 4
A gilded owl, a champagne sage
A child who mistakes clout for age
The real ones left, they know your name—
A chorus boy in Caesar’s game

Final Chorus
So sing your ache, your faux malaise
Your scented pain in greyscale haze
But when they carve your marble bust
It will read: “Here lies a man no soul could trust.”

That first one is great

 

[ChapelHillSooner]

So I was reading a reddit thread where someone pointed out that ChatGPT could not answer simple questions like how many words were in its previous response or how many s’s are in Mississippi. The theory was counting requires some logic that goes beyond what they were designed for. This thread was about 2 years old.

Anyway so I tried both and it was able to answer both correctly.

That makes me wonder, did they train it specifically to be able to answer these questions (or even hardcode the logic needed to do so) or did it gain this ability simply from increased complexity of the models?

 

[superrific]

So I was reading a reddit thread where someone pointed out that ChatGPT could not answer simple questions like how many words were in its previous response or how many s’s are in Mississippi. The theory was counting requires some logic that goes beyond what they were designed for. This thread was about 2 years old.

Anyway so I tried both and it was able to answer both correctly.

That makes me wonder, did they train it specifically to be able to answer these questions (or even hardcode the logic needed to do so) or did it gain this ability simply from increased complexity of the models?

They have added reasoning modules. I don't really know how those work, but they are probably responsible for that shift (if it was real). They have also tweaked the conversation processing algorithms.

I doubt they trained it specifically for that nor hard coded it in. BTW it will answer questions about itself.

 

[gtyellowjacket]

So I was reading a reddit thread where someone pointed out that ChatGPT could not answer simple questions like how many words were in its previous response or how many s’s are in Mississippi. The theory was counting requires some logic that goes beyond what they were designed for. This thread was about 2 years old.

Anyway so I tried both and it was able to answer both correctly.

That makes me wonder, did they train it specifically to be able to answer these questions (or even hardcode the logic needed to do so) or did it gain this ability simply from increased complexity of the models?

It's possible but unlikely. I guess you could test it with different questions to see if just those questions are hard coded. The models are updated three to four times a year and get better each time.

Maybe the early models would have trouble with those types of questions, especially the second one, because they stored less history, but the new models would likely find those questions pretty trivial.

 

[gregh1]

When downloading a model to self-host (from some vendor or hub), what's the actual file type or content? I think of most models being developed via Python. But you aren't downloading .py files, right?
Some model developed with tensorflow or pytorch libraries, what is it you'd actually download if you wanted something locally for having something for predictions requested/responded entirely locally? Is it just some type of serialized pickle file?

I guess part of the question - is the content just a complex data structure, arrays of arrays? Or is the content source-code + data?

 

[gtyellowjacket]

When downloading a model to self-host (from some vendor or hub), what's the actual file type or content? I think of most models being developed via Python. But you aren't downloading .py files, right?
Some model developed with tensorflow or pytorch libraries, what is it you'd actually download if you wanted something locally for having something for predictions requested/responded entirely locally? Is it just some type of serialized pickle file?

I guess part of the question - is the content just a complex data structure, arrays of arrays? Or is the content source-code + data?

The models are written in either python or C++. You are likely going to want to use python to actually run the models.

To run it locally, you will need a GPU. You can run it with a cpu, but the answers will take 10 minutes. If you don't have a GPU, there are free cloud solutions you can use.

If I were you, I would ask Claude or Gemini for step by step instructions on what you want to do. It will help you select the right model for your hardware and your task. There are plenty of open source models to play with like LLAMA and some Grok models that will be free for you.

The other option is YouTube. Find a video that will step you through. It will take you longer but you will learn more on how to do this stuff and how they work. That's how I learned it

Let me know if I can help.

 

[Geisterfahrer]

Musk is demonstrating a genuine risk of AI — he wants to rewrite human knowledge to match his own beliefs. Once the AI is “trained” by creators who warp information to their own world view, the people who rely on that AI in the future are captives to all the errors, misinformation and disinformation upon which the platform was built. Meanwhile, Musk and other AI evangelists are training generations to believe that AI has all available human knowledge and should be relied on over humans for all information (and eventually decisions). Meanwhile many of the AI creators are openly seeking to control and distort the AI (while others are advocating information in from all sources without filters for credibility, which has its own distorting impact).

