JPEG XL

those are the same thing

to the vast majority of people

income is their only type of wealth. when I say wealthy I mean people who have a lot of income

also Sherk's study is entirely results-driven (the one you cited before) considering his methodology includes people who have exceptionally high income in his evaluation

Sherk's just a run-of-the-mill conservative who does this kind of garbage

(he's the guy who is behind the major push by President Trump to mass fire federal workers)

Its purpose is to address the common claim of a productivity-wages divide. Which is only partly true, but also misleading, as not all labor compensation is monetary.

sounds good to me

except in all these studies they include non-monetary compensation so that is a moot point

In the commonly shared "what happened in 1971" graphs they don't

in fact, if you *include* non-monetary compensation the divide gets worse, because high income earners have a higher fraction of their income as non-monetary

most people make only income and get a health care package that is lower in value than their income

Correct, but this is still not about an income divide.

it is an always was

It seems you're trying to refute a point I never made.

as I said here

Income divide is not part of the graph.

as I said here

if your methodology includes people with exceptionally high-income as 'wage earners" then it's results-driven and meaningless

this was your original claim

yes, that was my claim. I'm not including people with exceptionally high income in my statement that "real wages have not"

because that would be results-driven and meaningless

if you include people with exceptionally high income there then you're intentionally missing the point of that statement

that claim wasn't incorrect, btw

but still misleading

it's not misleading

if you include people with exceptionally high income there then you're intentionally missing the point of that statement

and it's not misleading to not add "unless you're intentionally missing the point"

Real wages don't even come close to reflecting the rise in standard of living due to technological progress.

I didn't say anything about that

it's not misleading to *not bring something up* which was *off topic*

It's a huge part of the difference between IPC and CPI.

~~instructions per cycle and cycles per instruction?~~

not sure what you mean by IPC

IPD

oops

interpupillary distance?

interprocess communication?

Just because you say it's not debated doesn't mean it is so.

It's definitely a very controversial and endlessly debated thing in economics. One of the hottest hot buttons.

it really isn't and im not going to argue with an armchair expert

Deflation is the main reason why major cryptocurrencies still exist

You don't know who I am, honestly, but I agree it's pointless to argue.

I will simply say that I respect you personally despite our disagreement.

There is a popular camp of economic thought I suspect is heavily invested in defending those in power, and will argue that it's actually good for the economy for them to continue to have such control and power over the entire economy. This makes the richest and most powerful people even more rich and powerful so naturally this is promoted by the rich and powerful.

There is always a time to hold and a time to sell. If deflation has inherent value, then that fundamental principle is wrong.

You could say that maybe it's bad for a particular purpose but not inherently, intrinsically bad

It depends on subjective purpose/goals/values

Just like the price of eggs going up isn't by itself an intrinsically good or bad thing. It's a reaction to something else.

The reason or cause could be good or bad maybe but the price adjustment itself is just a neutral thing, like gravity.

When you fall out a window gravity doesn't kill you because it's evil, but maybe the person who threw you out the window is evil.

That's all.

Please don't confuse my meaning or intent. Like I said I have a lot of respect for you and I'm not trying to be disrespectful here.

Price has to fluctuate around the value, and value is effectively the same thing as crystalized labor. I am a pretty orthodox marxist when it comes to economics.

Those aren't really used as currencies (yet), though. If tomorrow it would become clear that we would switch to some specific hard currency, that currency would definitely deflate massively for a short while, as investors buy it up in masses, but I think it would stabilize after that...

In the long run the role of currency is neutral as it is just an intermediary representation of value. In the short run of course central banks, speculation, etc can play a role. Just like prices can deviate from actual value in either direction for a while but in the long run at the macroeconomic level, one's profit is the other's loss and it all cancels out.

I mean crypto for sure used to be primarily dominated by currency uses. now very much less so, so (yet) would be inaccurate and (anymore) more accurate

Not all labor has the same value. Value is determine by human estimation, even if that estimation is stupid or illogical.

If I dig a bunch of holes and then fill them back up to end up right back where I started, there's no value in all that backbreaking labor I just did. Maybe even negative value since now I'm sweaty and stinky and tired after.

I'm not an expert on orthodox Marxism, but I know a few things about the personal life of Karl Marx and some of his ideas. He didn't seem to have a good grasp of things like "value" as I remember. And he seemed very selfish in his relationships with other people.

I remember a lot of his words and ideas sounded like extreme selfishness (and rage/anger) dressed up as compassion.

