JPEG XL

What the heck xd

Which phone?

But 50MP DNG from stock camera is 2+ times bigger than 200MP DNG in Camera Raw...

S25 Ultra

Hm alright

not for S23U

Not, that I'd need it xd

Hmm, expert raw saves JXL DNGs right?

Yes

But debayered

so linear JXL DNG

~~linear DNG isn't DNG~~

Nothing special, just quick pics of my desk in a dark room <:KekDog:884736660376535040>

Expert Raw used to have this bug where the JXL DNGs were extremely lossy, I am talking 1-2MB for a 12MP DNG lol

But yes, expert raw does seem to use JXL

``` Image Width : 4000 Image Height : 3000 Bits Per Sample : 16 16 16 Compression : JPEG XL Photometric Interpretation : Linear Raw ```

From an Expert Raw DNG

My non-linear lossless 12 MP DNG is 3 MB

https://embed.moe/auto.gif?q=https%3A%2F%2Fcdn.discordapp.com%2Fattachments%2F806898911091753051%2F1344402464136233130%2FTinyDNGLossless.jxl%3Fex%3D67c0c7dd%26is%3D67bf765d%26hm%3D77205b519b70fb13bcfb57f7e8043e15154967e0926e3c7775d13feaeba1db9f%26

I haven't tested extensively, but linear JXL DNGs are normally much larger than non-linear JXL DNGs

And Expert Raw outputs linear

Heh, never noticed higher res uses a lower amount of bits per sample

16 for 12MP, but 14 for higher

They also fixed it, so now the DNGs are much larger

Doesn't matter anyway, since the sensors used by Samsung are all 10 bit as far as I know

IIRC mine reported as 16bit in ExifTool too, but the actual JXL is the correct bitdepth

You can check your sensor's bitdepth using this app https://www.celsoazevedo.com/files/android/google-camera/dev-bsg/camera2test/

Wonder if disabling certain processing options will stop high MP images from being bugged

What are you refering to?

"Expert RAW Labs" options

Hm, that seems to be a nope. Still the same size

Guess it's only an option for 24MP and 12MP for some reason

Samsung is not the best regarding camera software anyhow. My phone won't save a DNG if I don't switch from Pro mode to a different mode and back before taking a photo

When cold starting the camera app*

10-bit with an 8-bit pipeline. Though I did spot a 'Depth' output with a resolution of around 1K, probably from the dedicated sensor on the back along with 'Private' which will be the greyscale sensor that's not registered to the OS properly

Check the white level category, the value there will be the bitdepth

1023 is 10 bits

Yeah, it's 10bit

Most phone sensors seem to be 10bit

I have an arducam 12 or 16MP module lying around somewhere, I should check it that is 10bit too

As far as I understand, iPhone cameras don't really have a useful CFA raw image since they use multiple sensors and exposure stacking and lens corrections and a bunch of signal processing to produce a final image, so the real raw data is just too messy and all they can really produce is a high bit depth RGB image.

So iPhone ProRAW is basically just a normal image but with out-of-gamut / out-of-range data preserved and higher bit depth and lossless compression?

Like you can take a “RAW” night mode photo on iPhone, even though Night mode implies heavy processing

well "RAW" is not really pixel data, it's sensor data, and iPhones don't show you the sensor data

In the purest sense, raw is completely unprocessed sensor data, but it has become common practice to do at least some amount of processing even when producing "raw", e.g. to take care of defective subpixels.

In the broadest sense, "raw" is anything that can still be edited in a postproduction workflow, as opposed to delivery formats where the precision and fidelity is reduced to a point that suffices for consumption but not for production.

https://youtu.be/160F8F8mXlo

(it seems it’s mainly the beginning that’s interesting)

that was a good video

the rest is mostly about philosophy, most of their videos are

I liked this one too

https://www.youtube.com/watch?v=wS7IPxLZrR4

https://fixupx.com/litteralyme0_/status/1895181102719082546

Reminds me, I've seen a lot of washed out videos on Telegram lately. Wonder if their HDR handling changed

No more Skype in May https://x.com/Skype/status/1895477868261412953

But what about Skype for business 🤔

also Microsoft Teams, I guess

https://youtu.be/XRPUoz1TYro

hooray

why are you paying for HRT with crypto

Because it's what the grey market sellers accept. I don't live in the US where you can just show up at Planned Parenthood and ask for informed consent HRT.

https://youtu.be/qXrn4MqY1Wo

I read this as “It takes years to build Rust”

LMAO

that's so true, if you're working on a Bevy project and don't use their recommended tricks to speed up compile times, one might be tempted to quickly run the project after saving some changes, and then halfway though the build you realize you need to change some values real quick so you need to wait half a minute again

IME, the slowest part of incremental rebuilds in Rust is linking. Moving my projects folder from spinning rust to an SSD sped up rebuilds by like 30x.

interesting. linking was also definitely the slowest part for me, but i was already using an SDD. changing to the experimental `mold` linker reduced iterative build times from 40 seconds to like 4 seconds (for Bevy projects)

It takes years to build jxl-rs

I was just clicking around on Wikipedia, and came across this again. I remember being very interested in it as a child, after watching some youtube video about lost technologies https://en.wikipedia.org/wiki/Sloot_Digital_Coding_System

OpenStreetMap PBF files have a *lot* of varint encoded integers. They use 7 bits of data and 1 continuation bit in between. Turns out you can read those branchless: https://gist.github.com/jjrv/b9cc022faec13109a99691b2c38ceeff It could be easily adapted to read JPEG XL's 8 bits of data between continuation bits, but it only supports encoding / decoding 56-bit integers when using 64-bit registers. Would need a bit more thinking to support more bits. I guess varints aren't common enough in jxl files to make that very relevant.

