Honestly, AI technology has gotten good enough that those types of things won't fool it anymore. Captchas take longer for a human than it does for an AI model with vision, and the human is more likely to make a mistake.

The point is, sure, their system may suck right now, but if adversarial patterns became common, they're a model update away from fixing it. Might need better hardware on the device, so at most you'll cost them some money, but it won't get us our privacy back.

There is no such thing as "space", apart from "spacetime.". You can go, "there's a lot of space between our sun and Alpha Centauri" but if you were moving in that direction at close to the speed of light, instead you'd go, "hey, these two stars are super close to each other." Arbitrarily so: you can make them as close as you like, you can put as much energy into changing your frame your reference to make them centimeters or millimeters apart if you want to. Where did all that space go?

You can't really talk about space expansion in the instinctual model we have in our brains as 'space.' That's not "reality" it's a useful model of reality that our brains create to help us accomplish things, but using that to counteract general relativity results *which are testable and have been verified in a variety of ways* is like saying a description of what's going on in CPUs and GPUs don't make sense because it doesn't match up with the video game you're seeing displayed on the monitor. That's an interpretation of states in the computer, in the same way everything you see and experience is an interpretation of reality.

It's a way of reframing sloppy practices as a signal to your investors that your software has become powerful. Most of the horror stories about AI showing signs of becoming hyperintelligent are coming from the companies themselves, which should tell us something.

While the energy source of wind and solar are free, building and maintaining the plants is not free.

If all they were building were a single 100MW data center, sure: renewables would be cheaper to build. But they evidently are planning a major campus that will consume most of the plant's 615 MW output over the next 25 years.

A solar installation, in a favorable location, capable of supplying 600MW around the clock (with battery backup) would have to be roughly 1800MW in capacity. It would be among the largest inthe world, on the order of ten thousand acres in size at a build cost of maybe 2 billion. The battery backup system to ensure 99.9% uptime would be 9x the size of the largest li-ion grid storage system ever built, and set you back on the order of 4-5 billion dollars.

While it's probably cheaper to go renewable than build a *new* nuclear plant, if you can reactivate and one for just two billion that looks like a bargain, if you have a use for all that power. Inability to save money by load following is the financial Achilles' heel of this generation of reactors, but if you have a guaranteed customer for most of your output, years in advance, that's as close to an ideal economic case for them as anything could be.

ToS is the strongest argument, but the distillers are not parties to the ToS. They get their data from data brokers. It is possible that data brokers are violating the ToS, but it would be hard to write ToS that precluded running queries for third parties without creating problems for consultancies and other businesses. Even presuming the ToS could be written to preclude the data brokers doing that, it doesn't affect the resulting model.

But the general shape of the argument brings us right back to unclean hands: we worked hard on this model and it's not fair for you to profit off our work in a way that doesn't have our permission.

I think the fair-minded position is to see what people think when the documentary comes out. You're essentially arguing a negative here -- that there *can't* be anything of value that hasn't been said yet.

Also, I don't think saying something *new* is necessarily critical. Sometimes saying something obvious but in an interesting way is worthwhile.

In this case, Holmes gave access to the filmmaker. Obviously she has an agenda. If the filmmaker is smart, he has an agenda that's different from hers. If he knows his business he'll find something interesting to show you from that conflict.

Now looking at the trailer -- holy shit that woman gives off batshit crazy vibes. That's new to me. I expected her to be slick, persuasive, like Saruman in Lord of the Rings: someone. you'd need real strength of mind to resist being persuaded by. If I'd just handed someone like that my business card, I'd get it back on some pretext then head as fast as I could for the door.

It doesn't make it *right*, but it does make claiming it is *wrong* inconsistent with their own behavior. This could prevent the US companies from suing the Chinese companies seeking an injunction (due to the "unclean hands" doctrine), and probably blocks them from seeking monetary damages in most US jurisdictions.

And suing may undermine the US companies own intellectual property claims by exposing their shaky foundations. An AI model isn't *expression*, so it can't be copyrighted. Insofar as the service allows the underlying model to be deduced through regular usage, trade secret protections don't apply because that's *reverse engineering*, which trade secrets don't prevent.

This leaves violations of terms of service. But setting aside the dubious enforceability of anti-reverse engineering provisions, the Chinese companies may not even be parties to the ToS agreement if they are obtaining the model data through a network of contractors and shell companies.With only moderate paranoia, they can effectively shield themselves from some kind of US tort claim.

Which leaves them with the model they've created from the US company's model, but which the US company has no IP claim upon. If I download a DeepSeek model that's been (hypothetically speaking) trained on Claude, Anthropic might not like that, but what can they do about it?

It's a strange situation. US investors are spending cumulatively on over half a trillion dollars a year in the hope of owning a breakthrough frontier model that is the end of the economic world as we know it. But the essence of the model might not be intellectual property at all.

Think of the categories as fixed, and the people moving *through* them. Over 65s are exiting the category at a lower rate.