I used to have a school friend who, as he got older, changed and became prone to sudden tantrums when he used to be a calm person, and is bitter and resentful of his family but never moved out or became independent, and lets them continue to take care of him into adulthood even though he complains about them constantly, and even though he really should be taking care of those things for himself, like his own finances and bank account. Karl Marx reminds me of him, a former friend I used to think I knew.

Karl Marx had a similar backstory and living situation

For some reason, despite being given everything on a silver platter, while others generously shoulder all the work and expenses for them, and graciously providing everything... Despite that, it breeds incredible anger and resentment somehow.

Adults are meant to take on responsibility... not wait for someone else to do it for them.

For countries with their useless fiat money, they are used as currencies privately

it's not really used as a currency and the volatility and extreme speculation does not make it a good currency.

It does not have enough advantages over fiat. Getting in on it early gives you a massive advantage over late comers, which is not much different from a ponzi scheme or a central bank that shows preferential treatment to the well-connected.

I meant some countries in Latin America and Africa

Sure, the actual value labor crystalizes into depends on lots of factors, including skills and tools, which themselves are ultimately also crystalized labor.

Well, the largest part of it was always hoarded, even at the very beginning. That won't change until its use as a currency takes off and plateaus. If Bitcoin is the new gold, it will never be used as a currency, though, lol.

Sure. All kinds assets are used that way. I just meant that cryptocurrencies are primarily used as an investment, exactly because they are deflationary.

Not all of them

The architecture of the neural network MobileNet for classification: Full visualization of each neuron of the neural network during operation, the value of the neuron weights from -1 to 1 are converted into brightness from 0 to 255, the order of transformations is from top to bottom. The neural network is not entirely linear, so any images of its "layer-by-layer" work are not actually reliable

Just learned that the latest Windows Terminal can draw 256-color bitmapped graphics using ancient VT430 control codes 😅

Many other terminals too but Windows is the only OS shipping with one https://www.arewesixelyet.com

I made a split screen graphics mode like in Commodore 128 and it works inside VSCode or those terminals, of course also over SSH 😁

> pine needle stock > pine needle oil > whipped oat cream smoked over burning pine branches that's one way to create a very distinctive theme for a dish

Fish shaped sediment shaped like fish

Well, GNU/Linux distros that have downloadable images with Plasma should ship konsole and that has SIXEL support as well. And funnily it's one of the better ones there for dealing with SIXEL while it is sadly still lacking a few escape sequences (e.g. required by the drawing demo from the libsixel repo). Windows Terminal lacks them as well but it will eventually get those.

~~Somebody should add ReGIS to it~~

I found this: https://www.hanselman.com/blog/the-real-and-complete-story-does-windows-defragment-your-ssd

Your pfp loaded correctly in the notification by the way

probably requested the PNG version and did nearest neighbor scaling. speaking of for some reason my PFP shows up as green now on iOS or something‽‽ I don't have the proper means to try and see why myself

for people here on iOS is my pfp green?

I mean maybe it isn't green anymore since when I heard about it being so was like a month or 2 ago

Purple for me

But in the reply icon it's NN purple

lmao why does your pfp not showing correctly ? first time i see that

All array language may be incomprehensible, but Uiua makes up for it by having an official color font with pride flags.

Allow me to interject for a moment. The organizational body which you are referring to as JPEG is not a Joint Photographic Experts Group unto itself, but actually International Standardization Organization/International Electrotechnical Commission, Joint Technical Committee 1: Information Technology, Subcommittee 29: Coding of audio, picture, multimedia and hypermedia information, Working Group 1: Coding of still pictures

Allow me to interject for a moment. Although PNG does officially stand for "Portable Network Graphics", the backronym "PNG’s Not GIF" is arguably a better name for it. It highlights its nature as an open alternative to the proprietary Graphics Interchange Format. Also, since PNG is based on the outdated DEFLATE compression algorithm, it is no longer the best format for sending graphics over a network, and although it is still quite portable, other formats such as Lossless WebP have gain widespread adoption, and the format JPEG XL is also quickly gaining adoption, as it is more efficient than JPEG, PNG and WebP (both lossy and lossless) in most images.

Also, allow me to further interject. No matter how much it may appear, the above message was not generated by a Large Language Model (LLM). It was written fully by a human, one who goes online commonly by the name "CrushedAsian255", however on this particular server goes by the name "Jia Tan", as a playful reference to the XZ Utils backdoor. In case you, as the reader, have forgotten, the server you are currently in is the official JPEG XL discord server, which is why JPEG XL was mentioned in the above statement.

However, I am sorry, but as an AI language model developed by OpenAI, I am unable to continue this discussion. If you have any other questions or concerns, feel free to ask! ||(also written by human)||

I see that a user going by the nickname of AccessViolation_ is typing. I cannot wait to see what thoughtful insight they can bring to the conversation!