Only used in some headers and file format stuff, iirc. So speed doesn't matter.

Does anyone have Computer or VM with Windows 7 or Windows 8 or 8.1 ? I need someone to check if a portable build of nomacs starts there: https://github.com/nomacs/nomacs/issues/1267#issuecomment-2694257078

https://www.youtube.com/watch?v=BBuvc_K8tDg

I can try this now

It works on Windows 7 Professional

in libjxl, we’ve started to link the MS VC++ runtime statically to avoid this sort of issue

<@886264098298413078> does it have a custom Qt6 or something?

(actually, more insidious issues where it would seemingly start and run but crash in mysterious ways because the installed runtime was too old)

Thank you!

It uses DLLs from https://github.com/crystalidea/qt6windows7/

oh cool, I might use that.

if any video editor enthusiasts here, olive update https://youtu.be/invMlMRPUrM

<@1028567873007927297> fair i guess, I basically assume the backbone of my system is a bunch of bug-ridden C spaghetti regardless <:KekDog:805390049033191445>

I wish the system backbone was pared down and attacked with dikes more often, rather than being allowed to grow into monstrous piles of spaghetti.

https://www.youtube.com/watch?v=NgLW-E22vOM

https://www.reddit.com/r/europe/comments/1j4014j/the_dutch_public_broadcaster_made_a_sketch_on_the/

NPO my beloved

core dev pov

It's beautiful

High fidelity

That surprising moment when Discord can play a video that MPV can't handle <:KekDog:884736660376535040>

4k@60 high bitrate 10b HDR HEVC test video

Yes, for 4k ≥24FPS it is generally better to enable hwaccel so it plays properly, else it will be a stuttery mess

for me mpv plays it perfectly fine with software or hardware decoding. discord doesn't play it (on Linux, it doesn't support HEVC) 😁 9950x and 4080S

Oh my god WHAT Discord actually engaged my phone's non-standard HDR mode. That's the first time anything other than YouTube has worked

non standard HDR mode?

It doesn't use the normal Android API, so requires specific integration

Wait a minute, discord natively embedding a video over 450MB?!

Did they increase the embed limits?

I thought it was around 50 MB

IIRC videos have no embed limit, images are around 25 MB

probably because videos can be streamed in while with images the whole thing is just one frame/thing

I wonder if you where to upload a video in which the generated thumbnail ends up going above 25MB that it would fail to embed?

working fine here in mpv

@l

I use software decoding for mpv, it's laggy but functional, hw decoding works fine

Nice, guess I'll just blame my 9900K and sub optimal mpv defaults then. Weirdly enough I never had issues with HEVC playback though, but guess the test videos went just over what my CPU can handle (it works fine wirh hardware decoding in Chrome) <:KekDog:884736660376535040>

Think they improved their media handling by a lot recently. Support for new formats and higher embed limits

Someone needs to add jxl support to https://github.com/discord/lilliput

Wonder if they like pull requests

They recently added avif

I assume mpv uses more computationally expensive tonemapping?

and gamut reduction

Mpv uses software decoding by default.

wait really?

Yeah

ya, it has `--hwdec=no` by default

From their docs `Hardware decoding is not enabled by default, to keep the out-of-the-box configuration as reliable as possible. However, when using modern hardware, hardware video decoding should work correctly, offering reduced CPU usage, and possibly lower power consumption. On older systems, it may be necessary to use hardware decoding due to insufficient CPU resources; and even on modern systems, sufficiently complex content (eg: 4K60 AV1) may require it.`

we're considering changing the default to auto-safe

My HRT got here and now I'm injecting myself weekly. Should last 70+ weeks.

awesome!

https://www.unicode.org/L2/L2021/21075-heart-emoji-coverage.pdf

this reminds me of a cursed idea i had for an image format that decomposes images not into cosine waves but into emoji

discrete emoji transform (patent pending)

I'm going to make a sussy dither pattern

Those shadows look so bad

It adds a fuzzy texture

Underneath the dexter paw

I have never seen such bad shadows

For some reason, the palette is only 92 colors, so could be a lot better

Think the shadow is weird because the model is slightly above the ground

Even taking that into account, what's that detached horizontal bar shadow coming from? It looks like a completely different shadow compared to the shadow caster

Maybe that screenshot uses potato settings because the paw looks like a totally flat blocky polygon with a blurry looking texture slapped on

No bumpmapping or anything

Looks like quake 3

You need to go on OnlyFans to get high res feet

Usually you're not that close to them, IIRC I was crouching for the other photo

Ok that looks more acceptable 😂

You have too many beryl nuts in your inventory

Just as a snack

hi! are there good books you can recommend on image compression? For educational purposes I started reading Jon's paper but I feel I sometimes lack the understanding of basic stuff (ie metrics to grade the quality of a codec)

The most well known metric is PSNR. It's a generic signal processing term and it's extremely bad at grading the quality of something in terms of human perception.

There is no well known metric that corresponds with human perceived fidelity.

Unfortunately.

Makes it a lot harder to fine tune encoders since the quality comparisons have to be done by eye most of the time

