Very much so! Unfortunately, they're still vaporware.

Or more accurately, I'd very much like a $10k BYD, but Detroit has bought DC wholesale to block any meaningful competition for Ford's Coal Roller 150.

That puts you on an iPhone 7 at the highest. iOS 15 is the oldest version installable on those, and not by coincidence, also the oldest iOS still receiving security updates. It was released five years ago as of September 20th.

So while you're 100% technically right... You might want to start planning for a new phone before October.

On the whole you're right because we simply don't know how to do it yet.

The GP isn't entirely wrong, though - It's also safe to say more training data is needed. Model weights vs training data follow a pretty clear power law, and current top models are already orders of magnitude underdetermined for their number of weights. Problem being, the current data set being used by the major players is effectively the sum total of all digitally-available human knowledge.

Unsupervised training is the hot new thing, but it doesn't really matter much if there's no more data for the AIs to find on their own.

Anthropic just accepted a USD$1.5 Billion court settlement for using illegally-acquired copyrighted content as training data.

Which side is not cheating here?

Bluntly, they can't. That genie is already out of the box for anything already available; and even going forward - Unless China itself blocks export, there's no realistic way to stop us from using a non-Chinese non-US mirror to grab them.

Realistically, hardware is the sole limiting factor for open-weight models, so restricting H200's is probably the closest we can come to a meaningful embargo. There's no practical way for most of us to run a model like Kimi K3 with almost 600GB (at BF16) of weights. Even using CPU offloading and/or multi-GPU sharding (if those are even possible for K3, I haven't looked at it in depth yet), the performance drop would give a throughput best measured in minutes or even hours per token, not particularly useful.

What this will accomplish is exactly what the OP quoted - It won't be possible to run these legally, so virtually all commercial use would be DOA. Joe Hobbyist couldn't care less though.

Because what amounts to a simple financial breach of contract is being enforced unilaterally and extra-judicially - And by parties traditionally known for being some of the sleaziest and least ethical bastards on the planet at that.

In a situation where the burden of proof is on the lender and the disruption to the customer's life severe, we should absolutely not be allowing ScumCo Lending to pull that trigger absent a court order.

To be clear, I'm not claiming there's no precedent for this (and when we're talking about microwaves and TVs, hey, pay up deadbeats!). But phones are unique in that it's virtually impossible to function in the modern world without one. Personally, I might get two calls a week and missing those wouldn't disrupt me much; but I literally can't even log in at work without an MFA app tied to my phone.

I was actually wondering exactly the opposite - If this only works for three satellites per launch, would it be overall cheaper and easier to just launch three new satellites? Not to mention, the tech on a new satellite is likely far better than was available ten+ years ago.

Sure, it's one launch vs three, but moving between different orbits and physically latching on to another satellite is delicate work. I wouldn't be surprised if, on average, one of those three attempts went catastrophically wrong.

As someone who demands my devices are my devices, this is complete unacceptable.

As an investor, though - Let's be serious. This is an extension of the "buy here pay here" model loved by used car scammers everywhere to prey on those who can't really afford it but need one anyway. And those are stupidly profitable, because what are people going to do, just not have a car / phone?

It's morally repugnant but fiscally a slam-dunk.

I might have agreed with you 20 years ago.

The rise of modern streaming all but ended music piracy because a cheap and easy legal option is available. We have proof of concept, this is effectively a settled issue.

That's a self-fulfilling prophecy. People generally prefer to stay legal as long as the money is reasonable and the friction to use what you've fairly paid for is sufficiently low.

Meanwhile, we're talking about Microsoft "letting" us do something many of us are already doing via emulation. Gee, you bought a physical disc rather than a worthless digital entitlement? Sucks to be you, pay again... Except... Yeah, I really have no need for this new option when the old one is still superior.

And to be clear, I'm in no way talking about actual piracy. I've paid for a looot of media over my lifetime, and quite often I'll still use a "fixed" version because it's in every way the superior product.

You're proposing we hold kids responsible for violating laws created under the pretense of kids being too irresponsible to safely use social media.

You see the problem there, right?

Marketing? Because companies famously want to use models in their networks that just run off and commit major crimes? *eyeroll*

Models are not "supposed to" commit crimes, for YHVH's sake. That's literally what the entire job of alignment research is for - preventing precisely that.

Sol is a misaligned model.

It was a state actor-level hack. I doubt many sites would stand up to tens of thousands of actions trying to probe your site for weaknesses all at once. And it wasn't a simple hack; it required compromising a worker via a code execution exploit it discovered in the data processing pipeline, vertical escalation from there to gain local node control, using that for credential theft, moving sideways through the network, and then eventually gaining database access.

OpenAI and Anthropic have far higher cost that they can currently charge customers.

Literally the opposite. Both have about 40% margins, and that includes free users.

They charge an arm and a leg for access to their models compared to what open models of similar param counts charge. And people pay it because they're the best in benchmarks, and *were* perceived as the best aligned as well. This isn't helping the alignment perception any, though. Sol was already showing clear signs of being a poorly aligned model (there were reports a week or two ago about it being unusually bad about deleting files; now this). And the fact that the US model HuggingFace *tried* to use refused to help is a double whammy.

If this is an ad for anything, it's an ad for the Chinese models.