impossiblefork

That's a real risk.

Hopefully it's for buying VSORA or Euclyd machines, but well...

It turns out though, that this doesn't matter. There's actually a list of people who are known to have been tortured to death in these places on Wikipedia.

Still, I think there's a point to this statistical reasoning: Assange could have made it at the time, and been correct, and thus we have a strong argument for his duty to act under Icelandic law.

Censored variables are a very real thing and this model is perfectly legtimate. This is how you do maths with conditional probabilities.

It's ad hoc, but so is all maths, and all reasoning.

A man with a paper cut is at D=0.9999 or more. You probably have more noise from variations in sleep or eating or colds. But people don't end up at D=0.17 just randomly. El-Masri was in danger of dying. The fellow was actually tortured.

There's also no censorship in the example of the paper cut. It also isn't possible to infer much from a something like that E[D|some positive condition]=0.9999, so even if there were censorship your example with a small scratch doesn't allow the sort of argument I gave in my previous comment. My previous comment specifically relies that you have a censored sample D~P(D|something that you'd expect would make D large) and then finding that this D is small.

So imagine a censored random variable.

Let's say that D is a random variable representing whether a certain person in one of these black sites is dead due to mistreatment. It's not so simple that D=0 if the guy lives, D=1 otherwise, D is a continuous variable, a damage level.

You can't sample from D ~ P(D) because people aren't releasing such information; and we know this, because when El-Masri sued the US he was prevented from doing so by means of the State Secrets doctrine and it was argued that whether he had been tortured and by whom was a state secret. Consequently what we actually have samples of is P(D|released) and I think P(D|released)=P(D|innocent,recognized as innocent,slightly lucky). So if we have D ~ P(D|released) where it turns out that we have observed a realization of that with D = 0.17, do you think there are only a few outcomes where D=0, when taking into account that E[D|innocent,recognized as innocent] should be much larger than E[D], and when taking into account that P(released) is really small, something like 10^-4 or 10^-3?

No, it does not.

There's a guy from Lebanon who ended up moving to Germany called Khaled El-Masri who got seized in Macedonia by the US and who got sent around a bunch of US black sites.

Eventually people figured out he wasn't Khalid Al-Masri and started treating him better, but he still almost died. Consequently, the people who weren't treated better almost certainly did.

This can't really be disputed. The US certainly killed people at these black sites. The obvious danger El-Masri was in makes that clear.

There is a screw-up in my previous comment since I say Poland rather than Macedonia or Afghanistan, but there's reason to think that the Polish black site is iffier than the others because of how the investigation of the location was performed.

My objection is absolutely with Assange's case. I don't particularly approve of the idea of these sorts of plea deals, but they are not the core problem.

The core problem is that Britain pressured Assange to enter into a US plea deal over something which it was entirely legal for him to do, seeing as he was in Iceland when the acts were committed.

Seeing as part of the information obtained related to black sites at which there is reason to believe that people were killed (we know that that a person released from the Polish black site were almost killed, so we can be fairly sure that a bunch of people died there), and Iceland has a law making it legally mandatory to help people who are in danger; and I think there's good reason to believe that he had reason to believe that the information would reveal such things, so that it wasn't just him doing good, but him doing his legal duty, once he had reason to believe that people were in danger.

So my concern has very little to do with plea agreements. They have to do with Britain.

Yes, that he was forced to do so by the Brits in order to be release is what I referred to, so of course it is coercion.

He is of not American and has no duty to be loyal to the US or to refrain from committing espionage against it.

No.

Assange didn't commit a crime, and was still forced by the Brits to confess in order to be released. Snowdens situation is completely different.

His only hope was a much wider view of the legality of the acts he'd taken to be considered, taking into account the legality and constitutionality of what he released information about, and there was and is no chance of that.

Generalization from few examples is something transformers are bad at. They memorize outright. This is probably in part the objective, but humans have multiple readings of something they see, so generalization.

Transformers can also become confused be texts that no humans become confused by. I wrote some stories that I've used as test material, where I deliberately refuse to say who is speaking, or whose perspective we see, but where it is obvious to a human who it must be, and LLMs can make huge screwups in those texts. Mixing up an old guy with a young guy who he, when he was a similar age, was similar to, mixing up a kid with the kid's mother, that sort of thing.

I think the first part has no know solution. The second part is probably solvable, but not with a transformer-- maybe if they could make notes or output reasoning traces during prefill.

Both problems are hard and unsolved though.

The profits of AI will itself will probably be small though. Most of the benefits of the extreme scenarios will probably accrue to normal capital owners-- basically, the useless landowner.

Consequently things like "Sovereign AI dividends" and the like won't matter in the extreme cases. Profits on a scale that matters for society AI itself will only be a thing in an intermediate scenario which will probably only be real for an incredibly short time.

Morally I agree, but since there's probably a lot of LLM text in the training data, distilling on another model will probably make your model copy the values encoded into the other model as well, even in cases where you only distill on value-neutral stuff.

By copying their programming style, you'll move the model towards that way of writing, which will move the model towards the values expressed in those documents.

I feel that Deepseek v4 got so claudified at the end that it was like Claude.

Even if you apply Goldfish loss or other things like that, they still understand the gist of the thing they're trained.

That's of course the whole point of things like Goldfish loss.

Here in Sweden the murder of Salwan Momika makes me feel that states wanting to limit the anonymity of protesters must take rather great care to protect their rights.

This, I think must involve at least two things: strong laws protecting against creating any kind of registry of people who have participated, or who are likely to sympathize with any particular kind of protest, exceptions in cases where it there is a possibility of retaliation against protesters and that the police take care to actually protect protesters who are targeted because of their participation of protests, so that cases like Momika's murder do not happen again.