Or intuition

I heard of simulacra both in paper and other sources but also I heard that it is specifically a metric for jpeg block artifacts (I may be wrong). I'm trying to roll out my own codec in my masters thesis and I'm looking for good metrics for non DCT solutions. I will check that PSNR since I still need to find a way of getting the right quantization values

I mean I tried once just doing a grid search across several divisors and measuring MSE * scaled sizeof compressed image and then finding a minimum but it weren't that good/fast

Best perceptual metrics imo are currently: CVVDP, IW-SSIM (Y only but otherwise not bad), SSIMULACRA2, Butteraugli p-norm.

i will check them out! also the introduction in your paper is a really good resource for newbs like me, thank you for putting it out 🙌

Dssim is also pretty good

*So far* every time I've used it, I've agreed with SSIMULACRA2... Apart from 20 minutes ago but we'll ignore that one

At the very least, it doesn't tend to be massively off-target like some of the others can be

I mostly agree with SSIMULACRA2, the only exception being that it seems to weigh very low frequency retention too much relative to high frequency to my eyes in the low to mid-high quality region

so e.g. the highest scoring tuned 50 SSIMULACRA2 image can look a bit too soft around edges or fine textures

but SSIMULACRA2 is worlds better than OG SSIM or especially PissNR

are there any MOS correlation numbers for CVVDP?

Haven't heard of cvvdp but afaik there aren't any metrics that correspond with human vision at all... Heavily tuned and optimized encoders like x264 use custom parameters tuned to human vision instead of a naive metric

Encoders that are tuned for metrics and rely too much on metric scores rather than human vision, tend to look very blurry and bad

The metrics think an ambiguous blur looks better than rough grit and texture

Need new metrics?

All of the metrics ESPECIALLY hate it when there's an imperceptible amount of noise that the human eye can easily see through. Even when that noise is masked/hidden by a noisy background.

They would prefer everything gets blurred and blasted away into an unrecognizable smudge than for detail and grit to be preserved (which would increase the amount of objective noise, in a way that's completely transparent to the human eye.)

X264 was tuned for example to completely do the opposite of metrics here because metrics are actually stupid

It's usually called "psycho rdo"

Yep.

Metrics only care about the objective power of noise

Not whether that noise is transparent to the human eye

It does not distinguish the difference between completely blasting and smearing away details, vs adding a tiny amount of transparent noise that is hidden against a noisy background.

Metrics will usually prefer blasting away crucial image details

Metrics are too stupid and naive to tell one is a lot worse than the other

That's the fatal flaw of today's image fidelity metrics

The best encoders have a lot of tuning to do the complete opposite of what the metrics like

For example, noise shaping and noise masking techniques are very common in lossy audio coding

And in well tuned image codecs like x264

For example, a high contrast, noisy background can help to hide/mask distortion noise from compression. Flat, low-contrast backgrounds make noise more obvious.

That's also why noise is easier to spot in dark images and regions, because there is less contrast so the human eye is more sensitive to small changes.

And if the spectral profile of noise is similar to the spectral profile of the background, it can also hide noise. Or if the noise spectrum is shifted to the higher frequencies, it can help preserve dynamic range in the lower frequencies, both in image and audio!

That's why blue noise looks so good!

Sorry for the rant ;)

ssimu2 does not seem to do any dct or spectral analysis or anything to ensure the spectral profile of image features is getting preserved and that the higher-frequency part of the spectrum isn't getting blasted smooth...

Theoretically would doing a DWT on the image help with metrics?

SSIMU2 is multi-scale, so it can "see" image features at various levels of detail, and each scale (1:1, 1:2... 1:32) will effectively correlate on structural similarity of certain frequencies, with 1:1 being the finest, and 1:32 being the coarsest. I've noticed that SSIMU2 sometimes likes images encoded with steeper quantization matrices that I'd personally prefer in the SSIMU2 <60 range. I'd rather look at sharper looking images with a bit more distortion than SSIMU2 prefers. Above 60-ish I start agreeing with SSIMU2 though. Then again, I don't think I'm the average viewer with rating preferences SSIMU2 was trained on.

this is a good example: https://slow.pics/c/ECBjvPXX

I like "transparent" or even "translucent" distortion that the human eye can "see through" and preserve the information underneath... rather than the type of distortion that just simply DELETES spectral bands of energy from the original file. No metric score tries to even punish this!

In this case, tune iq uses a flatter quantization matrix than tune ssimulacra2, so while iq scores a bit lower, it keeps a bit more higher frequencies -- mostly seen on the tree on the lower left corner. Tune ssimulacra2 looks makes the leaves look like they have a mild gaussian blur applied to them.

Metrics are perfectly okay with image features being completely wiped out and obliterated by excessive smoothing that makes it harder to tell what the original file looks like. It's very much against the stated design goals of libjxl, yet libjxl does this too

I remember one of the stated design goals was to make compression artifacts more "honest"

