Discuss AI Here

[Eudimorphodon]

Things can be taken care of simply, it needs to be a lawful misdemeanor to try and patent/trademark/copyright non-human generated stuff. Period.

Up that to “firing squad” and maybe we’ll be getting somewhere.

 

[Zare]

There's a person who engaged into copyrighting all possible intervals and arrangements in Western music system in USA lately, to show the absurdity of the system.

It has been proven in music litigation, multiple times, that record companies are able to copyright the simplest intervals in the most common scales of our system, one high profile case (Katy Perry) was about three near-by notes on a synth, played one finger style, in a very simple rhythm. If you put an infant in front of a keyboard for an hour, he would hit that pattern.

The song they sued was not even in a same key.

"AI" will just abuse this system in the same manner record companies do with manual labour right now. Produce a ton of generic shit and copyright it. Perform illicit derivative work in a way to just slip over the rules - if you can't copy 4 notes in a succession, copy 3, then make mass variations of it - remodulate, reharmonize into all possible modes, publish/copyright. Tomorrow the author comes, jams over his own songs, comes up with a slight variation of his own stuff that's now fresh enough to be a new song, publishes it...boom, copyright strike from AI.

 

[Svenska]

That the EU has such different laws and regulation from the U.S. (including the examples you've given yourself) demonstrates that people have a very meaningful influence over businesses.

Fair enough, although I still don't see much of a point discussing it here.

If I ask you to create an NN that will never "decide" that a flag is preserved over a POP AF instruction, you can't do that, except to wrap it in traditional code.

Well, I can provide a curated dataset with only that information and train an NN only on it. Given the small space, I can even fully reason about its outputs. It's just that such a network will be completely useless for any questions not related to "POP AF" and flags.

And it's made all the worse by the fact that LLMs deliberately introduce randomness in their outputs.

There are two layers of randomness introduced. The initial state is randomized once ("seed"), and the sampling process introduces additional randomness ("temperature"). Reducing the temperature substantially reduces the "creativity" of the output, making it stay closer to the question at hand.

Your kind of question benefits from sampling with a low temperature. In my experiments, that reduced the tendency to hallucinate - but increased the tendency to form loops. There is a trade-off necessary, and the previous example of asking "the boss" vs "the wife" the same questions is a real-world analogy here.

To my knowledge, ChatGPT 4 is not a single LLM - it is a group of multiple LLMs fine-tuned for different areas. We will likely see more of them, which means better results in common areas and more spectacular failures elsewhere.

That you cannot debug some compiled code given to you by someone else is not surprising, but that's not at all what I'm talking about.

I know. But since your argument doesn't hold in large parts of the real world, I treat it as a theoretical argument. And my counter is that LLMs are fully deterministic (assuming pseudo-randomness, which is true for all systems I looked at), so they are in theory as inspectable as any other code.

In practice, LLMs are a lot less inspectable than traditionally programmed computer systems, but in a world of software-is-a-remote-controlled-and-constantly-updated-service, I don't see them as inspectable either.

That right there says that there's something vastly different from reason going on: if you reason out the rules for addition of numbers they always work.

Learn by example only will not provide good rules to follow. Learning languages show that quite well, as it takes most people unreasonably large amounts of time and effort to get to a point where their output is indistinguishable from native speakers.

To my knowledge, all current LLM training processes work by example only. I see that as a limitation of the training phase, not the inference phase. Also, brains struggle at complex reasoning, so I expect any system modelled after brain activity to struggle in similar ways. If you don't want that, don't use such a system - a logic solver or adder chain is going to be more efficient.

There's a person who engaged into copyrighting all possible intervals and arrangements in Western music system in USA lately, to show the absurdity of the system.

Related question: Do you see this as a problem with AI or as a problem with existing copyright law?

 

[cjs]

I know. But since your argument doesn't hold in large parts of the real world, I treat it as a theoretical argument.....
In practice, LLMs are a lot less inspectable than traditionally programmed computer systems, but in a world of software-is-a-remote-controlled-and-constantly-updated-service, I don't see them as inspectable either.

Ah, there's the issue. You're looking at a minority of the software out there, and doing you're analysis as if that's all of it.

There are vast quantities of software out there, ranging from financial systems to embedded software to safety-critical systems, where the developers do not work in a "move fast and break things" way, and where people are quite careful to reason out exactly what their software is going to do in virtually every situation.

Also, brains struggle at complex reasoning, so I expect any system modelled after brain activity to struggle in similar ways.

