simiones

> If the subject can utter the action, then it can perform it.

This sounds like another layer of weird terminology that doesn't mean anything for someone who is not familiar with whatever capability system you're thinking of.

Say I am a user who can see a particular directory on a shared setup. I try to upload a file in this directory, using the same method that worked on another directory. The question of AuthZ is: will I be allowed to do it or not? In the plain sense of the words, I can absolutely "utter the action", I have all of the "verbs" (upload) and "nouns" (the file, the destination path). Still, I should not be allowed to perform the action if I was only given read-only access here.

Now sure, you can say that "upload to dirA" is a different verb than "upload to dirB". But this is just confusing terminology, it doesn't enlighten anything.

Do you have proof Apple is facing unexpected demand for the Mac Mini because of AI? Do these news sites?

This comes off as claiming that no one could possibly read Harry Potter if they haven't read The Illiad yet, nor the Illiad if they haven't read the Epic of Gilgamesh.

Note that it's not uncommon for even more widely prescribed medication to have no firmly established mechanism of action. Acetaminophen/paracetamol is probably the best example.

Ultimately medication is prescribed based on evidence from clinical trials, whether or not the mechanism of action is fully understood. SSRIs work to help with depression in clinical trials when compared to placebo, so they get prescribed.

> SSRIs do not have the same level of supporting evidence, and the effects are not as durable.

To the extent that we can measure depression, SSRIs have been widely proven in gold standard phase III clinical trials to help with the treatment of major depression. What exactly is supposed to be lacking in the supported evidence?

You are assuming a lot about the competency of some specific doctor who knows where in the world...

> Are you sure about that? At that time, hard drives were constantly spinning. Getting them to suspend was a rare feat.

I am pretty sure, though, of course, memory is highly fallible (and note that I was a kid back then, and didn't understand that well what those sounds or the blinking disk light actually meant).

> 20 years ago, Windows 98 had long been abandoned. At that time I had been running a Linux desktop for many years already.

20 years ago means the mid-to-late nineties. That's my position and I'm sticking with it /s

But I think it's undeniable that Linux in the 2000s was a massive pain for anyone but the most dedicated tinker, and even then only for pretty carefully chosen hardware. The Year of the Linux Desktop meme existed for a very clear reason. Virtually the entire PC user base, both Consumer and Corporate, was running Windows XP in the 2000s.

> This allegedly violates the GPLv3 clause that requires the release of the source code for the entire work if they are combined to form a larger program in or on a storage or distribution medium, such as the flash memory of the STM32F407 micro-controller on the board, and more recently in the update images Prusa supplies

This makes 0 sense as presented here. The GPL explicitly distinguishes mere aggregation from combining in a way that makes it a larger program. Putting both the proprietary bootloader and the GPL OS on the same disk / in the same system makes 0 difference to whether the bootloader is covered by the GPL or not.

If this were true, then any phone, any PS5, any Linux laptop, and almost all consumer hardware would be in violation of the GPL (note that the wording quoted there is also part of the GPLv2, it's not GPLv3 specific).

Do you mean physically modifying a printer you bought from Bambu, and selling that on? Potentially as a service, where you buy in bulk and re-sell with this modification? If so, I don't see how that would change anything about the software you're allowed to run on it.

Or do you mean manufacturing your own printer that is a copy of theirs with 10mm extra build volume in each direction, and with your own software? That may be illegal if they have any patents, but otherwise should probably be ok, and would of course not have anything to do with their software - but you still couldn't take the software from one of their printers and running it on yours, most likely.

Nothing stops anyone from briging a suit and representing themselves.

Can you share a link?

This would be a vindication of the SFC's position, which holds that what people commonly call "TiVoization" was in fact already not allowed by the GPLv2, and that the GPLv3 makes a slightly more complicated requirement.

Specifically, the SFC generally claims that the GPLv2 doesn't allow you to distribute hardware running GPLv2 software on it and then refuse to allow customers to modify or update this software themselves*.

Instead, they claim that what TiVo actually did that necessitated the creation of the GPLv3 is more complicated: you were perfectly well allowed to modify the Linux system on your TiVo box, BUT the proprietary TiVo software would thereafter refuse to run. Basically, you could either use the TiVo with only official TiVo software, OR you could use it as your own Linux PC, but NOT both.