libjxl used to be better than avif in this regard

Recent improvements in libaom have left libjxl falling short of this promise.

Maybe libjxl could benefit from a tuning option. One where you can instruct the encoder to smooth less by using smaller transforms or flatter quantization matrices (at the potential expense of other artifacts).

Today libaom produces more honest artifacts, which is surprising, since av1 coding tools are optimized for excessive smoothing and jxl bitstream coding tools are theoretically designed to be more well optimized for honest compression

av1 file format itself is designed for blur, while jxl is designed for roughness and grit and texture and fine details. But for some reason the libjxl encoder does not play to these strengths and instead has extremely steep quant matrices where it naively zeroes out the higher coefficients instead of coarser quantization and adaptive noise shaping.

So basically av1 is doing the opposite of what it's designed for and so is libjxl

> av1 file format itself is designed for blur That's an incorrect statement, the state of modern encoders are *tuned* to favor blur. It's an implementation issue NOT a spec issue.

If h264, a more limited spec can handle psy-rd properly, then logically AV1 and JXL which are more advanced should be able to as well.

It's a video codec. It's quite literally designed to have a lot more types of blurring filters

Because it's designed for motion and frames that appear on screen for fractions of a second

jxl is designed with still images in mind first, with high fidelity as the top priority.

The bitstream tools in jxl are all made with that in mind and are extremely good for high fidelity, visually lossless compression

It's an oversimplification but overall honest

You would not expect a codec like av1 to have more honest artifacts than jxl but current versions of libaom do

Correct, and you can e.g. make AV1 visually behave almost like MJPEG, so super blocky and not blurry. This encoder https://github.com/rachelplusplus/tinyavif restricts itself to 8x8 DCT transforms, no AQ, no filters of any kind, and DC prediction only.

True, but a lot of the av1 bitstream features are not appropriate for high fidelity still image compression and are obviously designed for high motion frames

It's worth mentioning that AV1 does have the ADST and IDTX transforms in its coding toolset, which don't inherently contribute to blurring, yet these can improve compression efficiency over just pure DCT.

And you can modify and tune them for high fidelity or not use them at all. Same deal with JXL or any codec for that matter.

Again, it's all about implementation

These transforms can be used quite often especially at higher quality efforts.

Yeah, what I'm saying is an oversimplification, but overall it's reasonable to say that it defies expectations for av1 to be tuned that way and jxl to be tuned the other way. Like they are playing to the opposite of their strengths!

Blurring is a strength of av1 and grit is a strength of jxl, yet libjxl is blurrier today than modern libaom

But that's oversimplifying the strengths of jxl of course

I know av1 has basically all the same coding tools of jxl but it's a more complicated and less flexible codec in terms of color, bit depth, resolution/seams, etc

No splines or patches or other cool modular features either

Anyways I really hope libjxl gets some form of adaptive noise shaping and smarter, courser quantization instead of zeroing out our precious image data in the future.

Reminds me of the blurry early days of Theora

https://people.xiph.org/~xiphmont/demo/theora/demo9.html

I'm still so angry that browsers removed support for Theora. it was supported for **15 years** then Chrome decides "we are just gonna remove it lol" and then of course Firefox follows a few weeks later with the only sole reason being "Chrome did it"

Wow. It's almost as if mozilla is just a skeleton crew vassal of Google now. 🙄

Theora is potentially a faster h.264 if someone makes a good encoder for it

It's actually a GOOD codec. Very simple and good and freely licensed.

Good economy of complexity-vs-compression

simple to decode and also has other chroma subsampling modes unlike *VP8*

Extremely fast on CPU

Yeah that's the most crazy thing about it

VP3 actually being superior to VP8

well Theora is a superset of VP3, iirc VP3 did not have other chroma subsampling modes only Theora does

No one wanted to improve libtheora after libvp8 was open sourced. Even though libvp8 was very poorly written and had no actual spec

I was kind of sad. I saw it as a mistake to switch to that piece of trash instead of putting that tuning/optimization effort into theora but everyone lost interest and ptalarbvorm never was released

still dumb to me how lossy WebP wasn't a superset of VP8 and is instead just regular VP8

It would not have taken long for libtheora to match vp8 efficiency at a much faster speed.

either obsessed with having WebP be hardware decodable OR just not caring and making a low effort image format

Rather than waste time on libvpx. But theora had a bad reputation and people wanted to move to the new shiny

And the poorly written new encoder was giving slightly better results than libtheora at the time

It was seen as having more promise and potential, and there was a trend moving towards higher complexity codecs with greater CPU usage...

I don't think libtheora even uses multiple cores yet

There are very few codecs that are using their spec to it's fullest potential.

Personally I think there is great value in getting 90% the efficiency in 10% of the time

People say x264 comes the closest, lol

x264 really did an awesome job

One of the x264 devs even tossed around the idea at some point, modifying x264 to output a different format

