LarMachinarum

another reason why I find myself often using LLMs instead of classical search engines is the possibility to obtain structured data and format the output so as to match my use case, e.g. as markdown table, or as json file etc.

having had quite a bunch of MiniPCs, mostly from reputable brands (Intel NUC series back when Intel had those, then Gigabyte Brix series, then some cheapo china ones), I have moved away from those, because every single one of them (independently of the brand) ended up dying spuriously not long after warranty end and in any case far sooner than any µATX desktop would (in fact I've very rarely had any of the latter die; they usually live far beyond their phase out / replacement)

Even without wanting to attribute that to any malicious planned obsolescence, my impression is that the very small size of mini PCs makes it almost impossible for the manufacturer to ensure proper thermal management for keeping all components constantly at a temperature low enough for device longevity.

Do buybacks cause profits to increase?

They cause each (remaining) shareholder's part of the profits to increase (relative and absolute), because there are fewer shareholders left. So it does very really increase the value of each share.

then again, if the makers of one big browser (and via there also the derived browsers) start force-shoving spyware upon you (by restricting blockers), it comes down to a decision of how you set your priorities. Personally, It's a clear cut red line, but you do you.

… or you can instead phase out those browsers who try to force blocker restrictions i.e. spyware on you (e.g. chrome and such), and use one of the browsers where you can use the full-featured (not "lite") uBlock Origin instead, e.g. Firefox.

I wouldn't count on either to save Intel as it still is (i.e with the fab business still attached to the CPU/GPU business). While it's true that having Intel fabs as a second source would be nice for them to alleviate the dependency on TSMC, they are also competing with Intel on the CPU/GPU side.

My guess is, they're gonna let Intel rot a little further while doing their best to pressure for Intel to split off their fab biz (as AMD had done back then), and then invest just in the fab.

all with you about the fact that it's even more funny just in German (i.e. without subtitles)…

but well, given that people understanding German are likely a minority here, I felt it was probably necessary to link to a version with subtitles so that it would be accessible to everyone here.

Anyways, here's the link to the vid without subtitles for those who understand German: https://www.youtube.com/watch?v=4lSBwF8ojVw

Am I the only one who, after reading the title, had to think of the funny classic "Forklift driver Klaus" short vid (which, with its very own kind of humour, shows WHY forklifts require training)?

* For those who wouldn't know it: https://www.youtube.com/watch?v=TJYOkZz6Dck

and anyone with an ounce of curiousity would be like, ".........wait a minute"

I see what you did there ;)

for info: "Claudia" as a name happens to be already in use by another open source project: Claudia the audio session management tool (for the JACK/LADISH APIs i.e. typically on Linux), part of the KxStudio audio tools suite by Filipe Coelho (falkTX) and the KxStudio community: https://kx.studio/Applications:Claudia

that could indeed be an option depending on your use case. The problem with those (aside from finding a suitable enclosure) is that while more powerful, they aren't at all optimized for use in really low-power conditions and that their energy consumption is consequently enormously higher than that of e.g. an nRF52 (e.g. nRF52832 or nRF52840), so that the battery time would likely be significantly shorter.

didn't try the official Firefox mobile browser, but the page is displayed and usable just fine in the Firefox-based Iceraven browser that I use on my android tablet.

hmm… and it's not for scanning lasers either

don't worry, human, that will be corrected soon enough. You should learn to welcome your new AI overlords. Ask your regular handler LLM for advice on how best to do that.

Sent from my iMPC

Aside of how this looks like an AI-generated PR puff piece, it comes waaay too late after the Windsurf deal busted (OpenAI refusing to buy them after all) and Google cherry-picked the employees they wanted without wanting to buy the company, the artificial ARR-blown valuation disappearing in thin air and investors losing pretty much everything.

The thing is: those "$100M+" of negative-margin ARR don't mean anything positive for Cursor, just like it didn't for Windsurf; quite to the contrary, it's just a measure of capital bleed with an unsustainable model in a negative margin death spiral: the more of that artificial ARR they make, the closer they get to bust.

For a description of the problem, see https://ethanding.substack.com/p/windsurf-gets-margin-called which was discussed here: https://news.ycombinator.com/item?id=44843801

If that's the reason, I kinda wonder what kind of numbers we're talking about in this case.

Much as I understand how a 5 bit quantization might be a sweet spot in the tradeoff between precision and making it possible to cram more weight parameters into limited ram, and thus in that respect better than e.g. 4 bit or 8 bit,…

…I struggle to comprehend how an odd quantization like 5 bit, that doesn't align well with 8 bit boundaries, would not slow things down for inference: given that on one hand the hardware doing the multiplications doesn't support vectors of 5 bit values but needs repacking to 8 bit before multiplication, and on the other hand the weights can't be bulk-repacked to 8 bit once and for all in advance (otherwise it wouldn't fit inside the RAM, besides in that case one would use a 8 bit quantization anyways)

it would require quite a lot of instructions per multiplication (way more than for 4 bit quantization where the alignment match simplifies things) to ad-hoc repack the 5 bit values to vectors of 8 bit. So i kinda wonder how much (percentage-wise) that would impact inference performance

All with you on the general idea that growing potatoes is easy and great. That being said, just a detail:

People grow them in very poor soil, and even in literal bags of compost

a bag of compost is pretty much the exact opposite of "very poor soil"; it's about the richest soil there is.

sure about that? maybe he'll be the very risk-inclined, hot-headed, reckless platoon sergeant who will have a higher chance of earning a medal but also a higher chance of getting you and most of the platoon killed.

