Yes of course someone pays for it, in this case your deductions as you say.
But I think there is a fundemental difference to employers paying for health insurance in that it doesn't depend on your job. So if you lose your job you don't lose your healthcare so companies can't use that as a way to retain you.
And the actual cost of healthcare to the organisations paying for it is actually far lower than the US system, probably partly because it's more regulated and also because there is far less litigation so insuranace for doctors is cheaper.
So I don't think the US system is "more efficient", unless by "efficient" you mean in extracting money from patients / their insurances. In the US hospitals exist to make money, in the EU it's more about providing treatment.
There's no effective way of enforcing export controls on local software like PGP etc. Whatever they say someone will leak it.
It is possible to shutdown access to hosted services, as happened with Fable, but it can't really be done selectively. The US government wanted to allow it for US nationals only but Anthropic couldn't do that and so shut it down for everyone. Even if they did tie Claude accounts to nationality some people would set up "proxy servers" to allow access (either for montetary gain or because they don't agree with the restrictions)
They can restrict it to US citizens only: just do KYC and enforce it. Any citizen exporting ITAR capabilities would be committing a felony.
I think it is inevitable that more capable models will require export controls, KYC, background checks, verification of credentials, etc. Even if you don’t buy Anthropic’s marketing of their current models, there is a future where models are capable of developing chemical weapons, biological weapons, cyber weapons, etc. which are genuine security threats governments will care about controlling and preventing. Maybe this prevents big commercial labs from developing more capable models until they figure out appropriate safeguards, but this is a good thing, a world where you can ask an LLM for an airborne rabies genome and it will simply produce it for you is not a world we should want to live in.
If a model is really capable of that type of stuff (creating biological weapons etc) the problem isn't solved by export controls - I'm sure there are plenty of home grown US terrorist organisations that would like that capablility and wouldn't be subject to such controls.
Yes governments can, and probably should, restrict some things completely to all those outside of official government institutions but once something is public in one country it is effectively, even if not legally, public in all countries.
But for "physical" things like biological, chemical or nuclear weapons the know how / information part, that AI can "help" with is far from the complete picture. Articles have already been published on how to make a nuclear bomb. The knowledge isn't really the blocker that's more access to the required materiels and equipment. However in the cyber space then yes AI could indeed give the "bad guys" much more capability.
Identity theft is a thing. And if you gate a desirable commodity behind an identity it will become even more of a thing. There are 100's of millions of identities to steal.
You could quite effectively set up access to Fable with know-your-customer, attested clients/workstations, and then inspection of what the session is actually used for to detect out-of-country use. It's not impossible.
If they had said only governemnt and a few other approved institutions can use it then yes that could work.
But they didn't say that. They said "only US citizens". Once something is available to hundreds of millions of people there will be leaks like it or not. Some accidental and some voluntary.
And even if it could be done it wouldn't solve the purported security issue anyway. Does anyone believe there are no criminals or terrorists with US nationality?
Interesting that those supporting the motion claimed it was because there was no space left for new arrivals and that it put too much pressure on infrastructure like trains and yet the largest support came from the countryside which proably has less overcrowding and the cities were greatly against it.
Makes me think that "overcrowding" wasn't thre real reason...
Probably a pretty big selection bias there. People who don't like crowding and want space are more likely to live in the country. People already living in a dense City are less sensitive to more of the same.
I have a friend who voted yes and he lives outside of Basel because he thinks the city is too crowded. If you've been there you might find that amusing. His most vocal complaint is about too many people on Swiss hiking trails and difficulty in Booking huts.
More recently, he was laid off by a multinational pharma company where most of the Swiss office is now German immigrants.
That's not to say culture isn't a factor, but it would be a mistake to apply American ideas of racism to it
This is a collective action problem. Its too easy to claim the answer is "we'll hire our competitors juniors when they become seniors" and then every company wants to do that and not train their own juniors. Soon no one is willing to train juniors because theyll just leave for the companies that dont train juniors.
It's actually worse than that. It's not just financial depreciation or that the existing hardware becomes obsolete due to being less powerful than new hardware but also that hardware being run all the time at high load actually has a limited lifetime of a few years so it will physically break...
Data centers come down to performance-per-Watt. Electricity accounts for 20-30% of a data center's operating cost [1]. I don't know the exact breakdown but the GPU part of that is probably the majority given how power hungry GPUs are. The B200 is upwards of 1200 Watts [2]. The B200 is rated at ~4.5PFLOPS of dense FP8. So you're getting 3.75PFLOPS/W. We don't know what the next generation will look like. The A200 (Hopper architecture card that preceded the B200) had ~4PFLOPS apparently but also lower power consumption. Obviously this changes depending on whether you're looking at dense or spare and FP8 vs INT8 vs INT4 vs FP4, etc so we're just using FP8 as a yardstick.
Imagine a fictional B200 successor, the T200 that has 8PFLOPS of dense FP8 at 1000 Watts. Well then a DC built on that where the T200 will likely cost similar to what the B200 does now, you'll get nearly double PPW so the same size DC and same electricity load is going to be like 2 of your old DCs in operating costs. That's a big deal when you've laid out a trillion dollars.
Seems to be a thing in the US to need specialised software, an accountant or AI to file taxes.
In most of Europe individuals at least don't need any of that.
I'm in France and it's just a connection to a government run website to enter a few figures, takes less than an hour most of it is already pre-entered (salary etc), the main thing to add manually is charitable donations.
If you're running a business then yes an accountant can be good (or be required depending on the legal form of the business) but not for individuals.
In the US you don't need software or an accountant either. People have been convinced it is hard, but the reality is taxes are basic math for most people. It isn't hard to copy the few figures from a couple forms and add them up, the look things up in a chart. It is tedious, but not hard. That said, I use the software anyway because once in a while I make a human mistake and those are annoying when the IRS catches latter.
There are a few people who have a complex situation who need an accountant. Most do not, but since they have never looked at what is really involved they don't realize what is going on.
It will become more expensive to fix than replace. Also more energy intensive than newer generation to operate. MBTF is significant the older the fleet gets higher the failure rates .
A typical node today is 8 GPU node today , you have to keep replacing failed GPUs by cannibalizing parts from other GPUs as nobody is selling new GPUs of that model anymore at higher frequencies.
In addition to outright failure there are higher error rates in computation in graphics it tends to be flickers or screen artifacts and so on.
Azure operated K-80s and P-100s for 9 and 7 years respectively but they were running at 2 GPU nodes and of course were much simpler compared to today’s HBM behomouths on 2/5 nm processor nodes . Google operates their custom ASIC TPUs for about 8-9 years .
With custom inference ASICs like cerebras hitting production the cascading of training NVIDIA chips to inference to get the 5-6 year useful life is also not clear.