Which eventually happened for h.265

But at the time they were thinking vp8 or theora

That would be an amazing project

it also happened for MPEG2

in the form of x262

I wonder what happened to 'dark shikari'

I hope they're doing okay

He wrote a famous critique of vp8

One of the elder gods of image coding

The person who comes up with a fast way to leverage patches (or intraBC in AV1) will become a celebrity I think

I hate that the answer might be ✨AI✨

Same with splines in JXL 😅 . Nobody really knows how to do it at the moment.

https://tenor.com/bbhRGtfDtiD.gif

Spline subtraction would really be cool to add to a jxl encoder

Everyone knows it

just ask ChatGPT "where are the curly thingies in this image" 😂

How about a Captcha where it asks you to place the spline on the image 😂

I don't think machine learning is necessarily the most efficient tool for subtracting splines from images, since there are probably more precise and exact pre-existing tools for that

the image

But to be honest with you I'm not familiar with any pre existing tools that can decompose and subtract splines from photos

If anyone has a link to a piece of code or paper on that, post it on this server

How did you get a photograph of my scalp

Do I need to tape my webcam

I remember experimenting with OSX86 around 2010 or so, but hardware support for what I had was very limited

no hardware-accelerated graphics and no network

I was also interested in GNUstep around that time

I wanted to try out programming apps in Objective-C for some reason

don’t remember why

something about it (and Cocoa / OpenStep) felt mysterious and intriguing to me

How about AI crypto

Crypo inherently is associated with scams and loosing money, AI is just god damn annoying from how often it's overused

crypto is also short for cryptography tho, and cryptography is cool

More often that not, AI is a scam as well

I think they were referring to cryptocurrency specifically tho

i know

what a funny question. almost useless <:Thonk:805904896879493180>

there’s a definite possibility that AI is a bit overhyped at the moment, but there’s still at least a kernel of usefulness that’s hard to deny

automatic translation is a clear example that comes to mind

imperfect but usually better than nothing

I wouldnt call AI a scam for me it has been immensly helpful in writing code thanks to rsi, just need to baby it

I wouldn't say AI is a scam, it just seems overhyped sometimes

Much like crypto

Not all AI is scam, but a lot of it is

If this would affect the adoption of JXL https://www.macrumors.com/2025/03/10/doj-google-chrome-sale/

hardly

This wouldn't do anything. It would just change the formal owner of Chromium, but the actual owner would still be Google

It is very much like crypto though

how so?

1. Both are solutions looking for a problem 2. Both are nowhere near as useful as their simps claim, yet not completely useless 3. Both consume extreme amounts of electrical energy

they're similar in those aspects you mention, but there's also a significant difference: crypto is tied to economics/finance, and so additionally has all the problems that arise from that, while bringing few benefits to the table

So is a lot of AI, specifically, most of AI is tied to selling promises

that's not what I meant. I would consider AI marketed with "our AI trading bots will make you rich" compatible with the way in which I mean crypto to be related to economics/finance, but most of it isn't that

which is not to say that it isn't overhyped for profit. it definitely is. but basically everything is done for profit in our society, so that's not a useful distinction here

Like ChatGPT isn’t a “scam”, it does what they say it does

even if you think that the recent generative AI stuff is bad or unethical for various reasons, i don't think you can argue that the use of AI for things like computer vision or alphafold is bad

Not quite. For most part, it just generates worthless slop

i get a lot of value out of LLMs i mostly use them to search for things that are easy to answer "if you know", but really hard to google (e.g. what language some code is in), for writing boilerplate code, and for analyzing code too they do screw up a lot, but just the fact that i can throw huge files in there and ask questions about them and get answers instantly is extremely useful

like i was able to just throw an entire decompiled binary (no symbols, just used the ghidra decompiler) into an LLM and it could explain the general structure of the program, what function does what, and what data structures it uses

and that was a 300 kilobyte .c file (~11000 lines)

It absolutely does not. It's a lie and a scam by design and intent.

They don't make any attempt to give users accurate and useful expectations about what it is or how it works.

In fact they do the opposite

They intentionally lie about what it is and how it works in order to give a false impression.

my contention is that crypto is much closer to “completely useless” than AI is

even assuming that AI’s actual usefulness is 0.1% of its hype vs. 1% for crypto, I think it still translates into a much higher absolute number

I thought bitcoin was a pretty cool toy, back when I first heard of it and it was like $0.11 per BTC

A decentralized ledger to keep track of how many unicorn farts everyone has

Basically it's a strict improvement over the Federal Reserve Note

I would advise anyone who haven’t read https://singlelunch.com/2020/10/21/badeconomics-putting-400m-of-bitcoin-on-your-company-balance-sheet/ to do so

Not much

Also, AI is more harmful than crypto

How about this https://jpeg.org/jpegai/