If I were seeing LLMs struggle in similar ways, I'd be encouraged. But they're entirely unable to do simple reasoning, so it doesn't really matter how they struggle on complex reasoning.

Related question: Do you see this as a problem with AI or as a problem with existing copyright law?

I wasn't going to bother weighing in on this but since I'm here anyway: it seems to me very clearly a problem with copyright law. You don't need an AI to come up with songs that are the same "attack" on other composers.

 

[Zare]

Related question: Do you see this as a problem with AI or as a problem with existing copyright law?

Existing law problem AI is exuberating.

I wasn't going to bother weighing in on this but since I'm here anyway: it seems to me very clearly a problem with copyright law. You don't need an AI to come up with songs that are the same "attack" on other composers.

It costs pennies for AI in compute power to come up with a variation on a music theme. A professional would charge about 1000x as much, and would take 1000x in time to come up with the end product.

Yes the mass attack is possible without AI, but it isn't feasible. AI makes it economically feasible. Tomorrow the musician I was talking about is going to win over every case at every service, he's going to get his copyrights back. That means the Chinese scammer lost whatever he invested in that endeavor. And because his investment is next to nothing, he can lose 999/1000 of these cases, and still be in profit from the single poor artist that hasn't managed or just didn't bother with the litigation.

The low-effort clone and copyright attack is an AI-era thing.

 

[Kouji Matsui]

In my experiments, that reduced the tendency to hallucinate - but increased the tendency to form loops.

It's a very interesting difference. Since LLM is modeled on humans, I'm imagining that humans also use this kind of thinking process unconsciously.
In other words, when thinking about a certain thing, it seems like it's a matter of whether we think deeply or take into account all kinds of unimaginable possibilities.
(my understanding of LLM is superficial, so I may be off the mark).

 

[Agent Orange]

Hmm. Robotic mission controller R2D3 is faced with a huge problem; i.e. continue on course and crash into the Steve Allen asteroid belt, or otherwise veer over and slam into Skorbic 9's outer ring. Else could be heard from 3rd mate Alfie (bleeds like you and me), who leaned over the engineering bridge railing and could be heard shouting "how about turning this buggy around 180 and just head back home". Back in a minute and now a word from our sponsor . . .

Rather that the tired ole left-right, up-down logic, we now have the new champion decision maker, ta-da, the 'qutrit state'. It's neither here nor there but somewhere in between. Another way out! You can almost her R2D3 screaming "why don't I have that?

 

[hunterjwizzard]

Here's a question:

I'm a writer. All of my writing is set in the same continuity or story-world. I have thousands of documents containing notes, unfinished stories, ideas, world-building, etc. Without having it be uploaded to the internet in any way shape or form, can I plug this into an AI to answer questions about my own writing for me?

Not tone or story ideas. More along the lines of "what is this character's mother's maiden name?". I have all this info stored across thousands of documents. It can be frustrating to find it.

Just a random idea I am curious about exploring.

 

[MrImprovement]

Here's a question:

I'm a writer. All of my writing is set in the same continuity or story-world. I have thousands of documents containing notes, unfinished stories, ideas, world-building, etc. Without having it be uploaded to the internet in any way shape or form, can I plug this into an AI to answer questions about my own writing for me?

Not tone or story ideas. More along the lines of "what is this character's mother's maiden name?". I have all this info stored across thousands of documents. It can be frustrating to find it.

Just a random idea I am curious about exploring.

Yes, I think the feature you want is called RAG ; you can download and run some AI models, some even on CPU (not GPU) but of course they will run slower.

 

[Svenska]

Since LLM is modeled on humans, I'm imagining that humans also use this kind of thinking process unconsciously.

Biological neurons have a lot of inherent randomness (sometimes they fire when they shouldn't, don't fire when they should, have a long cooldown after firing), which is not modelled by technical neurons directly. Use of alcohol (or other drugs) makes them less predictable, so maybe that's an equivalent to "high temperature sampling".

All of my writing is set in the same continuity or story-world. I have thousands of documents containing notes, unfinished stories, ideas, world-building, etc. Without having it be uploaded to the internet in any way shape or form, can I plug this into an AI to answer questions about my own writing for me?

A RAG system provides access to data sources (some data base or your documents) to LLMs. It's an AI-enhanced search engine using your own data.

You could also use an LLM directly, but if you have lots of documents, they won't fit into the context. So you would fine-tune a base model on your documents, actually changing the model by teaching it your data. The result will know your world, but it will still keep its base training; that makes it usable in more creative ways, but less trustworthy for facts (expect the usual hallucination and unreliability issues).