According to the SFC, this is the only difference between GPLv2 and GPLv3 in this area - if the functionality of the product you received is allowed to be degraded intentionally because of modifications you make under your GPL rights. Linux with its GPLv2 license allows this, but much of the GNU utils, with the GPLv3 license, don't at this time.

* unless the hardware itself is incapable of loading modified software, of course, such as the case where the software is burned into a ROM and no one can update it.

It is actually a very relevant real-world definition - even though I agree with you on the intangible value of time.

The reason why it is very relevant is that, in the way we have unfortunately structured our economy, the amount of work you have to do to be able to eat and enjoy your time will always expand to grow to at least force you to maintain (if not increase) your economic productivity. That is, if AI enables engineers to deliver 10x more features in the same unit of time, then their employers will expect them to deliver 10x more features to continue earning their current pay.

So, while the AI did, in principle, allow you to just reduce your work schedule while getting the same pay, as you're providing the same results in 1/10th of the time, the economic system does not - your previous work is instead simply worth less than it used to be. Assuming you're not happy to take a 90% pay cut, you'll have to continue spending 40+ hour weeks, just using AI as well now.

> My machine 20 years ago was not an SOC and didn't have an SSD either. It was fast as lightning to open a window.

I never understand what people exactly mean by claims like this. I very vividly remember double clicking some executable, and then immediately hearing the HDD start to spin, and waiting with baited breath to see if it will spawn a window or silently fail, as programs on Windows 98 (the only option for personal computers at the time) often did.

So what exactly was snappy? Nothing that runs off a HDD is ever snappy.

In the sense measured in economics, only if the product at the end of the day is sold for the same nominal amount of dollars as before. It doesn't matter how many features or lines of code or bugs or anything like this you produced - it matters how much revenue your work produced.

So, if the product isn't selling any more expensive even though you're adding ~8 times more features, or if you're still getting paid for the full 40 hours even though you only really work 5, then economic productivity hasn't actually increased.

Becoming and then staying homeless for a significant time is often almost irreversible. The key is to prevent people from ever getting there, as much as you can - and that is the point the poster above was making. The fall from house + job > house > no house makes it very hard to ever get another job that actually pays enough to go back to having a house.

Per that link, this runs a Linux kernel as a VM on top of an L4-based hypervisor.

> but they need a native seL4/Linux if they want to honestly claim they are improving systems' security with their capability model.

Are you using "seL4/Linux" in the style of "GNU/Linux"? Because then it should be "GNU/seL4" - that would describe an OS exposing the GNU core utilities on top of the seL4 kernel. There's no way to mix the Linux kernel with the seL4 kernel, other than using one to run VMs of the other.

I don't think it needs to be so dramatic. You could have a proof that the algorithms don't contain data-dependent logic, and perhaps ensure that no data-dependent instructions are generated; and that all branches have the same number of instrution-cycles (adding padding if not). You'd then simply rely on the architecture-specific instruction timing differences to be respected by the compiler.

Would something like this guarantee that no side-channels are possible on any architecture? Perhaps not, but it would still get you most of the way there.

> If not, then what would be the result of the multiplication of 3 with "m"?

The answer is not, and the result of 3 multiplied by m is 3m. Just like 3 multiplied by pi is 3pi; or, perhaps more accurately, you can view m as a kind of vector unit, and 3m as the scalar product. Of course, none of this is exactly matching - dimensions are different from irrationals, vectors, complex numbers, etc, they are mostly a thing of their own.

> What this gets you in the end is a type algebra, but that is also not exactly a new concept.

Sure, that's why I said specifically programming language types. I am aware that type theory has way more complex operations on types. I think some of these may even be expressible in Idris or Haskell + some appropriate extension. But in almost all programming languages, even ones like OCaml, SML, plain Haskell, Rust, C++ with template magic, Scala, F# and what have you, there is no way to specify that the result of multiplying two values of type A is of type "A * A", especially not in a way that then allows you specify that the division of a value of type "A * A" by A has type A. So types as exposed in any of the common programming languages are horrible for modelling dimensions as used in even high school physics.

