You're a little bit ahead of things here, considering that is not slated for completion until 2040.

https://www.nationalobserver.c...

This will be the nail in the coffin for coal plants, and probably a fair number of natural gas plants as well. Once this AI bubble bursts, coupled with the ability to process AI demands much more efficiently (on silicone designed specifically for this purpose, instead of using much less efficient GPUs and the like), there will be a huge surplus of power. The shiny new, and extremely expensive to build, nuclear power plants certainly won't be the things shutting down when there is too much energy.

The environment will be much better off for this.

You must be new to this. This kind of sentiment has existed since computers have become prominent around the globe.

10 years from now we'll be hurting, and not because AI replaced humans in so many roles. We'll be in bad shape because it didn't replace humans in various areas we expect it to. In addition to that we'll have to deal with cleaning up the messes from, and maintaining the crap that was spat out by, various AI models.

I'll give you a perfect example of something AI screwed up without actually having done anything whatsoever. Starting about 10 years ago, and peaking about 8 years ago, it was all over the various news and tech headlines that AI image recognition had gotten so good that it could read various X-ray, CT, MRI, etc scans and detect various problems. That it would soon replace radiologists. So guess what happened? Some small percentage of medical students considering the field of Radiology chose something else. Now, nearly 10 years later, we have a serious shortage of radiologists, as there has been a deficit between those retiring and the new radiologists finishing up school. It's getting worse, and will continue to get worse. Now we actually *need* AI to do what was claimed, and help read images so that the radiologists that are still working can be more efficient.

We're going to see this exact same thing in many other fields, as young adults avoid various careers most threatened by AI. Then there will be a shortage several years from now when we truly realizes the limitations of (and probably more important, the legal liabilities resulting from) AI.

Just a few articles going back 5+ years regarding radiology:
https://subtlemedical.com/ai-i...
https://www.healthcarefinancen...
https://www.medtechdive.com/ne...

Why does this remind me of TempleOS in some way? I have a very strong suspicion that if Terry Davis was still alive that he'd be all over AI in this way.

Business have liabilities - legal, civil, on and on. AI will be awesome at saving businesses money by replacing employees - until it results in a massive liability that costs them far more money than it saved them. Especially if the court systems have no sympathy for these kinds of business practices and don't cut them slack when it happens.

We've already heard of what is just the tip of the iceberg, where support chatbots hallucinate things, like telling a customer they will get a full refund for their car because it isn't running right. Once this kind of stuff gets worked out in court, that these kinds of things are legally binding, you better believe a lot of businesses will be very, very afraid of using AI.

I have two questions. One, why would this not be instantaneous? I thought the point of quantum computing was that all states were visited in parallel, with it collapsing on the final state pretty much instantly. You set up the starting state, and it collapses into the result. At least that's what I've always heard.

The second, is how do you know when you've successfully decrypted the data? What if you end up with data that looks correct (like a credit card number, or a valid sentence that even makes sense), and it produces the correct checksum (because there are many, many collisions in checksums), but it is not the original data? So does this produce every valid output (valid via checksum or other method) for the input, and thus you still then have to brute-force test each candidate?

Linters and other tools to analyze code have taken over this role. Back when Java came out the only option was baking it into the language itself, which I agree is too rigid.

By accumulating the energy until it reaches a threshold where it is emitted. Like most everything else that can be "charged" by light and release it by glowing later.

This isn't converting the wavelengths of IR to some other wavelength like a filter. It just kind of soaks up IR light and glows it out as visible light. And even then it is taking an extremely strong source of IR light from an LED to work. You know how you can close your eyelids and still see a bright light like the sun glowing non-distinctly? It would be a lot like that. The glow is so faint that it really can't be seen in a bright light environment, which is why it works better with eyes closed (again, with a bright spotlight-like source of IR beamed at the eye).

Neat, but this isn't going to turn you in the the Predator seeing in IR or something.

It means the other 17 hours it produces unusable nonsense that superficially looks correct that a human then has to spend 40 hours sorting out and fixing.

You must not have a lot of experience with unions to think there's anything odd about this behavior. It's pretty well expected for them to pull something like this. It's not about protecting the artist, it's about getting their cut and maintaining their control.

Like executing extra code to check for various memory safety rules?

It's not about any controversial topic or language. Sure, the paper may sound totally reasonable, but is the data real or fabricated? Is the methodology sound? It sounds like it wasn't.

I'm not an AI expert, but I believe this pathway of creating a monolithic AI that is in and of itself the reasoning, knowledge and computation is the wrong way to go.
Stephen Wolfram (of Mathematica / Wolfram Alpha fame) had insight that LLMs would be best served as an interface between user queries and the knowledge / computation engines, instead of attempting to be the knowledge and computation engines themselves.

An analogy is asking a person to add up 100 numbers in their head as you call them out, versus asking a person to use a calculator to add them up as you call them out. The calculator is pretty much infallible in this regard, as that is what it is designed to do, while a neural net may decide 1 + 1 isn't actually 2 on occasion (partially due to temperature for one thing). It would also require drastically less computing power, as the calculator is optimized to do that task, while a neural net has to build up a bigger and bigger token memory to hold the contents of those calculations in a much slower and inefficient way.

They are already heading in this direction, for example you can see LLMs are now actively accessing the internet seeking out relevant information that is the source of what they produce, as opposed to that information having been cooked into the LLM's weights during training.

Great, this will make it much harder for us to stop Skynet.