Airgaps are not flawless defenses. Ask, say, the Iranians about Stuxnet.

Even at the bare minimum, *information* is flowing through said airgap (otherwise, what's the point?), and the output information could be deliberately tainted.

There are legitimate concerns with the use of an AI model from a hostile state, esp. given how much today is agentic.

That said, there's no reason for a complete ban; surely having a standard set of post-training data and requiring that Chinese models first be posttrained should be sufficient? Unless someone finds a trick to embed harmful behavior so deeply that resists removal by post-training (and also resists detection in a way that can be abliterated)...

To be fair, AI is the fastest growing segment; 70% of new datacentre capacity is for AI. The main question is whether one things that demand will keep growing exponentially, whether there will be a "pop", whether people will increasingly choose "cheap models for the same capability" over "better models at the same price", and whether Jevon's Paradox will consume all of the efficiency gains (~100x in just a few years in terms of power and cost for a given model capability). These lead to very different long-term pictures. One can only really talk about the situation as it stands today, and to a lesser degree, over the next 1-2 years. Beyond that is pretty speculative, and it depends greatly on your assumptions.

That said, if one's assumptions are "AI is massively successful, Jevon eats all efficiency gains and then some, and people keep choosing the latest and greatest models", then one also has to assume the side effects of that scenario: massive increases in the economy, massive increases in tax revenue, and massive increases in the ability to solve the world's problems. So even if you assume a scenario where AI scales to consuming several percent (or more!) of the world's primary energy, you're also assuming a scenario where the world has vastly more resources to dedicate to fighting environmental problems (if it so chooses). Honestly, the real environmental threat in that scenario isn't the AI itself; it's that if the whole global economy is growing so much thanks to AI, then overall consumption of goods and services is correspondingly growing, with all the environmental consequences that entails.

And also to be clear, this is in no way to dismiss the sorts of challenges / threats imposed by AI - disruptions, wealth concentration (particularly if proprietary models win over open models, and if no political action is taken), humans using AI in warfare, humans using AI to manipulate other humans, etc etc.

Selling your own stock does not create a GAAP profit.

ED: "That's more than all datacentres combined " - that was supposed to be deleted, but I didn't proofread :P

Video gaming was 347 TWh/yr in 2024, and growing fast. That's more than all datacentres combined in 2024 was 415 TWh, and 485 TWh in 2025, but the majority of that is crypto (mainly Bitcoin), not AI. AI in specific in 2025 was 155 TWh (that's for both inference and training, all AI tasks combined) (It's quite trivial to show that it can't be much more than that because there simply haven't been enough servers produced). Total global primary energy in 2025 (not just electricity, but also including e.g. transport fuels, heating, etc) was ~177,000 TWh. I'll repeat: AI is a blip.

Also, I'm not sure why you thought you couldn't look up these numbers?

AI is a blip on global carbon emissions. Videogaming itself is a much larger power consumer.

Humans still review and authorize the specific decisions. And the bot will have spent far more time trying to meet the user's need than the user would ever have had time to do so.

Meh, we're talking something that should in bulk be the price of a low-end car and works 24/7 without breaks. Hard to make the case that that's not "economically practical". And as for realism, I'd take a mixed stance.

The truth lies somewhere between you and the GP.

Against the GP: the plan IS to have robots do those sorts of things. And there's real progress being made.

Against you: it's not simply a matter of "being programmed to do those things". Robotics is the least advanced AI field at present - in part because there's no convenient internet-scale training datasets to work with for real-world interactions, and also because accumulating physical real-world interactions impose time and cost bottlenecks. Beyond all issues of longevity in real-world (physically taxing / complex ) environments. I think a bet that we're *not* going to get their in our lives is a very bad bet (unless you're currently on palliative care!), but talking about it like it's just "some programmer needs to code a specific job" just isn't where we are - carefully curated and parameter-defined stunts aside.

Think, for example, of automated driving as another example. Very real progress over many years. Also still not to the mass deployment stage. And that's a much more "controlled environment", with much larger datasets, than general-purpose robotics. I must stress that "general purpose" aspect, because that's what's needed for the sort of jobs in question here. It's not simply "robot arm that's bolted down, target gets inserted into the target location, series of precise actions taken on it". The real world has an endless series of "unexpected things" in it, and a robot doing real-world physical chores absolutely must be adaptive to them.

Reality: No. They're forming a wholly-owned LLC, not a 501(c)(3) charity. They're treated as a disregarded entity for tax purposes. Transferring shares from Anthropic to the new LLC is just moving money from Anthropic's left pocket to its right pocket. You don't get a tax deduction from that. To get a deduction from IRC 170, you have to make an irrevocable transfer to a qualified 501(c)(3) charity.

If Anthropic later donates from this LLC to actual charities, it could then claim a corporate charitable deduction (against up to 10% of their profits - and note that word, profits; also, there's no breaks for UBI programs targeted at individuals). But the formation of this entity and transfer of 5% of their shares to it does not give them any sort of a tax break whatsoever, and there's no difference between donating from this structure to a charity and donating directly from Anthropic. In fact, by forming this structure, they're skipping a potential tax break that they could get by forming and donating to a 501(c)(3) upfront.

(If your next question is "why not a 501(c)(3), then?", the answer is because there's all sorts of IRS restrictions on 501(c)(3)s, including distributing at least 5% of assets annually, excise taxes on investment gains, strict limits on cash transfers to individuals, etc)

I get that it's popular for mod points to be cynical, but at least focus your cynicism on factuality. Try this: "We found a PR cheat! And how to turn things our CEO wants to do anyway and might not do until many years from now into a pre-IPO press release!"

Why?

I run a Shopify store, and if anybot wants to buy from me, by all means, and please invite all your bot-friends!

I think the only merchants who should be fearing this are ones relying on brand recognition rather than low prices and good products, since bots have more time to "shop around".

BTW, this also applies to old games. Have an old game that no longer runs on modern operating systems? In Europe, you have the right to task an AI to fix it so it will run.

Adding *new features*, mind you, is trickier. You're not allowed to decompile and alter a program to add new features, but you *are* allowed to decompile for interoperability purposes. So you are within your rights to task an AI to decompile a program to figure out how to link a mod into it and then to write a mod.

In Europe, it is. Article 6 of the Software Directive grants an "interoperability exemption". The lawful licensee of software has the statutory right to strip DRM in order to achieve interoperability. Any EULA terms that prohibit reverse engineering for interoperability are automatically treated as void. Article 5(1) of Directive 2009/24/EC also grants the ability to decompile to fix bugs that impair the ability to use the software, even if the supplier prohibits it. So long as you don't then start sharing your "offline mode" version of the software, you're within your rights

People in Europe are within their rights to task an AI to decompile and tweak a game they own make an "offline mode" version of it, or to create their own servers or whatnot if the manufacturer plans to shut them down or otherwise strip features that the user paid for.

Of course, the US is less consumer-friendly in this regard.

As for Islamic countries, there's only two that ban women from driving: Afghanistan and Turkmenistan - both de facto / arbitrary / adminstrative bans rather than official bans (e.g. refusing to issue/renew licenses, requiring guardians for long-distance travel, or things of that nature). Saudi Arabia revoked their ban in 2018. And Iran has always allowed women to drive (though they have to cover their hair while driving).