The naming is irrelevant here. The point is that sin(x) ~ x for small x, whereas sinrev(x) ~ rev * x for small x, which is much uglier. And similar things happen to the derivative of sinrev() vs regular sin() and so on. So switching to preferring to express angles in revs instead actually complicates most formulas, at least in some domans.

Dimensions and units are separate things, though. For example, 1 minute and 1 second are different units of the same time dimension. Similarly, 1 rad and 1 degree are both dimensionless, but they are both different units.

This is precisely why (programming language) types are poor model of physics units, despite often being touted for this exact use case. 3m is not the same thing as "the value 3 of type meter". It is the multiplication of the dimensionless scalar 3 with the special "m" constant for meters.

That's why pow(3m, 2) = 9 m^2, and not `the value 9 of type meter`. Of course, you can define the type `square meter` as well, and define `pow -> meters -> positive integer -> square meters`. However this quickly becomes overwhelming once you start doing more complex expressions with multiple types. What is the type of `pow (3kg^2 * m/s, 3/2)`?

Edit to add: also, there is a simple fact that "sin(pi/2 kg)" is just not defined, in programming or math or physics or any other useful system. It's definitely not 1kg, just like sin ( (pi/2) * 2) is not sin (pi/2) * sin (2).

An important nitpick: Wikipedia requires reputable secondary sources, not primary sources. You can't cite a politician's personal social media, for example, you have to cite reporting on those statements. This is part of Wikipedia's defense against novel research - it's supposed to be a record of what is commonly known, not a replacement for, say, Reuters or arxiv.

It has been happening since people have existed. It used to be called propaganda before these, and rumor spreading since before the written word even existed.

> All those methods are applied with the specific goal of improving output quality and are applied to the extent that they do this

Yes, that's the goal that was used, but they are quite simplistic and crude methods, not some specifically designed function, with carefully fine tuned parameters or something. So, if a basic function like top_k can improve model utility, it's not impossible to imagine that watermarking could also happen to do so, or at least not have a significant negative effect. So whether the effect is deleterious or not is an empirical question, not something we can assume ahead of time.

> You're basically saying that if we train LLMs on watermarked text they'll be really good at producing text that looks watermarked

No, you're misunderstanding how the training works. If we train the model's output so that it minimizes the error function after the watermark is applied on it, the model will learn how to produce the best output it can given the watermark. It will produce better text that happens to be watermarked, not "more watermarked text". Same as if you train the model on minimizing `top_k_error(input) = |top_k_sampling(model_output(input)) - desired_output(input)|`, the model will learn to produce better output under top_k sampling, not learn to produce output that's "looks more top_k".

My main point is that sampling with a modified distribution compared to the one produced by the model is already being done, and it is generally found to increase quality, not decrease it. So there is no reason a priori to assume that the watermarked distribution would be lower quality than other schemes for altering the "raw" output distribution (such as top P, top K, temperature, etc).

My second point is that the training of a model by definition maximizes the fitness between the final output function and the training metrics. So, if the model is trained with the watermark applied, the training process will minimize the function `model_error(input) = |watermarked_sampling(model_output(input)) - expected_output(input)|`, by definition. This means that a model trained in this way will perform better when sampled using the watermaked_sampling method than if using, say, top_k sampling.

No one uses a pure random function over the whole probability distribution described by the LLM's output. For example, there is exactly 0 probability that the chosen next token by any common API or even local LLM runner would be a token whose final value is "0.0001" if there exist at least K tokens whose value exceeds "0.7".

Also, as long as the same sampling strategy is used during training as the one used during inference, then the LLM will actually do much better with the biased sampling strategy than it would with a fair one - because that is what it was trained to optimize.

An arena is a type of allocator that allocates efficiently and never deallocates (`arena.free(ptr)` is a no-op). That is basically what defines it. Since there is no free(), doing allocation via a simple pointer bump is the most natural implementation.

Not sure if they tried to do this in the UK, which I believe has very different standards to the USA for what it means for something to be a legitimate religious practice. In the USA I believe they actually succeeded in obtaining various religious exemptions, such as wearing a collander on their head in driver's license photos.

I'm personally atheistic, and thus have no religious practices whatsoever. However, just like it's legal to build churches, to ring church bells, to wear visible religious imagery, it's also legal and highly protected to talk to others, even strangers, about your religious beliefs (or lack thereof).