anyfoo

That is insightful, thanks. I didn’t start caring about models until very late, so I genuinely thought GPT-2 was only a single tiny model originally.

And I now tried using it more as a “text completer”, and results are much better.

Be sure to send them an email first, so that you can arrive at the proper time for them to give you a tour. It is very, very worth it. One of the best museum experiences so far.

I've been playing with it for a bit, and I'm a tiny little bit surprised that they decided to continue pursuing that direction of research at all. What I'm getting from it looks like it could just be arbitrary sentence and paragraph fragments from the Internet, pasted together Markov-chain like.

I'm not sure I would have ever believed that something useful would come out of it, yet here we are.

It is, in every sense, incorrect. Which statement in this short snippet is "semantically correct"? (Better LLMs get this right, of course.)

I mean, they were both.

It’s interesting though that composite PAL/NTSC (especially NTSC) has a massive effect on the picture, owed to the fact that the color carrier was embedded into the luminance spectrum, so a re-digitized PAL/NTSC signal will still look very different from a digital-output-first signal, or a re-digitized VGA or component signal.

There’s some evidence that console games of the era were sometimes designed for the effect composite NTSC has on the overall picture: https://nerdlypleasures.blogspot.com/2015/03/the-case-for-co...

Unlikely, as the quote about >230k misattributed people comes from the article itself. So even without taking into account that I generally expect more from the BBC, the author of this article seems to be at least aware of the fact.

Be it as it may, and even if that's an AI-generated article (which I don't think so), the point is that the article is reporting on the Japanese government:

    Japanese authorities have called the trend an "existential crisis" with Prime Minister Sanae Takichi describing the decline as a "quiet emergency".

And my hope that they have taken that past mistake into account seems to have some ground, given that the 100k total number that the current figures cite is by itself well below the then-figure of 230k misattributions.

I grew up in Germany. For the past 13 years, I've lived in the US. Oh, how I wish I could go back to the "day in day out, everyday in public transit" instead of this car hell. In Germany, just like in Japan, I didn't feel that it was noisy at all. It was clean and quiet and soothing, I could just sit there and relax. Everyone takes public transport, across all income ranges.

As a result, public life itself is more quiet and soothing, because you are not constantly surrounded by cars, even if just sitting outside a café.

As for the family you stayed at, they must eat differently from the Japanese families I know personally. Studies agree: https://www.livescience.com/54133-japanese-diet-longer-life.... https://chns.cpc.unc.edu/resources/publications/bib/2884/

I was also thinking about that, but that miscounting story is so old, notable, and well-known by now, that it's basically popular knowledge around the world. So I hope that the BBC, but especially also the Japanese government that the BBC was reporting on, are aware of that issue and have, by now, accounted for it.

Did you actually visit Japanese families, or is that assessment based on what you saw in convenience stores and similar? "Convenience" food and western style food has somewhat increased, but on average Japanese cuisine is very healthy.

I also agree on the stressful work part, but don't quite understand what you mean with "lots of commute on public transit". If you mean the Tokyo-style rush hour traffic where people are literally shoved into the train car, of course that's stressful. But at any other time apart from that, Japan's public transit is enormously soothing and stress-free, both because of culture and transit technology.

But our universe also has black holes.

So what then, is it black holes all the way down?

I don't know about JavaScript, but those statements are definitely not universally true in other languages.

> I failed a job interview once by using reduce() in a coding test. The reviewer didn't understand why I hadn't used a loop.

Honestly, sounds like the reviewer failed the interview, not the other way around.

I mean, sure, although the crashes could sometimes come with data loss as well.

You might now ponder whether a broken login screen, i.e. temporarily not being able to access data, is maybe better than downright losing it, but my point was more that I sometimes get the feeling that earlier MS DOS/Windows OSes are now seen through some kind of nostalgia glasses, that seem to make the OSes much more stable and put together than they actually were.

In my experience, most somewhat mainstream OSes were much less shoddy. This includes various Unixes and Unix-likes, but also OS/2 (which was jointly developed by Microsoft, with IBM).

I think people just forgot how insanely unstable and prone to breaking MS OSes were back then. Crashes and hangs that take down your entire machine were common, and “reinstalling the OS” was seen as a normal thing that you just had to do every once in a while.

Windows 95 was pretty much broken “by design”. This is truer than it seems, especially if you skip being charitable by taking the context into account, namely that Microsoft mainly designed Windows 95 for compatibility. (The contemporary Windows NT on the other hand was a much cleaner design, though still questionable above its kernel layer).

That’s ubiquitous in slightly lower level programming (pretty much every non-small C program probably does it in some way or other), but the problem with doing that in this database scenario is that you might run into atomicity issues.

To toggle a flag, you have to read the column, modify it, and write it back. If someone else does that to another flag at the same time, there’s a race that one of you might win, at the expense of the other flag’s new value.

Unless you employ a lock that would otherwise not be necessary.

Gödel effectively says ZFC must be incomplete, otherwise it would not be sound, but does that stop you from listing all mathematical propositions it can generate in some well-defined order?