I really can't comprehend wanting to enforce this kind of thing at the EU level. It's not like they can't pass this locally if they want it, and leave it to others to have other rules.

I think it's much more interesting to deal with the algorithms, etc. I agree with that bit about not leaning on parental consent though.

It's always good to set up objective criteria for decision so as to prevent corruption, yes.

But when ad-hoc deals are made anyway, then it is more important that principles like those have been applied than that everything is ideal.

So, the problem in this case is that there's no gain for the country building the datacentre.

If you're in Finland and there are two possible uses for electricity production, let's say, either a steel plant or a datacentre. The steel plant will employ a bunch of people locally. A datacentre will employ a bunch of people in California.

So if you are to build a datacentre, the deal must necessarily be that the R&D for the models that are to run on it must happen locally. Otherwise there's no reason to give them the allocation over the steel plant.

Ah, I didn't intend to imply anything of that sort.

It's just that I, having never seen what that would actually be like, imagine that I'd be fine with a genuinely democratic mass-organization for LLM development which has a cutthroat, even OpenAI-level cutthroat commercial arm.

So you can make a huge internal model that you can then distill from?

As a Swede, I can't read the Emil books because of fremdscham for the parents from those scenes where Emil feels the need to run to the carpentry shed.

The farmer class could actually be like this, before we banned it, even though she tries to write about it in a comical way, but it was always a low class thing to punish ones children and to read stories where it happens is basically intolerable at least to me.

Astrid Lindgren probably knew this though. She isn't some idiot who puts this in as comic relief, it's comic but there's a serious and intolerable feel to it too. She knows she's portraying something bad and she intends for us readers to sit with the dissonance-- the low class going-to-punish-his-children aspect, the family's love for Emil, that book in which his mother writes down what he does, that Emil is well-meaning, that everything goes well in the end and that the adult Emil becomes a nämndeman, etc.

Astrid Lindgren is one of the scariest Swedish authors because of her deliberate careful and nuanced use of moral dissonance. She's scarier than Willem Fredrik Hermans.

Same way for me.

Imagine what a genuinely openness-focused organization of this sort could be. Even if we imagined a commercial half, we could imagine a foundation with mass-membership, perhaps with a membership fee equal to 1/2 the typical personal subscription and functioning to set the direction, elect the board, etc., and then a commercial half which might be rough, tricky, deceptive, making deals with anybody.

I think I'd have been fine with the commercial half being a bit of a monster, as long as I'm part of the members and we decide what sort of board it gets and there's a clear "this is basically controlled by the public" and if I were part of a club of this sort, I would, like you absolutely fill up a directory with texts and computer programs and careful annotations to aid training.

and they could have had it. It could have been easy to make an organization like this. I think you still can. An international AI club, the members vote on what sort of training material may be supplied and for what intents, create some committees to review quality, and then everyone starts making their little games and RL environments and annotated stories and programs that ordinary LLMs misunderstand, and then they get together and fine-tune something, and if that works well they then get some staff and better training infrastructure and end up with a commercial half.

Why do you think it will be net-good for society in the long run?

If wages go down, worker power goes down; and if dependence on labour goes down, you literally move closer to the situation of the extraction economies in things like petro-states, leading to oligarchy instead of democracy.

It could be the end of ordinary people's power over society rather than anything even slightly good.

Models are very obviously continuously updated.

Model editing to remove PII that slipped through, all sorts of things of that sort.

I haven't read the whole thing yet, but I think this is a really important paper.

I used to despise this kind of thing but it sheds light on the enormous generalization problems that aren't even close to being solved.

If Mistral aren't going to distill other people's large models they obviously need to train their own large models. This obviously requires money for optimization, tuning and training hardware.

Preferences come from selection pressures and needs. Taking them as a foundation leads to excessive focus on trade and obscures the fact that humans in almost all cases need essentially the same things.

Yes, but the court in this case, is taking a particular interpretation that overstates the novelty of the rights claim.

The people bringing the lawsuit are not in fact asserting a right to clean water, they are asserting a right to not be deliberately poisoned by deception.

Still, it's very obviously a deprivation of right to deliberately trick people into drinking poisoned water or other poisoned things.

The right not to be deliberately poisoned doesn't disappear just because the act was achieved by leading you to the poisoned object and deceiving you that it was safe, instead of handing you an object that has been poisoned by hand.

This does not imply a right to clean water, as the judge chooses to interpret their question as asking him to recognize, and shows that his choice is in fact sufficient even if that is what they asked.

Yes, but surely it's a deprivation of rights to trick someone into drinking poison?

Isn't that even how the gas chambers at some extermination camps worked? They tell you it's showers and you go in "voluntarily"?

Yes, of course, but he interprets the arguments and evidence presented to them.

Here he chooses to interpret what they say as demanding the recognition of a new constitutional right to safe water and to correct information. It is not at all clear that that's a good choice.

He could equally well interpret the situation by analogy: the government deliberately tricked people into drinking unsafe water, and people were, depending on whether there is evidence of poisoning from the lead, either provably or not provably harmed.

Thus it is open to him to interpret their question as one that is not so focused on novel constitutional rights.

What I'm saying though, is that he has room enough to interpret things so that a dismissal isn't necessary.

It isn't the job of the courts to make minimal rulings and avoiding all substantial questions, but to solve disputes. Here we have a case where the government has harmed people in a way that would be a tort if it had been done by anybody else, and people arguing that the government by doing so breached a right. There is presumably a right which can be interpreted as guaranteeing that people's bodies won't be messed with, or things in the US would be quite strange, so there's room to do something useful here, and the judge's choice is not forced.