You are just butt-hurt not everybody is in love with your new fetish. How pathetic.

Hahahaha, no. But the US military capability is routinely vastly overestimated. The reason for all those bombing campaigns? Because that is essentially the only thing it can really do.

Agreed. (Again? What is happening?)

If you have reasonable worker protections in place, unions are redundant and mostly cause problems. In Europe, worker protection is typically at least reasonable. In the US, it very much is not.

I think a lot of people just fetishize information technology to a degree that they lose all perspective and all sanity.

Lets remove all the bullshit form your statement and state the core, shall we?

AI is a tool. It need to be used competently to produce good results. It is also not the thing that solves every problem.

You are mistaken. You cannot fix LLMs via guardrails, that is fundamentally impossible.

That is a massive problem. If you trade speed for quality, you will build up technological debt. Tech debt is a silent, slow killer. At some point it cannot be fixed anymore, you can just throw away the thing. And that may kill you.

But there are uses where speed is king, like malware creation. We have seen malware created from patches very fast with LLMs. The malware will be low quality code, but that does not matter for malware. Because once you are in the target, you install a (very much not LLM created) root kit to take over.

Indeed. Example: LLMs are capable (with low reliability) to create attack code from patches. The attackers can then try a lot of variations, often even automated. Attack code does not have to be reliable or secure or maintainable, hence speed is a massive advantage. We have seen very fast attacks based on that several times now in the wild.

That unreliability is also why LLMs are not suitable to directly create secure code and LLM code with security requirements needs careful, competent review by a human (!) expert to be usable in production.

The problem with LLMs is that a lot of people are using them for things they are not good for. People are building houses of cards, because they get blinded by the speed and ease of use. I expect that some of these uses will end up killing enterprises.

That said, for all kinds of search uses (like that protein folding), where the results get verified by other means, LLMs can sometimes find other things than previous approaches. Just do not expect LLMs to find things reliably or find all of them. But finding more is a good thing. If you do not lose sight of the limitations, that is.

My impression is that many (most) just use LLMs because of FOMO (which was aggressively created by the LLM profiteers) and to give the appearance of being modern and innovative by following the hype. These are obviously really bad reasons.

Sure, and we are looking for thing that cannot.

We are? We're talking about somebody using the term "quantum material", not looking for evidence supporting quantum physics.

Even so, you should be able to pretty easily notice deviations from Maxwell. It's a pain in the ass with film but pretty well any modern CCD or CMOS sensor should be able to detect shot noise from a double slit, crossed polarizer or just plain old light source experiment at tabletop scale. If not, add an ND filter.

'd be very careful about interpreting QM as "how it works".

Yeah, I didn't say that. It's pretty easy to show that both quantum mechanics and Maxwell are not how it works. Your claim that a polarizer is not a "quantum material" because you can explain it's behaviour with a classical theory (in very limited circumstances) really is assuming "that's how it works" though.

Or Slashdot will tell you you need 10 terabytes of the fastest VRAM and a few dozen of the very latest processors from Nvidia.

There it is!

You don't need any RAM at all. You can read each weight from an SSD, a hard drive or a stone tablet. If it's an SSD it's not necessarily "incredibly slow" if you're just playing around and aren't ready to spend fifty thousand dollars on your chatbot. There are benchmarks for running big models on Raspberry Pis for example.

The 50 B generally means you're going to be able to do things like hold most or all of the matrices in memory while you multiply them, which is usually a pretty big performance increase.

Convenience comes before the rare occasions when it's actually pleasurable to drive a standard day to day. I do not miss a clutch crawling along in traffic or cruising on the highway. And for the in-between bits my car has both a shift by wire stick and paddle shifters if I want to make vroom vroom noises and pretend I'm in F1.

Nah. Learn Chisanbop, Korean finger math, instead. An abacus you'll always have with you. Hopefully.

No, it is not "the slop". People fail in completely different ways than AI and that is why "slop" is reserved for AI. At least by sane people.

Hmm. Maybe that would actually work.

The most likely current scenario is that even without all the damage it does, LLM-based coding is significantly more expensive (when you have to pay the real cost) than competent human engineers doing the same job. Of course, if you add all the defects LLM code has into the mix ...