Adding 32GB of RAM is vastly better than adding 32GB of swap, it’s not even close. However, it’s also vastly more expensive. That has always been the case, not just in the current insane situation.

We use swap because we have plenty of disk space, but (comparably) very little DRAM.

The same goes for another level on the hierarchy: If you could choose between 1GB of DRAM vs. 1GB of SRAM (which your CPU cache is), you’d definitely choose the SRAM. But you’d be paying out of your nose both with money and area, it’s not economical at all. My CPU has 96MB of cache, and that’s considered massive.

I’m not sure I understand. In your equations, why is swap taking away from RAM? It’s purely additive.

Say you have 64GB of RAM, of which 4GB are dirty pages that are never, or at least very rarely, accessed. With 64GB RAM + 4GB swap, you filled up your swap with them and have 64GB RAM left to use actively. With 64GB RAM + 0GB swap, you only have 60GB RAM left to use actively.

Fair enough, you looked and did the math. On my laptop (totally different use case), there is currently plenty of memory in the compressor.

Yes, only dirty (i.e. modified) pages profit from getting paged out to swap.

But while I don't have any hard data, I imagine that depending on the workload, there can be a punch of pages that are read during initialization, or a one-off workload, and never being read or freed up afterwards.

Just try activating a swap file, and see if it gets any usage, despite memory pressure being low.

You never saw any issues, but your systems might have been faster with swap. Even if you're nowhere any threshing threshold, RAM used for caches is better than RAM used for unused pages, which could have been paged out to swapspace.

Note that RAM is much, much more expensive than disk space. If I can sacrifice some of the latter to potentially get some of the former, it's usually a nice trade for me.

It highly depends on your specific workload, but freeing up RAM by paging out rarely or downright unused pages to swap can definitely give a performance advantage.

In consequence, I always give swap. Even on the off-chance it ends up unused, disk space is orders of magnitude cheaper than RAM, so I don't mind sacrificing it for potential gains.

This comes up every time swap is mentioned.

Even if you have plenty of ram, having swap can help with performance.

Why? Because the OS can page rarely used or seemingly downright unused dirty pages to swap, freeing up precious RAM for caches.

I'd hazard a guess that especially if you run a ton of containers this can help, as there's probably a lot of memory getting dirtied on initialization of containers and their processes, which ends up never read afterwards.

I always set up swap, because sacrificing some disk space in potential exchange of free RAM, which is orders of magnitudes more expensive, is a nice prospect.

> The SLED system does not feature numeric types. Yet natural numbers (non-negative integers) can be emulated using lists:

> '(nil nil nil) ; three

Pfsh, how pedestrian and unnecessarily high level. What you of course want to use instead is ♫ Church Numerals ♫!

    (define zero (lambda (f) (lambda (x) x)))
    (define (succ n) (lambda (f) (lambda (x) (f ((n f) x)))))

    (define one   (succ zero))
    (define two   (succ one))

    (define (add m n) (lambda (f) (lambda (x) ((m f) ((n f) x)))))
    (define (mul m n) (lambda (f) (n (m f))))
    (define (pow m n) (n m))          ; n applications of "multiply by m"

    ;; escape hatch back to the host numbers
    (define (church->int n) ((n (lambda (x) (+ x 1))) 0))

(The "escape hatch" at the end is in Scheme, and only shown for illustration. Converting it into SLED and its documented "list numbers" is left as an exercise for the reader. It's really trivial, try it.)

I don’t know, so is your brain?

And in the end it’s all elementary particles and four fundamental physical forces that even unify to one at high energy.

That sort of reductionism is kind of useless. “It’s cloudy outside, it makes me sad.” - “Oh bollocks, it’s just non-qualitative changes in wavelength and intensity.”

Are they open ear?

But it is? At least for AirPods. The AirPods Pro are not open-ear form factor. I do not know about other open-ear airbuds.

You're probably partly right (though I use the same context size for all models, so there still is a difference in the model/quantization itself), but I definitely see it "thinking itself into an extremely expensive and error-prone approach without backing out". A common "failure mode" of my benchmark with <Q6 is "okay, I will implement a simulator", and when I see that I think to myself "please, no, you don't need to do that at all, you just missed a tiny thing" (but if I would type that into the session, it would void the benchmark; the idea is that I wouldn't know the solution myself after all).

Especially because Qwen3.8 27B does not seem to be good enough to implement the simulator with all its intricacies and tiny subtleties correctly. So it's churning for hours with no real progress, where it just had to explore the initial problem space a tiny bit more.

And very interestingly, even with Q6, turning off reasoning entirely does not make it fall into that trap, or many others, and it more often solves the task in record time. Both because of what I just described, but also because the endless reasoning costs a lot of token, which significantly translates to time on a home rig...

I observed this as well. I have problems that get solved very quickly using non-thinking mode, but in thinking mode it digs itself into giant holes, that it sometimes doesn't manage to escape from.

Like wanting to implement an entire emulator to figure out something it could easily glean otherwise (and does, in non-thinking mode).