Running LLMs locally is reasonably easy. When run on CPUs, it is usually faster to use only a subset of cores (LLMs are limited by memory bandwidth, not compute speed), and performance is acceptable for medium-sized models. Using GPUs is obviously much faster, but the limiting factor is video memory size (the LLM plus its working memory must fit into VRAM). It is possible to combine CPU and GPU execution, but I've never tried that. When I played with it, the performance characteristics were noticably different between CPU and GPU execution ("time to first output token" was very slow on CPU implementations).

 

[Zare]

Biological neurons have a lot of inherent randomness (sometimes they fire when they shouldn't, don't fire when they should, have a long cooldown after firing)

It is still under debate whether brain contains quantum systems. E.g just because there are counterarguments and criticism of some scientists work (Penrose/Tegmark), doesn't mean it's not a valid hypothesis.
The notion of brain being a fully classical system was always sketchy to me.

Also you've framed perfectly what scientists avoid to say to not lose the funds for further research. "This is random" is a placeholder for "stuff happens when it shouldn't", which is another way of saying, "we don't have enough understanding of the situation"

 

[hunterjwizzard]

A RAG system provides access to data sources (some data base or your documents) to LLMs. It's an AI-enhanced search engine using your own data.

You could also use an LLM directly, but if you have lots of documents, they won't fit into the context. So you would fine-tune a base model on your documents, actually changing the model by teaching it your data. The result will know your world, but it will still keep its base training; that makes it usable in more creative ways, but less trustworthy for facts (expect the usual hallucination and unreliability issues).

Running LLMs locally is reasonably easy. When run on CPUs, it is usually faster to use only a subset of cores (LLMs are limited by memory bandwidth, not compute speed), and performance is acceptable for medium-sized models. Using GPUs is obviously much faster, but the limiting factor is video memory size (the LLM plus its working memory must fit into VRAM). It is possible to combine CPU and GPU execution, but I've never tried that. When I played with it, the performance characteristics were noticably different between CPU and GPU execution ("time to first output token" was very slow on CPU implementations).

Speed is not a huge issue. I'm fine with giving it a few weeks to chew on the data and don't mind if it takes several minutes to answer me. These to me are minor inconveniences so long as the system can run locally and not report anything external. This is why I won't look at any sort of cloud-based solution, my ideas are mine.

What exactly do you mean by "less trustworthy for facts"? I'm primarily looking for something that can find me minor details I've forgotten but are written down somewhere in the myriad of notes.

I do have plenty of computing power at my disposal. I have a dual xeon workstation with 12x 3.4ghz cores, 64gb RAM, and a RTX A4000 graphics card. There's also enough space in the tower I could throw in a few of those things that are like a GPU but not. I forget what they are called but they're for this specific application.

 

[Svenska]

It is still under debate whether brain contains quantum systems.

Maybe, but actually modelling them is unlikely to be necessary. Quantum effects only become relevant at very small scales, and the structures we are working with here are much larger. Modelling electricity using voltage and current is perfectly fine unless we are talking nanoscale semiconductor technology, and one rarely uses that abstraction level to model computer programs.

Knowing that brain cells are inherently unreliable (and knowing that having more cells is an effective mitigation strategy) means that we don't need to model each neuron accurately. According to one of my professors, researchers used logical AND/OR functions on sequential binary data streams to implement technical neurons - because that was much faster than using multiplication/addition. It did work.

Using 32-bit floating-point math (carefully centered around 0.0f) is pure overkill, but still more efficient until 16-bit floating point formats became supported in hardware. With those, much larger models can be run in the same amounts of memory size/bandwidth, for only a relatively small decrease in quality.

"This is random" is a placeholder for "stuff happens when it shouldn't", which is another way of saying, "we don't have enough understanding of the situation"

That's not true in general. For some processes, especially at very small scales, knowing the relevant factors in sufficient detail is too hard or even impossible (related: Heisenberg's uncertainty principle). In most cases, one is better off modelling these as statistical noise rather than

What exactly do you mean by "less trustworthy for facts"? I'm primarily looking for something that can find me minor details I've forgotten but are written down somewhere in the myriad of notes.

Even if you fine-tune a model on your documents, it's still a large language model and has seen tons of other training data. You will need to deal with hallucinations, especially when you ask questions for which your notes do not provide a clear answer. In general, model quality scales a lot with the data quality used during both training and fine-tuning. We currently just throw tons of low-quality trash at our models and hope for the best, which shows.

