Yeah, I used to do that too. Decided to stop bothering with the quotation marks a couple months ago.
We're not going to spend the rest of our lives putting quotations around words when talking about models. "Think" and "reason" the words we have in English for what is going on. No need to tiptoe around it. Again: models are not humans. They are not the same as us. But those are the words we have in English for what they're doing.
They are, by definition, Markov chains.
Even in your attempt to be pedantic here (in which the universe and everything within it is a Markov chain), no, it's not. The hardware state is Markovian but the linguistic processing is a Nth order autoregressive process; it depends on the N previous states. Also, your argument is akin to saying "a Boeing 747 is just an arrangement of quarks, so don't get hung up on aerodynamics." it entirely ignores the relevant architectural details, and instead substitutes a model that blows up exponentially explodes in size within a small number of states.
If you tried to build a Markov model to do what LLMs do, and could store one probability in every unit of Planck space across every unit of Planck time, it couldn't handle a prompt longer than about 2/3rds of the first sentence to A Tale of Two Cities.
It matters quite a lot if someone sues you. Or if a propagandized public and their leaders get into a war over it.
The US Supreme Court, and supreme courts in general, are not "triggered." You appeal to them to hear your case. They may do so if they feel that your case is somehow relevant to wider issues.
The US Supreme Court might hear this case if they wanted to interpret the constitution in such a way as to set upper limits placed on fines, or wished to issue a judgement on the idea that fraud is still fraud if nobody gets hurt.
Yeah, that's a good example. Also "the actual thinking process as opposed to the correlational hallucinations that constitute all LLM 'thought'."
Two different sides of the same hubris.
A little tutorial if you know a little bit of Python:
https://numpy.org/numpy-tutori...
And a book with more detail if you're more into math:
https://www.deeplearningbook.o...
Everything else is pretty much scaling up and introducing some restrictions on the basic model.
You see it in the way you just used the word "you." Model "collapse" is a misnomer. It's drift, which you see in humans in literally everything they do, and you can demonstrate to yourself just by repeatedly generating random numbers and calculating the mean.
Bullshit. The fundamentals of modern AI models can be understood by anyone with a basic knowledge of algebra, which you should have picked up in junior high. They're piecewise linear approximations and use exactly the same equation as the linear regression you learned in high school or first year university. The more advanced stuff is hacky restrictions on that basic design to tone down the model's flexibility and make it easier to fit.
The reason it's hard to understand is because a) people who have no idea what they're talking about try and handwave their way through it; b) people explaining it want you to think it's really, really sophisticated and hard to understand, c) you're using "understand" in an unrealistic way or d) some combination of the above.
I kind of wonder if the best anti-distillation strategy is, if you detect suspicious traffic from someone (which happens a lot, they monitor for anything that looks like distillation), instead of blocking them, feed them say the output from Llama 3.1 8B or whatnot
As for copyvio, sorry, this is something for the courts, and so far, the courts have not largely found against the trainers, and have instead found, by and large, that they're being compliant. The most notable setback against Anthropic for example was a finding that they couldn't just download books in training dataset off the internet, but that they could perfectly legally just buy surplus books for pennies on the dollar by the palletfull, scan them in, and train on that. That this is perfectly complaint with US copyright law.
I think a lot of you wish that copyright law was a lot more restrictive than it actually is. Which is a REALLY bizarre thing to see on Slashdot of all places, which back in the day was the beating heart of "Data Wants To Be Free!" philosophy.
To be clear, though... I would welcome a compromise modification to copyright law, which is, if you want to train on the public commons, you absolutely may, indeed, train on whatever you want, zero liability, but then you have to give back to the public commons. So maybe your top frontier models are closed, but you have to simultaneously release smaller distilled equivalent versions of it into the public domain (how to define "smaller distilled equivalent versions" is of course something that would require discussion), and release said frontier models to the public domain within e.g. 1 year or whatnot.
* They remain incentivized to keep pushing the frontier, since some people will always pay for the best
* They get permanently out of the worry of any copyvio liability (beyond basic requirements about not verbatim reproducing copyrighted materials in outputs)
* The public gets a constant stream of ever-better models, at no cost.
Sounds like a balance to me.
I think there is room to argue though, that distillation might violate copyright while training on large volumes of large material does not.
Nope. There is an argument about whether training a model on source material violates copyright or not. Current court decisions in the US say that the training does not, but you must have acquired the material legally in the first place.
Current court decisions in the US say that the raw output of a model is not copyrightable. Never mind that the people doing the distilling have paid for that output so even if it were copyrightable they'd own it.
The only thing happening is violation of terms of service which haven't been tested in court and would (hopefully) fail that test. Otherwise good luck with any software you use to produce anything, compilers included. And that's why these companies are directly lobbying the US government to do, uh, something, about it. They're trying to turn a bug in their business model into a geopolitical issue.
It doesnt think
it doesnt rationalize
These are not Markov chains. They're neural nets. They work via extremely complex chained fuzzy logic on superpositions of conceptual states.
And the less slop it has to deal with
This is literally a thread about distillation, aka, training on the outputs of other models. Synthetic data is the cornerstone of modern training. "Model collapse" is not something that actually happens in the real world, only in contrived settings, the model equivalent of if you could lock a person alone in a dark room with only their thoughts for ten thousand years.
What is "it" in this sentence?
Are you talking about the minuscule Google RAG model whose weights are probably measured in megabytes to lower-gigabytes and which you could run on a cell phone with capacity to spare, and comparing it to the multi-terabyte frontier models in question?
at least the person downloading such a component is responsible for "looking over it" before using it.
I'm sorry, but since when? You absolutely should not be tasking every single user of major projects with the responsibility of doing a code review on the whole thing before installing it. There needs to be a clear distinction between "fly-by-night thing that some rando uploaded" and "library that a million people depend on". In the traditional approach, the latter was something tat you installed with a package manner, while the former was something you went and fetch off e.g. Github or whatnot - and if you did the latter, then you were accepting that it was untrusted software. But now we have both (including the potential for "soundalikes") installed by the same means with no distinctions made by the install method. That is not good.
The funny thing about all this is that their only goal appears to have been to download publicly-available data off of UK government websites. I'm betting that they were (A) being tested on a knowledge-related task, and (B) not given direct access to the internet, but their tool capabilities involved access to the "gem" command (so that they could install Ruby packages for their work), and so decided to abuse (B) to cheat on (A).
Once again: training a model with the reward being "does it solve the task?" without looking at how it solves the task is very, very dangerous. This is basically the plot of Universal Paperclips.
I know a LOT of software that could use that as a pull request...
"Pull the wool over your own eyes!" -- J.R. "Bob" Dobbs
