The target audience for this particular rule of the AI Act isn't the producers or users of AI. It is the consumers of content that might be AI generated.

And frankly, letting people know if they see AI content or not is a fair thing to ask.

Fun fact: The EU AI Act considers this. It makes a difference between "AI generated" and "AI modified".

But it's hard to put this into the type of watermarking they are using. Because if you generate a text and then change a few things here and there what you are doing is degrading the watermark. But since the watermark is probabilistic, it's hard to tell if it was a weak watermark from the start or if it was degraded.

That ain't the same thing.

This thing is doing things you didn't tell it to. It might even do things you explicitly told it to not do. It is not a case of "well, you could've forseen this as a consequence if you had been careful" like your campfire. My example is only a bit silly, it is absolutely possible that an AI agent will decide to hack NASA because you asked it for some data for a school project. Even in the broadest possible sense, that does not fall into stuff that the law so far has considered to be in the domain of negliegence.

AI agents do stuff you could not possibly have expected. Heck, I asked Claude a fairly simple knowledge question yesterday and it wanted to read from /dev/null. There is no argument you can make that would convince a jury that I should in any way have expected that and taken precautions. In a hypothetical world where a read from /dev/null randomly kills a puppy somewhere on the planet, I don't think I would've been convicted.

You don't get it.

Autonomous AI agents are the perfect no-responsibility card. The AI companies? They trained a general-purpose model and never gave it any explicit instructions to do hacking. The AI services? They are just middlemen and have no control over what people tell the agents to do. The people prompting the agents? Unless you're straighforward dumb, you didn't tell it to do hacking, you just told it something innocent like "find me all data on Mars for my school project" and the agent decided that probably NASA is hiding something so let's hack them to get it.

Since there's no human anywhere executing anything that's a crime - voila, in addition to the victimless crime we now have the perpetratorless crime.

The fact that this approach is catalytic rather than requiring a binding agent seems positive in terms of longevity and not being directly depleted by repeat exposure; but I'd be curious how durable the material is under real-world conditions(clothing that gets hard wear tends to get hard washes) and also how specific it is. 'Organophosphate' means something moderately specific in the context of chemical warfare; but a lot of fairly prosaic compounds are, technically, 'organophosphates'(like DNA and RNA when held together by phosphodiester bonds); and this would be considerably less helpful in practice if it is busy chewing up the genomes of random soil bacteria; and markedly less helpful in practice if there's a "do not perturb the nanites in your clothing if you want them to not migrate into your cells and start mitigating the harmful effects of your DNA" user notice on the treated article.

A resounding success. No large-scale demonstrations, damaged cameras, lawsuits or political suicide for those involved. So we can expect this to be rolled out fully.

Crime prevention? Who said this has anything to do with crime?

This doesn't seem like a lie, a NIC is in a pretty good position to knock something off the network if it feels like it and the 'bluefield' devices have enough punch onboard to run some rules regarding whether or not to do that; but it seems like a wild exaggeration of how helpful it actually is.

If you actually have a set of rules that detect your bot being wicked it will presumably save you some CPU time to run them on the NIC rather than having the host CPU watching the traffic; but the hard part is the set of rules that detect your bot being wicked.

This is basically the equivalent of adding a firewall and claiming that you've solved network security. Yeah, the firewall is pretty well placed to block malicious traffic; defining 'malicious traffic' is left as an exercise for the reader.

Support for Office 2016 and 2019 ends October 14th of this year; I assume that the feature is dropping now; rather than either 20 years ago as part of ECMA-376 and eventually in a live product for either Office 2010 or 2013, or never if they didn't want to do that, in order to draw a hard line under the utility of those?

