How do you measure these improvements in a quantifiable manner to gage whether or not they are plateauing?

All pure speculation that improvements are exponential, just like it's pure speculation that "intelligence" is some kind of emergent property of an LLM?

I love all the YouTube clickbait videos titles "these jobs won't exist in 24 months". Yeah, right. If an entire class of job was going to "not exist" in 24 months it would be painfully obvious as businesses began testing the alternative in advance of eliminations. Which of course would not happen like flipping a switch, we would be seeing 10% or more losses now, with measurable increases every month.

LLMs excel at generating large volumes of text that appears moderately plausible and coherent. This is literally the job of every journalist on the planet. So, naturally, journalists think LLMs are going to do all the things and they write about it.

And they can't do that reliably or repeatably. They can't do anything reliably or repeatably, at anywhere near the level necessary to replace a human.

I've recently been using Chatgpt for assistance with typescript and css for a work SPA project. One time I asked it if a very small function - only 5 or 6 lines - was the most efficient, idiomatic way to do a thing. It gave me a response, then asked if I wanted an even "more efficient" way - which, duh, of course I do, that's what I asked. The "more efficient" way was a literal copy of my original function.

I despise CSS so much that I'm sure I will continue to use LLMs to tell me why a layout isn't working, but it still takes multiple iterations to get it right.

(Crazy that we are three decades into this whole web experiment and the mess of css is still the "best" thing we have. But why fix it when AI can write it for you?)

None of the "models" can reliably and repeatably do what they are claimed to do. Humans are still the only reliable choice, and there is absolutely no evidence that this will not be the case for the foreseeable future.

Several of the Walmarts in my area have similar "gates". It's clearly just to stop people from running back out the entrance with stolen merchandise. If you need to get out just walk around past the cashiers and leave.

I think we have a decade of empirical evidence that the complexity of the issues facing modern society has exceeded the ability of the average person to understand. In the past, you could just go somewhere else and "farm", not worrying about others. That world no longer exists.

And CHA of 0.

Which is why refresh rates of 120 and 240 are "correct" - both are multiples of 24, 30, and 60 - the most common framerates in NTSC world. Instead of interpolation frames they are just duplicated. That's how it was done before someone thought they could sell you a new TV with "motion enhancement".

Kodi changes the refresh rate of the TV to match the content, which is probably the lowest energy solution.

Hires people for three months, from what I can tell. I also can't remember when I didn't hear about a 10k+ sized layoff at IBM.

But this is a much, much better decision than spending tens of billions of dollars building it yourself. Apple is where the eyeballs are, and everyone is going to want to plug their models and agents into iOS. OpenAI's "device" will probably never be anything more than vaporware, or at best a total flop. In the end, all roads lead to the iPhone and Android.

No, as I said, quantum mechanics is the best explanation we have, and it has yet to be contradicted.

At the quantum scale, there are truly random events. But all events are governed by probabilities, which just means that nothing is 100% certain. A glass of water could spontaneously boil at room temperature, but it's so improbable that it won't happen within the lifetime of the universe.

If you believe everything they say about AI, it won't need an OS at all, and one else will either. It will just inhabit any silicon of any kind, everywhere, all at once.

And in a few hours, the models will have ingested these comments and be able to "modify" their "behavior" a little bit more!

And all of those sound like "solutions" that would be in the training dataset known as "linux.rootkit.ru.2025".