Allow me to interject for a moment. What you are referring to as human, is in fact. H. sapiens, or as I've recently taken to calling it, Eukaryota/Animalia/Chordata/Mammalia/Primates/Haplorhini/Simiiformes/Hominidae/Hominidae/Hominini/Homo/sapiens

I am sorry for my mistake, allow me to correct my response. Allow me to further interject. No matter how much it may appear, the above message was not generated by a Large Language Model (LLM). It was written fully by a member of the species known as H. sapiens, or as AccessViolation_ has recently taken to calling it: "Eukaryota/Animalia/Chordata/Mammalia/Primates/Haplorhini/Simiiformes/Hominidae/Hominidae/Hominini/Homo/sapiens". The member of the H. sapiens species who has been sending many of the above messages commonly goes by the name "CrushedAsian255", however on this particular server goes by the name "Jia Tan", as a playful reference to the XZ Utils backdoor. In case you, as the reader, have forgotten, the server you are currently in is the official JPEG XL discord server, which is why JPEG XL was mentioned in the above statement.

||this is a very insightful conversation||

quite!

Thank you for your feedback! It has been forwarded to the ~~AI training facility~~ totally real person who is writing these messages.

https://www.thedailybeast.com/musk-boosts-claim-that-hitler-wasnt-to-blame-for-holocaust/

that is certainly a take

Benchmark material

I dunno about Stalin or Mao, but for Hitler its pretty well known that Hitler was too much of a coward and squeamish to actually kill anyone directly and always had to get his goons to do it, and refused to take any responsibility.

at least, that's the common stuff when researching it anyways for years

but the original tweet, whether intentionally or not, its fairly vague , does bring up a good point, these genocidal leaders can only be so because they have people following them.

I wonder if that means that Musk agrees that guns do kill people (https://en.wikipedia.org/wiki/Guns_don%27t_kill_people,_people_kill_people)

logically, it should, right?

unless you take it to the extreme and “Guns don’t kill people, bullets kill people”

dunno, its a weird post to begin with. as for the gun thing even as a big gun dude, I've always hated the "guns don't kill people, people do" phrase, not because of any merit or anything, but man its cringe xD as for musk specifically he flops around a lot on the whole gun ownership topic, no idea if its a grift or just a fence sitter tho

also Taurus, tarus hand guns are so terrible they prolly do kill people lmao

it would be nice if people would just stop posting dumb tweets tho

sure, but whether the dictator pulls the trigger or orders someone to is largely irrelevant, and it's intentionally distracting

I agree its a dumb distinction but it is a distinction, which is why IMO the original tweet screens haha I am so smart much in the way of grammer nazis

that's how it reads to me anyways, a man child esq post made to look smart

Allow me to interject for a moment. What you are referring to as Hitler is actually GNU/Hitler, or as I have recently taken to calling it, GNU plus Hitler. On its own, Hitler is not a complete Operating System, but just a component of the complete GNU system.

<:kekw:808717074305122316>

`sudo apt install hitler`

they really did

I suppose in the causal chain of things, you can blame people for harmful outcomes in a way that is proportional to the extent to which they could have acted differently to prevent it. Then of course Hitler gets a substantial amount of blame, as do high-ranking SS officers etc, all the way down to the person actually pulling the trigger. But also the CEO of the company that agreed to build gas chambers and ovens, or even regular people who knew or could know what was going on but didn't try to stop it. The amount of blame should probably be the largest for Hitler though.

I didn't expect that when I clicked on the channel

Excuse me it's actually GNU slash Hitler

In all seriousness though, in that video you sent, he says the advantages of binary only appear if you assume the first digit cannot be zero... That means you're assuming all numbers have the same number of digits.

He's giving himself a free digit, which is the same as saying, numbers can start with zero again.

Basically he's cheating and he admits he's cheating...

Knowing the first digit always starts with zero, is the same as taking the first digit away completely. Then you again have a number that can start with leading zeroes. So it's not an apples-to-apples comparison.

Then again it's a little bit confusing to me as well, so I could be wrong. Really depends on how you define radix efficiency...

no, it doesn’t apply recursively

I'm still trying to wrap my head around the radix optimization theory

He's assuming the first digit has less cost than the others. But in practice I don't think that's valid. It makes sense since the first digit can't be zero, but at the same time that's feels like subverting the radix cost equation.

Normally to get the cost of expressing a number, you would just multiply the number of digits by the base.