I think we have a pretty good idea what the capabilities of the most advanced models are. And they're not that great.

This is a really stupid bill, but the one thing about it that I kind of like is that it's going to put a stop to the AI companies trying to inflate their value by talking about how scary their frontier models are, and how they're "going rogue." None of these AIs are "super intelligent" or "going rogue." What you have is a (admittedly really cool) technology that can solve a lot of problems but also is non-deterministic, so makes a lot of mistakes. And you're giving it the keys to do stuff. And then it misinterprets the instructions and fucks up and does something with those keys that it wasn't supposed to. Then instead of admitting, "oh, we engaged in really bad computer hygiene here," they figure, "we can market this, and say it 'broke free of the sandbox and went rogue" to imply that their models are so advanced it had purpose and the will to act.

So now the old politicians who don't understand technology are going, "oh, that IS scary. We're going to ban it," backfiring on their scary tactics marketing campaign. So it will result in them stopping this stupid scary tactics marketing campaign.

Here's the actual situation: you know every story you hear about how a model deleted the entire codebase at a company? If you have any level of computer literacy at all, the reaction you have is not, "oh, using AI as a coding agent is dangerous." It's, "why the HELL was the agent running with permissions to do that?" And if you hear about the OpenAI's agent broke free of its sandbox using an exploit that it found on the database you set it loose on, the question is, "why the hell could it even run executables outside a whitelist of stuff for it to build". it's amateur hour over here, giving a thing that is non deterministic and capable of fucking up the ability to do do anything at all outside a very narrow scope. The models aren't going rogue, they're fucking up, like when you ask it to give you the emoji for a seahorse and it goes into an infinite loop of bad emojis.

It's more a case of the fox being in charge of the chicken coop; or bank robbers gaining control of the bank.

The job of the press in a democracy is to hold officials up to scrutiny.

The logic is economic autarky -- the belief that we're always better off having complete political sovereignty over every part of our supply chains than having to depend on imports for anything.

This is likely also behind Trump's trade war with Canada. His demands with Canada aren't about discirimatory tariffs -- those were negotiated away with NAFTA then re-negotiated in his first term. His demands are now infringe on Canadian sovereignty, to give US control of Canada's external trade policy, and even on its internal markets (e.g., the use of French in Canadian internal commerce). Symbolically, that's why he wants to rename Lake Ontario to "Lake America". HIs establishment of US control over the Venezuelan oil industry is his greatest foreign policy achievement -- although it's not clear he understands the legal limits on what he can do with their money.

There's a foreign and military policy angle to this too: Taiwan is absolutely critical to the US economy, so we are absolutely committed to defend them if China invades. Semiconductor autarky would mean it's not our problem.

The problem with autarky is that even if it's a good idea (which economists don't believe), you can't conjure it into existence overnight. It will take many years, possibly decades for the US to replace the Taiwan's semiconductor industry. In effect, the administration is proposing to inflict the very damage on the US economy that a Chinese invasion of Taiwan would inflict.

And, once the US has voluntarily self-inflicted that damage, China would be free to invade Taiwan.

Investors are chasing a company that achieves some kind of AI singularity. Let's set aside the fact that there's no reason to believe there is anything but diminishing marginal returns by making marginal refinements to current frontier models. Let's imagine someone hits the jackpot and gets, not even AGI, but a system that's as far ahead of today's frontier model are ahead of 2020's GPT 2.0 in performance.

Globally AI revenues are 150 billion, against a cumulative burn rate of 450 billion. A model that is a generation ahead of others would almost certainly capture the lion's share of that revenue.

If AGI magically appears as a Sam Altman has promised investors it will, a hundred million is way too low. Add, maybe, another zero to the revenues.

Conservatively, a safer assumption is that frontier models will get marginally better based on refinements in training and reinforcement and the other bits and bobs that go into these systems. The nobody is winning the lion's share of anything, at least overnight. But you have to define "safe". By "safe" I mean unlikely to lose money. But some investors are clearly defining "safe" as "having the greatest chance of owning a piece of the biggest thing ever."

I'm not following this super-closely, but if Anthropic is pursuing adding multi-step model based reasoning to their system, that could be the basis of a generational leap in capability. But if that is an approach that looks like it has a chance of working, then their competitors are no doubt pursuing the same thing. In that case you'd expect the revenue pie to grow as the scope of model utilty increases, but that growth to be split among several competitors. This could credibly result in a revenue stream for some of them that is as big as the entire industry's revenue stream today. But there's going to be hell to pay on the data center impacts end of things.

Ok, but some of us actually want to do something useful instead of just try to win the trophy of making a model run on the cheapest hardware. Unless you're generating at least ~30 tokens / second, it's not useful.

If you want to split the cost evenly, sure. But a team of 3 people working at a software company today typically pays the AI companies about $5,000-$10,000 a month, so I don't think they need to rely on the parts of the world where $400 is twice their income.