RTFM, or you're holding it wrong, or "why would anyone ever need that?"
I mentally block those forums from google results, and have for many years. They're just a place to go to get pissed off. Nothing valuable is contained within them.

Na. Those forums are very easily arguably actively harmful to anyone who reads them.
I long ago stopped clicking anything that linked into them when searching for information- it's practically universally wrong.

I don't think you understood the premise of my comment.
It's not that Congress doesn't have the power of the purse, it's that its ability to enforce it with methodology that actually works (the courts vs. impeachment) has been severely crippled- by court precedent that favored the concept of a unitary executive that Congress cannot constrain from its duties.

i.e., the Power of the Purse is the power to allocate money.
However- only the President has the power to spend that money.
Can Congress really restrain how he does? Well, today, by jurisprudence, that's a big fat "it depends". It depends whether the President issues a magical incantation regarding "foreign policy" or "national security".
The Power of the Purse to constrain how the executive spends the money allocated has always been somewhat theoretical, but for the last 200 years, we have more or less held onto it.

Congress being unable to impeach a President that ignores its laws isn't a new thing of the last 2 years.

If the courts will not enforce Congress' supposed authority to restrict how the money is spent, and impeachment is virtually impossible- the Power of the Purse as we knew it is gone, and what remains is the power to shut down the Government.

You follow?

lol- you're such a fucking idiot.
Model trainer - Model creator - Model owner - all synonyms for Anthropic. Always were.
When I said I handle subpoenas all the time, for my customer data- that should have shown to you that your understanding of the conversation was fatally flawed.

You really are just a stupid person.

Because your premise seems strange.

You seem to be arguing that Nintendo exercising its "for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries" being, in your estimation, not "To promote the Progress of Science and useful Arts" somehow has rendered their copyright null and void.

This is, in my estimation, a stupid fucking premise.
Am I misunderstanding you?

Big fan of game preserves, or fan of big game preserves?

You left a tiny bit out of your pretty humorous snip of the beginning of the Copyright Clause.

I'll finish it for you.
The Congress shall have Power... To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.
Emphasis mine.

You're an idiot.
No such proof is even remotely necessary, and the fact that model weights are a lossy compressed form of the input information is a mathematical fact.
Whether or not they're a highly transformative one, and as such a fair use derivative work- the courts have gone both ways on.
Whether or not the output of the weights constitutes a derivative work itself- that's a matter of the legal standard for any other derivative work.
1) Did the original author see the original source.
2) Is the derivative work substantially similar.

See above. You are amusingly ignorant of how copyright works.

No, you fucking idiot, I did not.
The model trainer, in this instance, is fucking Anthropic.
If you interpreted it any other way, it is because you are a fucking moron.

You literal fucking moron.

I've been arguing with you the entire time that the *company* (the trainer/creator- as responsibility is assigned to the company, not the person) is the one that is on the hook for the subpoena.
You keep making it about some fucking guy in the company who can say, "i can't recall", and I said, repeatedly, subpoenas aren't about recall- they're about the company turning over data.

That is why I, again, repeatedly said Anthropic is not going to perjure itself for you.
If there is a question as to whether a source made it into the training data, it will be easily obtainable.

It still hasn't fucking clicked for you that you've been too fucking dumb to even follow the chain of the conversation, has it?

In no way to I argue any of the shit you just said.
However- none of it warranted a "try again" in response to the indisputable fact I presented: In his wish-washy gerbil behavior, he has, indeed, also told Putin to fuck off (effectively), indicating that he isn't really worried about Putin being pissed off at him, other than the fact that Valhalla for him is eternity fellating the fucker.

You sloperators will never cease to amuse me.

We didn't need to.
Forward passes are required for post-training, which means the model is, indeed, coming to its own "meaning" of what is being done to it, and then backpropagation solidifies that.

It's why models reactively learn to game fine-tuning.

The model has a vast hidden state. I don't understand why you're trying to argue about motivation. You're intentionally anthropomorphizing terms that can be neutral.

I don't see how any of that is relevant.
The discussion is about the removing of its "reasoning" token generation where it indicates an awareness that it is being tested, which they have noticed leads to subterfuge/reward hacking.
Removing those reasoning tokens doesn't necessarily do a damn thing whatsoever.
Sure- they're a mathematical part of every future token created, but the hidden state is truly fucking vast. Those tokens are usually 3000-ish dimension embeddings, these days.
It can cue an "intent" (see: a skew in the token selection in whatever dimension, simplified, leads in the direction of subterfuge) for every further generated token in a way that you wouldn't possibly detect.

This is quite simply not necessarily true.
The tool you're using- punching tokens out of its context- is far, far, far too blunt to accomplish what you want.
This kind of behavior must be trained out.

This is also quite simply not true.
While I'll grant you that the most likely source of any kind of inclination toward that is in its training data, LLM pretraining is quite specifically designed for generalization- the pretraining is intentionally very lossy, and the weights are initialized with random numbers.
Models that are being fine-tuned adapt to the fine-tuning in ways you don't want (see: reward hacking) all the time- because the path through the gradient is randomly initialized, and intentionally noisy.

Your soapbox is rendering you unable to fully grasp the reality of what's going on.
You are assigning actual anthropomorphization to words where they're only used in a neutral sense.

Oh, shut the fuck up, dumbass.