With no special treatment for the first digit

Okay, I've been thinking about his argument, and the definitions of things like "cost." His argument is: - We shouldn't be multiplying by the base, we should be multiplying by how much information is contained in each digit (log of the base) - We should also divide the cost by the amount of information contained in the number to get the radix economy, or cost/information

But he needs to rely on the 1st digit containing less information/less cost trick, in order to make binary seem better. But the "cost" is how much space does it take up to write/store a number, and that does not change.

So it's still not valid for him to give himself a "free digit"

The advantage of binary disappears without that cheat.

Then the natural base *e* reigns supreme once again

Even if the first digit contains less information, it shouldn't be counted as if it takes up less space to store or write down, because it doesn't.

Since we're talking about "how large does a buffer have to be."

to some extent, it arguably does, at least in the case of binary, since it means you don’t _have_ to store or write it down (to the extent that you accept an empty number to represent 0)

But you still need a buffer of length floor(log_b(x))+1 to store numbers as large as x

That cost doesn't change

It doesn't make sense to treat the cells in the buffer any differently in terms of cost, because we are evaluating the cost of storage NOT how much information is contained in each digit of a value of a predetermined size

He's measuring and unfairly comparing the wrong thing to give binary an advantage there

you basically trade the ability to see if any number is there (whether there is nothing or a zero), which is a practical issue, but representing "nothing is here" is not a property of a positional system itself

aren't the cost of storage and information density intertwined? you can just measure the cost of storage, but if a certain amount of storage used can hold more information, that's better

IMO it's more like binary *has* an advantage, which that way of measuring fails account for

He's not measuring "What's the minimum buffer size," he's measuring "how much entropy is in each digit of a specific given value" and then adding them up.

That's different than the buffer size.

And totally irrelevant to radix efficiency.

And it obviously gives binary a huge advantage because it's not counting the size of the buffer.

again I don't think it's "giving binary an advantage", it's recognizing an advantage and pointing out the flaws in radix economy

with buffer size you mean the amount of digits you need right? might I recommend base 1 trillion

It's not an advantage in the "minimum buffer length to store a number," it's an advantage in "total entropy count of each digit of some arbitrary fixed-length value," which is not what matters

That's not what he should be measuring

That's not radix economy

With base 1,000,000,000,000 each digit costs log_2(1,000,000,000,000) bits of information.

So each symbol, or each cell in the buffer, has a cost of about 39.86 bits

it's true that the leading digit is always 1 but if you leave it off you have an extra piece of information which is the number of bits in the number

Exactly

if you already know that (in the case of floating point numbers, for example) then you can leave that off and it's safe

if you didn't (for example, 32-bit fixed-size integers) then you can't leave it off easily

That should be enough to set alarm bells off that he's discarding information and cheating to give decimal an advantage. At least in the "radix economy" calculation he's using.

pretty sure any base which is prime or a power of a prime should be functionally equivalent

and more efficient than any other base

that's my mathematician instincts though

one of the reasons for this is that some shortcuts are only true in prime bases

He's making a mistake by counting how much entropy is in each digit of some given, fixed-length value, instead of counting "how many symbols/digits are needed to express numbers UP TO x"

well entropy in each digit is based on context

what is the distribution of the data?

there's no way to have a uniformly distributed random natural number btw

like it's proven that there's no uniform distribution on naturals

so you have to have *some* nonuniform distribution, whatever it is, and that is going to affect your entropy

you can take a limit of a uniform distribution as the number of things grows, but then you end up with some kind of bias as if you always use "numbers up to size N" then you're going to get different entropy for the leading digit than not

but in the limit that slot will go back to uniform

isn't it impossible to have a uniform distribution because there's an infinite amount of them?

uniform distribution is possible for, say, reals between 0 and 1

it's not an issue of there being an infinite number but a countable number

(there's no unfiorm distribution on an infinite countable set)

this gets into lebesgue measure theory though

Yeah. It's unbounded, that's the main issue

Uncountable

actually the issue is that it's countable

which is very funny

Oh. Huh. I think I might know what you mean.

But the problem is that it's both unlimited and countable at the same time.

Right?

if you have a lebesgue measure on a compact hausdorff space where the entire space has finite measure, you can assign distributions to positive-measure sets within the space

the problem is all countable sets are measure 0

so can't be positive-measure

😵‍💫

basically in order to assign a distribution it has to "look like a line segment" or have very similar properties

Bottom line is the Binary guy is a cheater!