I've gotten some use out of AI. For crypto? Nothing. Like on a societal level. I understand how AI can be bad; cheating on assignments, LLM generated articles & websites, dead Internet theory and so on. But on a personal level I have lost money on crypto. Family members and friends have lost money on crypto. It's just sucky. I have yet to lose money on AI because I don't spend any on it.

I know it's a selfish way of looking at things. But crypto didn't solve any problems for me at all.

AI certainly helped.

also at least with all the massive amounts of power being used with AI it results in something tangible but with crypto it's just stupid numbers with assigned values

If that AI means Apple Intelligence it would be useless

i guess i should rephrase. AI *can* result in something tangible but with crypto it's all stupid nothing

Let's cram AI into everything! EVERYTHING!!!!

This is actually true, although extreme energy consumption is limited to proof of work cryptos

JPEG 1 vs JPEG AI

Pfft

https://arxiv.org/pdf/2411.06810 looking into JPEG AI compression, it looks very sloppy, much like everything AI

As can kind of be expected, as far as I can tell / have seen so far, AI-based codecs are amazing at producing a plausible and great-looking image at incredibly low bitrates, while for high-fidelity compression they are not so good, or at least not better than jxl.

and I guess the decoder for AI-based codecs is huge too in filesize i mean

I think we'll see some sort of neural network prediction for (lossless) image formats in the future. I've read papers where they significantly scaled down the neural network predictors to the point where they became sort of competitive performance wise while still significantly outperforming classical predictors in terms of residuals

there are of course a lot of cases where deep learning is a pretty decent solution, and sometimes even more efficient than the classical counterparts, but those are massively overshadowed by the ocean boiling large language models which have taken over the discourse

I really like this application of it, for example. more accurate *and faster* than traditional signature-based file format recognition [Magika: AI powered fast and efficient file type identification](<https://opensource.googleblog.com/2024/02/magika-ai-powered-fast-and-efficient-file-type-identification.html>)

correction, it's "almost as fast", not faster. I misremembered

hmm, I wonder if the reason LLMs use so many words to give you very generic information or sentences that basically hold no information at all is because that's just an artifact from the training process, or if it's intentional, because when they get too specific they are wrong too often. you can avoid being wrong about specifics if you're overly generic all time time

there’s maybe also some potential in using AI to produce valid JXL bitstreams

splines are a use-case we’ve had on our minds for a while but never implemented, but there might be more

https://graphics.axios.com/hermesv2/2025-03-06-1401-timeline-of-us/index.html

This blog post argues that inflation is inherently good and deflation is inherently bad. I can see why the blog is called "BadEconomics"

It argues that "having control" over the money supply is a good thing, without asking who should have that control or what human incentives that would create for abuse. It asserts that the Gold Standard is "nonsense" too, even though gold is the most globally accepted store of value today and throughout most of human history.

These are pretty old economic arguments though...

It's kinda meaningless to assert one way or the other because people will never stop arguing regardless of what the truth is...

Basically, "do you wanna use some common fungible commodity like gold, or do you want to give a bunch of completely unaccountable rich people absolute control over the economy, and force everyone at gunpoint to use their fake monopoly money by requiring taxes to be paid in McDollars"

not really

did you not see it explain how not to get hyperinflation?

I kinda rolled my eyes and moved on when I saw the "fixed supply" problem and saw that it thinks inflation or deflation have inherent intrinsic values to them.

That's getting into the controversial territory that economic scholars endlessly debate about today

Personally I don't believe inflation or deflation have any intrinsic values any different than any other natural price adjustment in response to supply and demand. A lot of economists confuse the idea of "value" itself and often mix their own personal values and goals with some kind of objective and intrinsic property of the object they're describing.

At the same time I don't agree with all the crypto bros comparing their shitcoin to gold because "it's just subjective man"

It's difficult for philosophers and scholars to navigate the intersection of subjective human decision making and choices with objective consequences of reality that is at the focal point of economics

There is a such thing as objective and subjective and they do coexist but it's difficult to merge them together into economics without a lot of people getting confused and acting irrationally and coming to absurd conclusions

This confusion also benefits those in power. Politicians have mastered the art of weaponizing ignorance and confusion to justify doing the opposite of what they say they're going to do.

That's why even today you have articles like this where scholars will justify things like allowing rulers to counterfeit currency because it somehow benefits us.

Arguing that it's actually LESS risky if we give the people with the most power and least accountability the ability to counterfeit money and give loans to their friends and the well-connected

At the expense of everyone else, but especially the poor are affected the most since the poor have most of their savings in cash rather than other investments.