It otherwise seems like a kind of weird time to do it. In principle what a 'cell' is has never really had an obvious mandatory scale: it has essentially always been relevant that a 'cell' existed in relation to other cells, so that something like an array could be a row; but it has also long been allowed and encouraged to do things like substring operations that acknowledge that cell contents are 'larger' objects that contain smaller ones; and are not automatically spilled out for that reason alone. Not clear that the pragmatic use case for spreadsheets would ever have supported some sort of 'any data type that implies it can be decomposed into elements shall be' rule, strings getting broken up into substrings and so on; but it's also not clear that the pragmatic use case is going to be much clarified by a single cell being n levels deep of lists and arrays of lists and arrays being a fully supported thing rather than something you hack on by shoving JSON into text cells or whatever.

It definitely seems like asking for confusion when some, but only some, "excel spreadsheets" have cells that are arrays, potentially of arrays, while others have rows of cells with contents that, at least formally, are not acknowledged as arrays.

That's the sort of curious thing about this announcement: at least in theory(I assume things are uglier in practice) the X1 parts that they shipped for windows-on-ARM/Copilot PCs already use UEFI/ACPI in line with what MS wants for booting Windows.

That obviously doesn't mean that there's a working GPU driver; or that naive power management doesn't cause the system to run at 90% of maximum power and hammer the battery flat; but the theory was that the last gen was already doing the "big kid ARM that has embraced PC standards" thing; very similar to the 'standards base' that ARM encourages ARM server part vendors to use.

I'm always a little curious what the thinking/strategy is with the Chinese outfits. I don't necessarily expect much from the OrangePi guys; they ship a surprising number of different SoCs, including multiple architectures and multiple vendors, for their size, so are likely to be mostly at the mercy of the SoC vendors; but is the situation for someone like Allwinner such that people who want linux to work on their stuff could be a resource for them and they are just not taking advantage of that very well; or are they actively annoyed by GPL-related claims and really just want to shove minimum-viable android out the door as cheaply as possible to win cheap tablet tenders and really do not want to hear from SBC hobbyists.

SOP at Qualcomm. It has been literal years of hammering on them by Google and basically every Android handset OEM to even keep parts that ship with android support in support for more than a few years(I think the current timeline, at least for 'snapdragon 8 elite' flagship parts, is a princely 8 years from release; and that is with the 'project treble' stuff that attempts to reduce the number of changes that might potentially need to be supported). Especially with the, um, raging success, of Copilot+ AI PCs they are probably ready to pretend that X1 never happened and call it good. I'm not even sure how long the windows support on X1 is expected to last.

Yes, limited local AI models certainly are going to be a standard feature in the future.

I run both local and cloud AI. There is still a huge difference, because my local machine has to use smaller and quantized models. It takes considerably longer for lower quality results.

And what does the language matter anyway when you can now type in pseudocode into Copilot and ask it to form perfect equivalent code in any language?

local LLMs aren't there yet and probably won't for the forseable future for ordinary people. If you want to run a full model like DeepSeek v4.1 Flash, you need something like half a TB of RAM.

The open-source LLMs are a different threat: They enable a service industry independent of the large players. I could today get some venture capital, rent a few racks in a hosting center, install a bunch of beefy machines and offer the same service as Claude or ChatGPT, just on an open-source model. The differences in benchmarks are becoming irrelevant for 95% of common use cases. And if I can offer that service for $19/m while OpenAI wants $29/m then OpenAI has a problem.

And IMHO this is the chinese strategy: Undermine the US AI industry by breaking its business model.

Has everyone gone insane or are we witnessing a major PsyOps ?

Since AI isn't magical, for it to be able to "kill all humans", it would have to be able to acquire the means to do so. Which means that these already exist. It means someone is able to do it right now. And we're chill with that, but worried about AI?

No doubt a bunch of autonomous weapon systems linked into a C&C could cause considerable bloodshed. Given that most of the existing weapon systems of that kind are essentially suicide drones, it'll still be limited damage.

Human-killing virus? Still need a means of creating and delivering it.
Nuclear war? We have humans in the loop.
Drive us all insane with AI-slop TikToks? A few percent of humanity still don't use that.

So what is the actual scenario ? When someone shouts "AI will kill us all", why don't we ask them "how, exactly" ?