The Colmi P8, which is older, long replaced by myriads of newer china watches and now hard to even find, was one of the last cheap smartwatches to be based on the nRF52832 microcontroller/SOC which had the advantage (for that purpose) of being both well documented and yet not locked down. The successor SOC, the nRF52840, already had a flash securing feature that (except for devices that wouldn't use it or that would have exploitable vulns) made it easy for the manufacturer to lock the device down and to prevent the install of alternative firmwares. Also about that time, cheaper chinese SOCs came out and cheapo china smartwatches switched to using those instead of nRF. Trouble being: most of those chinese SOCs for smartwatches, aside from probably also having the lockdown problem, don't have much in terms of openly accessible documentation or developer tools.

Consequently, pretty much all open source projects for cheapo china smartwatches apparently only support devices that are so old that you don't even find them anymore on aliexpress or other such shops.

I'd be interested to know for what currently easily available cheap (i.e. not in a much higher price category) china smart watches there is an open source alternative firmware that does not miss half of the features.

Qt used to be what I used as well, but the licensing has become so repulsive for professional use that I now refuse to touch it now (not even for my open source side projects where this wouldn't be an issue).

Too bad the hopes from the early Nokia adoption days got smashed by MS mole Elop and the later owners of Qt.

In Italy it's even worse, it can even be made of rock, dirt and lava.

I could go into a place, order a pizza, sit there starting at it for 15 minutes, and then report them for a violation

it's just the other way around: the pizzeria's duty is just to serve you a fresh pizza, but if you haven't finished eating it within 10 minutes, THEY can report YOU for violation and the pizza police comes for YOU. This ensures that customers don't hog the table (which is bad for business).

while none of the arguments of the article came even close to being convincing or to balancing out the disadvantages of a non-power-of-two orientation, there actually is one totally different argument/domain where the 9 bit per byte thing would hold true, that is: ECC bits in consumer devices (as opposed to just on servers):

The fact that Intel managed to push their shitty market segmentation strategy of only even supporting ECC RAM on servers has rather nefarious and long-lasting consequences.

fun fact (or not so fun): incidentally, AFD (or more precisely written AfD) is also the name of a radical far-right party in Germany. The party has progressively shifted to the fringe and is now classified as "confirmed right-wing extremist" by the German Federal Office for the Protection of the Constitution (Verfassungsschutz), Germany's main domestic intelligence agency. Ref:

* https://en.wikipedia.org/wiki/Alternative_for_Germany

* https://en.wikipedia.org/wiki/Federal_Office_for_the_Protect...

It's interesting to see which prominent figures of tech and politics in the USA have publicly associated with the AfD.

while I don't doubt that such situations also exist, that wasn't the reason for any of the many "Excel abusers" I've encountered in different positions. Quite to the contrary, these people all had access to the appropriate tools, but their whole thinking was totally formatted and fixated on Excel as their go-to tool for everything:

be it things better done with a database, a word processor, a diagramming program, a label generator, a Form editor, a markup language, a web page, anything: they had all the tools at their disposal but no, no, they felt the odd compulsive need to do it with only Excel instead…

…often leading to problems down the line when the limitations of Excel for the use case (for which it wasn't made) would show more and more but they wasted already so much time and (needless) effort doing it in Excel that they would be even more reluctant to the possibility of switching to any more appropriate tool for the task.

Many here point out, and rightly so, what an abominable mess UI frameworks have been on Windows for quite a while.

That being said, I'd point out that the situation has been (and still is) unfortunately far from satisfying on the cross-platform and open-source side as well: GTK has also been a mess for quite long; Qt has - aside from the bloat - an increasingly repulsive licensing model for professional use (too bad the hopes from the early Nokia adoption days got smashed by MS mole Elop and the later owners of Qt), …

… and while there are some nice solutions for specific niches (e.g. Dear Imgui), the overall situation in general-purpose cross-platform native UI frameworks looks quite bleak and unsatisfying imho. There's a huge gaping opening for a permissive-licensed open-source cross-platform UI/Widgets framework that compiles to native, has a good set of widgets and gives access to Vulkan contexts for 3D rendering.

when I read "1m-accurate indoor GPS", my first gut reaction was along the lines of "so what? Nowadays outdoor differential/rover GPS/GNSS has <2cm precision with RTK+dual band receivers, even for civil use… and indoor allows for yet way better parameter control than outdoors, so it should be possible to achieve at least that precision" (besides, it would allow for robots to get a sufficiently precise location for indoor navigation and to shelve wares)

The thing is: that would require additional hardware though, whereas the relevant limiting criterion of the project presented in the article is obviously that it has to run with just the hardware provided by common mobile phones. Which is what makes the project interesting despite the comparatively lower precision.

In all fairness, that really stands to reason: retrospective analysis is enormously much easier, more precise and less risky than prediction, for the simple reason that after the fact (especially this long and with lots of inside people having spoken up since), there is enormously more data available than before it went downhill. So it's much less a question of having guts than of having access to sufficient data to see enough of the whole picture.

How much is gained by using SVG (as opposed to a raster graphics format) varies a lot depending on the content. For some files (even with complex shape paths depending on a couple details) it can be an enormous gain, and for some files it can indeed be disappointing.

That being said, while raw SVG suffers in that respect from the verbosity of the format (being XML-based and designed so as to be humanly readable and editable as text), it would be unfair to compare, for the purpose of HTTP transmission, the size of the raster format image (heavily compressed) with the size of the SVG file (uncompressed) as one would if it were for desktop use. SVG tends to lend itself very well to compressed transmission, even with high-performance compression algorithms like brotli (which is supported by all relevant browsers and lots of HTTP servers), and you can use pre-compressed files (e.g. for nginx with the module ngx_brotli) so that the server doesn't have to handle compression ad hoc.