Verifying any output by grepping through your documents (assuming they are in a sensible format) should be fine, though. Specialized systems will spit out references, raw LLMs using simpler models might just invent some if pressed.

I'm fine with giving it a few weeks to chew on the data and don't mind if it takes several minutes to answer me.

Personally, I have never trained or fine-tuned a model myself, so I don't know how much time or effort goes into doing that. Your system specs are plenty good, I'd expect outputs within seconds rather than minutes.

A while back, I did performance testing of CPU-based LLMs (ggml). On Xeons, best performance was achieved when restricting the tool to either "even" or "odd" cores (leaving half the cores unused). Threadrippers performed best with consecutive blocks of 16 cores (0..15, 16..31, etc). Intels big/little architecture was complete garbage with all cores used; but it did well with only P-cores (again, even or odd cores only). Things may have changed since then.

 

[hunterjwizzard]

Well it looks like the worst case scenario I will be out a few hours of my time figuring out how to set all this up. And as we've established, my time is worthless. I think I'll give it a whirl and see what happens.

 

[Svenska]

Good luck and please report back on how it worked for you!

 

[hunterjwizzard]

Oh don't worry, I'm virtually guaranteed to hit some sort of technological snag and need help I'll keep you all posted. I've been one of the most vocal opponents of AI so this should be interesting.

 

[Zare]

Maybe, but actually modelling them is unlikely to be necessary. Quantum effects only become relevant at very small scales, and the structures we are working with here are much larger

The percieved randomness could have its source in the quantum realm. If a random value in nature occurs we do not know whether it had a wave function collapse behind it.
We're not sure and still debating about the term of the observer - the term is too vaguely anthropo-centric, the observer is anything entangled with the thing observed, a part of the same system.

There are hints of something like this :

How Migrating Birds Use Quantum Effects to Navigate

New research hints at the biophysical underpinnings of their ability to use Earth’s magnetic field lines to find their way to their breeding and wintering grounds

www.scientificamerican.com

I'm sorry to give out a popular article as I don't have time to find all the papers that these hypothesis compile in.
But if there's a chance that birds do use quantum mechanics to navigate, it would mean the Earth's magnetic field is entangled with particles in their brain.

That's not true in general. For some processes, especially at very small scales, knowing the relevant factors in sufficient detail is too hard or even impossible (related: Heisenberg's uncertainty principle). In most cases, one is better off modelling these as statistical noise rather than

Agreed about the modelling. High-energy particle physics does look back into the randomness after a model has been found. In discovery of particles of Standard Model I believe a lot of previous accelerator runs were looked back into, to see whether the newly found out pattern happened then and there.

What I was trying to state is that a biologist might look into a bird's head scan and say, these cryptochrome proteins change their energy state depending on the sunlight exposition, which is a normal behavior in interaction with the environment, and it results in some random chemistry in the eye/brain processes its a part of.

Also I am just talking generally - what you write about using computers to model the brain is completely right of course.

Well it looks like the worst case scenario I will be out a few hours of my time figuring out how to set all this up. And as we've established, my time is worthless. I think I'll give it a whirl and see what happens.

Good luck and please report back on how it worked for you!

Yeah please do, I'm also interested in the case of private document management. For me a discovery of things in the knowledge base would be perfectly fine, compiling and deriving and combining is a secondary but it would be great if an AI could give you some sort of a procedure that was sourced from multiple KB articles that weren't written step-wise.

I also have a box to test this out, so we might as well try it together.

 

[olePigeon]

As someone who doesn't have the mental capacity for programming, I've found it immeasurably useful. I was able to make a website using PHP without knowing a single thing about PHP. And that was with the public beta of ChatGPT. I haven't tried any of the more recent iterations, either from ChatGPT or any of the other AI companies. I really should revisit it because the code it produced was a bit buggy and I wasn't able to coax it into doing everything I wanted. It got stuck and kept making the same incorrect suggestions, so I stopped while I was ahead. I'm happy with what I have so far.

 

[hunterjwizzard]

Yeah please do, I'm also interested in the case of private document management. For me a discovery of things in the knowledge base would be perfectly fine, compiling and deriving and combining is a secondary but it would be great if an AI could give you some sort of a procedure that was sourced from multiple KB articles that weren't written step-wise.

I also have a box to test this out, so we might as well try it together.

There's already tools that do this, though they cost money. That's where I got the idea, we are experimenting with one at work.

I've found a few guides already for doing it on your own. Doesn't look too complex, just downloading and installing some software.