Ahh. TV shows before everything became political. Just two guys hating each other for very silly reasons completely unconnected to anything on earth.

Haywire. You don’t see that often anymore.

Text explaining why the neural network representation of common features (typically with weighted proportionality to their occurrence) does not meet the definition of a mathematical average. Does it not favor common response patterns?

Hmm. I’m not really sure why anyone would write such a text. There is no “weighted proportionality” (or pathways). Is this a common conception?

You don’t need it to be an average of the real world to be an average. I can calculate as many average values as I want from entirely fictional worlds. It’s still a type of model which favors what it sees often over what it sees rarely. That’s a form of probability embedded, corresponding to a form of average.

I guess you picked up on the fact that transformers output a probability distribution. I don’t think anyone calls those an average, though you could have an average distribution. Come to think of it, before you use that to pick the next token, you usually mess with it a little to make it more or less “creative”. That’s certainly no longer an average.

You can see a neural net as a kind of regression analysis. I don’t think I have ever heard someone calling that a kind of average, though. I’m also skeptical if you can see a transformer as a regression but I don’t know this stuff well enough. When you train on some data more often than on other data, that is not how you would do a regression. Certainly, once you start RLHF training, you have left regression territory for good.

The GPTisms might be because they are overrepresented in the finetuning data. It might also be from the RLHF and/or brought out by the system prompt.

I accidentally clicked reply, sorry.

B) you do know there’s a lot of different definitions of average, right?

I don’t think that any definition applies to this. But I’m no expert on averages. In any case, the training data is not representative of the internet or anything. It’s also not training equally on all data and not only on such text. What you get out is not representative of anything.

A) I’ve not yet seen evidence to the contrary

You should worry more about whether you have seen evidence that supports what you are saying. So, what kind of evidence do you want? A tutorial on coding neural nets? The math? Video or text?

That’s a) not how it works and b) not averaging.

Who exactly creates the image is not the only issue and maybe I gave it too much prominence. Another factor is that the use of copyrighted training data is still being negotiated/litigated in the US. It will help if they tread lightly.

My opinion is that it has to be legal on first amendment grounds, or more generally freedom of expression. Fair use (a US thing) derives from the 1st amendment, though not exclusively. If AI services can’t be used for creating protected speech, like parody, then this severely limits what the average person can express.

What worries me is that the major lawsuits involve Big Tech companies. They have an interest in far-reaching IP laws; just not quite far-reaching enough to cut off their R&D.

Why would averaging lead to repetition of stereotypes?

Anyway, it’s hard to say LLMs output what they do. GPTisms may have to do with the system prompt or they may result from the fine-tuning. Either way, they don’t seem very internet average to me.

You’re allowed to use copyrighted works for lots of reasons. EG satire parody, in which case you can legally publish it and make money.

The problem is that this precise situation is not legally clear. Are you using the service to make the image or is the service making the image on your request?

If the service is making the image and then sending it to you, then that may be a copyright violation.

If the user is making the image while using the service as a tool, it may still be a problem. Whether this turns into a copyright violation depends a lot on what the user/creator does with the image. If they misuse it, the service might be sued for contributory infringement.

Basically, they are playing it safe.

It’s all just weights and matrix multiplication and tokenization

See, none of these is statistics, as such.

Weights is maybe closest but they are supposed to represent the strength of a neural connection. This is originally inspired by neurobiology.

Matrix multiplication is linear algebra and encountered in lots of contexts.

Tokenization is a thing from NLP. It’s not what one would call a statistical method.

So you can see where my advice comes from.

Certainly there is nothing here that implies any kind of averaging going on.

mathematical average of internet dialog

It’s not. Whenever someone talks about how LLMs are just statistics, ignore them unless you know they are experts. One thing that convinces me that ANNs really capture something fundamental about how human minds work is that we share the same tendency to spout confident nonsense.

Well, maybe they can’t. This clause would probably not hold up in a lot of countries/courts. OTOH it would in others. It might take years of litigation to figure out.

So, if you want to work on this kind of thing, better consult a lawyer first. It will have a chilling effect and that’s something.

US situation: https://www.eff.org/issues/coders/reverse-engineering-faq

The EULA of the CUDA SDK bans reverse engineering output of the SDK to make translation layers (and such compatibility aids in general).

That makes it more legally dangerous and/or harder for devs. It has no effect on anyone not using the SDK.