If you’ve trained on “the entire corpus of human knowledge” then how would you know what’s missing? By definition it’s the entire corpus, i.e. nothing is missing.

 

[Geisterfahrer]

When downloading a model to self-host (from some vendor or hub), what's the actual file type or content? I think of most models being developed via Python. But you aren't downloading .py files, right?
Some model developed with tensorflow or pytorch libraries, what is it you'd actually download if you wanted something locally for having something for predictions requested/responded entirely locally? Is it just some type of serialized pickle file?

I guess part of the question - is the content just a complex data structure, arrays of arrays? Or is the content source-code + data?

Good question, better ask the open source AI

 

[JCTarheel82]

Good question, better ask the open source AI

A couple months ago, I was assigned some Microsoft Copilot training for work so I prompted Copilot to summarize Microsoft's training for Copilot.



But seriously, I did the training as well because the system tracks completion.

 

[gregh1]

Good question, better ask the open source AI

Of course:
you're not downloading raw .py code, but rather the trained model's weights and architecture, typically in a serialized binary format. Here's a breakdown:

Common Model File Types (Framework-specific)
PyTorch
File type(s): .pt or .pth

Contents: A serialized dictionary containing the model's weights (and optionally the architecture).

Serialization method: Usually Python's pickle, wrapped via torch.save().

Note: You often need the matching model class definition in Python to load the weights unless the model was saved via TorchScript (see below).

TensorFlow / Keras
File type(s):

SavedModel format (directory with saved_model.pb, variables/, etc.)

HDF5 format: .h5 file (especially for Keras models)

Contents: Model architecture, weights, training configuration, optimizer state (optional)

Serialization method: TensorFlow-specific serialization, not raw Python pickle.

ONNX (Open Neural Network Exchange)
File type(s): .onnx

Contents: Framework-agnostic model graph and weights.

Serialization method: Protocol Buffers.

Use case: Portability across different runtimes, like ONNX Runtime, TensorRT, etc.

What’s Actually Inside the File?
It’s not Python code, but a binary representation of:

The model's architecture (optional or required depending on framework).

The learned parameters (weights, biases).

Optional: training config, optimizer state, etc.

In PyTorch, these are typically Python objects serialized using pickle, so there is a potential security risk in unpickling arbitrary models from untrusted sources.

What Else Do You Need to Run It Locally?
To use the model for prediction entirely locally, you typically need:

The model file(s) (e.g. .pt, .onnx, or TensorFlow directory)

The inference runtime:

PyTorch: torch.load() and the model's class

TensorFlow: tf.saved_model.load() or keras.models.load_model()

ONNX: ONNX Runtime or compatible framework

Any preprocessing/postprocessing logic in Py
thon (often custom code that goes along with the model)

 

[gregh1]

The models are written in either python or C++. You are likely going to want to use python to actually run the models.

To run it locally, you will need a GPU. You can run it with a cpu, but the answers will take 10 minutes. If you don't have a GPU, there are free cloud solutions you can use.

If I were you, I would ask Claude or Gemini for step by step instructions on what you want to do. It will help you select the right model for your hardware and your task. There are plenty of open source models to play with like LLAMA and some Grok models that will be free for you.

The other option is YouTube. Find a video that will step you through. It will take you longer but you will learn more on how to do this stuff and how they work. That's how I learned it

Let me know if I can help.

Thanks I'm just trying to mentally unwind some stuff, sometimes I get stuck on something low-level B4 being able to grasp the high-level.

 

[gtyellowjacket]

A couple months ago, I was assigned some Microsoft Copilot training for work so I prompted Copilot to summarize Microsoft's training for Copilot.



But seriously, I did the training as well because the system tracks completion.

What was the gist of the training? It seems like it would be pretty close to "you know how to Google right? Do that but ask slightly more complex questions."

 

[ChapelHillSooner]

Thanks I'm just trying to mentally unwind some stuff, sometimes I get stuck on something low-level B4 being able to grasp the high-level.

I do that too.

 

[ChapelHillSooner]

Anyone think in the very near future most of our interactions on the internet will be with AI, whether we know and want it or not?

It has already happened with images and videos where the majority of them are now AI generated. We are also seeing more and more AI responses in conversational social media. At first I thought it was users feeding things into a LLM but I think it might be getting to the point that that is more and more automated.

And it is possible I am way behind on noticing this.

 

[1moretimeagain]