I mean, I *have* checks. Been years since I actually wrote one.

The funny thing is the 'bank' on the checks is like 3 mergers ago.

6 m/kwh is *significantly* better than most cars. Model 3 is ~4 right? 50% better.

I really want it to succeed, but it's way beyond believe it when we see it territory too. Vs a Model 3 that exists and is on the road.

One major difference, assuming you've got full platform support(should be the case on any server or workstation that isn't an utter joke; but can be a problem with some desktop boards that 'support' ECC in the sense that AMD didn't laser it off the way Intel does; but don't really care); is that ECC RAM can (and should) report even correctable errors; so you get considerably more warning than you do with non-ECC RAM.

If you pay no attention to error reports ECC or non-ECC are both rolling the dice; though ECC has better odds; but 'proper' ECC and Linux-EDAC support will allow you to keep an eye on worrisome events(normally with something like rasdaemon, not sure what other options and preferences there are in terms of aggregating the kernel-provided data) and, unless the RAM fails particularly dramatically and thoroughly, will give you much better odds of knowing that you have a hardware problem while that problem is still at correctable levels; so you can take appropriate action(either replacement, or on the really fancy server systems, some 'chipkill'-like arrangement where the specific piece of DRAM that is failing gets cut out of use when deeemed unreliable without having to bring the system down.

Sometimes you'd get a BSOD that was a fairly clear call to action; when the error called out something recognizable as the name of part of a driver; but that is mostly just a special case of the "did you change any hardware or update any drivers recently?" troubleshooting steps that people have been doing more or less blind since forever; admittedly slightly more helpful in cases where as far as you know the answer to those questions is 'no'; but windows update did slip you a driver update; or a change in OS behavior means that a driver that used to work is now troublesome.

Realistically, as long as the OS provides suitable support for being configured to collect actual crash dump material if you want it; it's hard to object too strongly to the idea that just rebooting fairly quickly is probably the better choice vs. trying to make the BSOD a genuinely useful debugging resource; especially given how rare it is for the person with useful debugging ability to happen to be at the console at the time of crash(rather than just an end user who is ill equipped to make sense of it; or a system that mostly does server stuff, quite likely not on actual physical hardware, where nobody has even touched the physical console in months or years; and it's more or less entirely useless to display a message there; rather than rebooting and hoping that things come up enough that management software can grab the dump files; or giving up and leaving the system in EMS so that someone can attach to that console.

Win9x and Win2k (and the other NT descendants) are fundamentally different operating systems. In general, NT had a much more robust kernel, so system panics were and remain mainly hardware issues, or, particularly in the old days, dodgy drivers (which is just another form of hardware issue). I've seen plenty of panics on *nix systems and Windows systems, and I'd say probably 90-95% were all hardware failures, mainly RAM, but on a few occasions something wrong with the CPU itself or with other critical hardware like storage device hardware. There were quite a few very iffy IDE cards back in the day.

The other category of failure, various kinds of memory overruns, have all but disappeared now as memory management, both on the silicon and in kernels, have radically improved. So I'd say these are pretty much extinct, except maybe in some very edge cases, where I'd argue someone is disabling protections or breaking rules to eke out some imagined extra benefit.

We should ensure his fat ass is hauled in one in the not soon enough state funeral

Yeah, it's not even worth considering for something like 15-20kg. A full pallet in this case is 464kg :) Years worth of supply for me.

The current "AI" is a predictive engine.

And *you* are a predictive engine as well; prediction is where the error metric for learning comes from. (I removed the word "search" from both because neither work by "search". Neither you nor LLMs are databases)

It looks at something and analyzes what it thinks the result should be.

And that's not AI why?

AI is, and has always been, the field of tasks that are traditionally hard for computers but easy for humans. There is no question that these are a massive leap forward in AI, as it has always been defined.

It is absolutely crazy that we are all very very soon going to lose access to electricity

Calm down. Total AI power consumption (all forms of AL, both training and inference) for 2025 will be in the ballpark of 50-60TWh. Video gaming consumes about 350TWh/year, and growing. The world consumes ~25000 TWh/yr in electricity. And electricity is only 1/5th of global energy consumption.

AI datacentres are certainly a big deal to the local grid where they're located - in the same way that any major industry is a big deal where it's located. But "big at a local scale" is not the same thing as "big at a global scale." Just across the fjord from me there's an aluminum smelter that uses half a gigawatt of power. Such is industry.

That "ruler study" was ancient. It's mentioned in peer review at least as early as 2018, and might be even older.

Believe it or not, people in the field are familiar with these sorts of things that you just read about.

Most of these new AI tools have gained their new levels of performance by incorporating Transformers in some form or another, in part or in whole. Transformers is the backend of LLMs.

Even in cases where Transformers isn't used these days, often it's imitated. For example, the top leaderboards in vision models are a mix of ViTs (Vision Transformers) and hybrids (CNN + transformers), but there are still some "pure CNNs" that are high up. But the best performing "pure CNNs" these days use techniques modeled after what Transformers is doing, e.g. filtering data with an equivalent of attention and the like.

The simple fact is that what enabled LLMs is enabling most of this other stuff too.

Jevons Intensifies

I agree with the sentiment, and hate the idea of putting my data and functions in somebody else's hands. Let us pause for a moment and remember *** Cloudflare ***

But why does it have to be Somebody Else's Server? This is what the Raspberry Pi Pico was born for...

I've done my first test of buying a whole pallet of filament straight from a Chinese manufacturer. It's a risk - it could be all junk - but if it's usable, the price advantage is insane. Like $3/kg for PETG at the factory gate (like $5/kg after sea freight and our 24% VAT). Versus local stores which sell for like $30/kg.

I'd love to see someone try to 3d print with a filament that melts at 162K. Where do you even buy xenon filament? ;)