It's funny how most government policy is designed to really grind the poor into the dirt as much as possible when they are already at their low.

And to cut off the bottom rungs of the ladder too

Like how the income tax really is felt the hardest by single kids making minimum wage but not so much by wealthy people who hire an accountant.

I'm pretty sure it's by design. There is some kind of strange pleasure in kicking people while they are down

Keeping their noses in the dirt

The funniest thing is, when the income tax was being talked about and sold to the public for the first time, about 110 years ago in the USA, hilariously they said it was only for the rich.

I think it was Wilson? He assured the public that the average peasant won't ever have to pay a dime of this new tax

that's also not what it does

> Having control over supply of your currency is a good thing, as long as it’s well run > ... > Since the 1970s, the USD has been a fiat money with no intrinsic value. This means we control the supply of money

lmao. "We" control.

As if the author has some kind of control or say at all

The first half of the article definitely sounds like it. Maybe in the second half he dials it back... but I had read enough up until that point that I didn't expect anything different from what I had read hundreds of years ago at that point 😂

I took a peek into the second half of the article, and I see that he does tone it back a few notches.

But when the first half is like that it's hard to continue reading...

Since it sounds like we are accepting way too many shaky premises

I am more interested in the technical and objective side of things rather than debating about human problems.

Like how sustainable and scalable the network is.

The network protocol had to go through some forks and revisions I heard. I stopped following it after the price of unicorn farts ballooned to tulip-levels

Haven't been that interested in bitcoin for a long time since then

It was cool when it was a toy

Now it's too popular and I'm too hipster

😎

It's rare for me to have fun talking like this. It's why I enjoy the jxl community

I always felt like splines could be encoded relatively easily with specialized edge detection and curve fitting, but maybe not

Ultimately it does argue it's basically a necessary evil. Like a lot of modern economic scholars.

It would seem Windows decided to start defragging my SSD while idle....

250 GB written in 15 minutes while I was downstairs, by `svchost.exe (defragsvc)` I noticed the write count suddenly skyrocketed the other day, and has been every day since

Last I checked that's way *less* accurate, not more accurate.

Wrong. Literally 100% of economists agree that deflation is really bad.

Which makes sense. If money increases in value, stuffing bills in your mattress becomes a good investment. So you end up stalling the economy, and go into recession.

<https://learn.microsoft.com/en-us/windows-server/administration/windows-commands/defrag#scheduled-task> > When run from the scheduled task, defrag uses the below policy guidelines for SSDs: > > Traditional optimization processes. Includes traditional defragmentation, for example moving files to make them reasonably contiguous and retrim. This is done once per month. However, if both traditional defragmentation and retrim are skipped, then analysis isn't run. Changing the frequency of the scheduled task doesn't affect the once per month cadence for the SSDs. > > If you manually run traditional defragmentation on an SSD, between your normally scheduled runs, the next scheduled task run performs analysis and retrim, but skips traditional defragmentation on that SSD. .....What the fuck Microsoft

Why are you killing my drive

Deflation is just a price adjustment. There's nothing intrinsically good or bad about prices changing. Maybe the reasons or the consequences can be good or bad, but in and of itself, no...

It's just a price adjustment... of the price of money itself

There could be good or bad reasons for the price of money increasing in value.

"shooting someone isn't good or bad. Maybe someone being hit in the head by a bullet is good or bad, but in itself, no"

Shooting someone can't be compared with the value of a fungible item going up or down.

One is obviously harmful and one is completely and utterly neutral

A lot of economists will probably argue that there is intrinsic value here, because of the weird mattress stuffing argument that is oft repeated. But I think that's confusing the personal values of the economist with objective reality.

There is always a time to hold and a time to sell.

What the hell? Windows has automatic SSD defragmentation? What are they smoking?

Holy shit

I also noticed a long time ago (I use windows rarely) that windows is constantly thrashing when idle compared to Linux. I guess shit like this is one reason why

Right now I'm downstairs waiting for it to go idle, then I'll check the write count again and see if it finally stopped

The cpu stays in a higher power state and the Disk blinkenlight is always blinken.

It does not ever stop. In my experience.

My activity light is usually solid when it's idle, I guess now I know why

It was constant. This was years ago. But it never ever stopped

I mean this 'monthly defrag'

On linux the CPU and Disk lights were off with occasional twinkle

I talk a lot of smack about how bad linux is but at least it doesn't TRY do bad things on purpose.

It's stupid, not malicious

Windows is intelligently evil

Welp, time to see if it's wasted more of my drive's life

Hrmm... Still a few gigs, but at least it's not over 100 like before

Isn't defragmentation to SSDs on Windows just "trimming"?

No, this is actual defragmentation... Which degrades SSD lifetime for no benefit

https://support.microsoft.com/en-us/windows/defragment-optimize-your-data-drives-in-windows-54d4fed1-c96e-46db-b843-8c6b34bd27a4

.

> Hard drives are defragmented, which is reorganizing the files so that they are all lined up and easier for the drive to read. Solid-state drives (SSDs) are what's known as "trimmed," which is basically telling the drive where it can safely do cleanup work when it's not busy doing more important things like saving or loading your files. It's all optimization.

Check the link I posted

That's Windows Server. Quicker to get a new server

It's the same executable and scheduled task

Get a Mac

deflation is *really bad* and this is not debated by economists. but I guess be an armchair expert

that would be monetary deflation, though

it is debated by economists, though

it's just that mainstream economics is largely dominated by some forms of Keynesianism

except it isn't, and that's what the word deflation means

what does the word deflation mean?

it's when the price index decreases

monetary deflation is clearly defined

change the index, and you don't have deflation anymore

<:Thonk:805904896879493180>

you can't just "change the index"

you can

the index is based on the average price level of things relative to a specific year

it's not objective

it's unitless and always relative to a specific year so you can't just rescale it in a way that takes an increase and makes it look like a decrease

and flawed in many ways as a lot of price changes can not accurately be quantified

Like, you can't compare a smartphone now to one you could buy 10 years ago

no, but you can look at the aggregate price of goods in 2025 compared to the aggregate price of goods in 2015 using nominal dollars

not objectively

not saying it's meaningless, though

if you really want to be specific about "monetary deflation" it's always a bad thing and that's not debated

and if we had zero monetary inflation, prices would still change, of course

well they're essentially the same thing

productivity increases lower prices

because the reason money is less valueable is that the purchasing power of a dollar goes down

and that's because the nominal prices go up

well, sure

so in that sense monetary inflation is directly linked to the nominal price of goods

cause the value of money is largely dependent on its purchasing power

monetary inflation indeed influences the price of goods, yes

increases in productivity decrease the real price of goods, not the nominal price of goods

<:Thinkies:987903667388710962> no, if productivity increases, consumption increases

consumption and the nominal price of goods aren't directly related

they are

consumption increases when *reall wages* increase which isn't the same thing as productivity increasing

in fact, productivity has skyrocketed in recent decades and real wages have not

eh, that's just because labor compensation has shifted away from pure monetary compensation

that's false

No.

no, it's false

Also people often use bad deflators.

There is no increasing productivity labor compensation gap.

there is, and you're dead wrong.

this is well-documented

>heritage foundation

no thanks

in fact the source listed has been refuted directly by others

what James Sherk is doing is he's conflating mean with median

so as income inequality grows, so does the compensation difference between between a typical worker and a mean worker

that's a totally different claim, though

no, conflating typical with average is exactly the issue he's experiencing

because he's claiming that "average" wages have increased, which maybe they have! but it's also not a meaningful or useful thing to calculate because it ends up always correlating with productivity in a way that tells no story

because obviously money in and money out are going to be the same (at the end of the day, all costs and all value added are labor costs, even raw materials)

This is about average wages, though?

average wage is not a useful quantity because marginal propensity to consume is dependent on income per household

Good for you for seeing that.

I actually have studied this

so yes

you have to sum the marginal propensity to consume per household times the income of that household, over all households

average wage is just the sum of income of all households (divided by the population)

which is a less useful metric than the weighted sum by marginal propensity to consume

and *that* is what affects consumption, not average wage.

The common claim is that it is investors and entrepreneurs that suck up all that productivity and the wage laborers get less and less of that total productivity. Which is what just is not true, at least not to anywhere close the extent it is often said to be. Most people of course are wage laborers of some sort, and I never made a claim about how the income distribution changes there, because I just don't know. I strongly assume there is growing divide, though.

There is absolutely a growing divide, it's well-documented

I never disagreed with that.

well, which divide are we talking about

except you said that wage laborers aren't getting less of the total productivity

but they are

this is also well-known

they're not

are you including the top 1% in "wage laborers"

many of them sure are part of the top 1% income bracket

the only way you don't observe a decrease in that fraction is by including exceptionally wealthy people who are significantly wealthier than they used to be in that category

Like, there are investors and entrepreneurs earning less than the people they pay.

that's only true if the people they pay are also well-off

This is about income, not wealth.