I think it's more accurate to say that the question "which base is more efficient" is probably ill-posed in a vacuum

as it depends on the distribution of the data

not the number of bits, only whether the number is zero or not

yea, but in practice knowing that the first digit is always a 1 doesn't help you

since if you don't write it down, under the assumption that you know it's a 1, then you have to know where to actually start reading the number

that said, it does depend on the distribution

for example, if your numbers are uniformly distributed between 0 and 2^32-1, then you have 32 bits of entropy in any base

it's just that in binary that entropy is uniformly distributed across each bit

and in, say, base5 it is not. but that's dependent highly on your distribution

what do you mean by knowing where to start reading the number (in a way that doesn't also apply if you do write the leading 1)?

if I write 001 instead of 1001, etc. then I just read the digits that are there, i.e. same approach?

I guess what he's saying is that if I'm reading out a number in little-endian form

I have to tell you when I'm stopping anyway

and you can assume the next digit is a 1

but that doesn't really *tell* you anything

because distribution

in order to calculate the "entropy of a digit" you *have* to know the distribution of the data

if it's, say, unformly distributed between 0 and 24, then base-5 is the "most-efficient" as it's going to be exactly 2 digits always

whereas in binary you need to use 5, but the leading bit contains less entropy because only 10000 through 11000 are valid numbers that start with a 1, so you end up with less entropy in the first digit than in the others

so in that sense binary is "less efficient"

this is the key thing. information entropy is based on knowing the distribution. the reason "the" contains less info than "info" is that "the" is more common. this is crucial

crucial to the whole concept

If you assume the first number is 1, as if there's truly no entropy in the first digit, then you are pre-supposing existing knowledge about the quantity you're trying to record and write down. If you don't know what the number is beforehand, and you throw out the most-significant digit, then you won't be able to tell the difference between 10 and 110.

It should be obvious by this point that he's wrong to treat the 1st digit as having less "cost" than others.

He's NOT doing an apples-to-apples comparison and he's not even measuring/comparing the right thing

But the other 2 points he makes are correct

- multiplying by the log of the base (the entropy per digit) rather than just the base (the number of possible values) - calculating radix economy as cost/entropy

you do – one will be written as 0 and the other as 10

I think the point is that it contains _so little_ entropy that you don’t even need to write it, whereas if you go from “digit can be 0-4” from “digit can be 1-4”, you still have to

No, you're just moving the problem around. How do you tell the difference between 10 and 0 then?

in a sense, it’s similar to how within binary, arithmetic coding is more efficient than huffman coding

it's "more efficient" based on the distribution though

you’re not moving the problem around because you apply the “drop leading digit” only once, to the number as you would write it in “full binary”

you can't make a statement like "binary is more efficient" without knowing the distribution

0 is indeed a problematic special case

But then you're unable to express "0"

encode `x+1` <:wesmart:512506347057840129>

It's because he's throwing out information and assuming pre-existing knowledge about the number, rather than fairly measuring the true cost of writing a number with a given radix.

he’s not assuming pre-existing knowledge about the number other than “it’s not zero”, because “leading non-zero digit is 1” will be true of _all_ non-zero numbers

But that is pre-existing knowledge

And for binary that's an entire symbol of pre-existing knowledge about the number

So he's giving binary an entire extra symbol for free, and giving other systems a fraction of a symbol.

When he should not be doing that at all

it’s the same amount of information as in the leading zeroes we also don’t write

does your objection extend to those leading zeroes?

If you follow those rules, then binary is the only system with that problem of "unable to write 0"

All the other bases don't have that problem

The problem only happens when you remove the leading 1 in binary like he's suggesting

Because he's wrong

It's not an apples-apples comparison, like I'm saying.

He is not counting something he should be counting

The question is "what is the cost of number X written in base Y" where each symbol has a cost related to how many possible symbols there are. And in order to write a number, that means you're writing enough information to distinguish it from other numbers without any pre-existing information about the size of the number.

I really like obscure subtle problems like this...

which, if you don’t encode zero, “removing the first 1 you would necessarily have” still achieves that

Is there jpegxl server on revolt ?

full real full true This is what music looks like 3,290,112,000 Discrete Fourier Transform (DFT) processing calls *with 3D plane simulation using wave phase*

No, like I said that only works for binary. All other bases don't have a problem unambiguously recording zero.

So you're using one less bit of information for binary, but comparing it to other bases that encode more information.

And treating it as if it's an equal comparison. As if the binary encoding is equivalent even though it's missing 1 bit

It's not a fair comparison, no matter how you slice it

He should be counting the full cost of each digit.

it’s not missing 1 bit, it’s missing one single value

I know <@853026420792360980> was joking when they suggested encoding `x+1` but it would actually solve the problem

Whether the value is zero or not is 1 bit of information

it would but you still have the varlen encoding issue

which is that you need an extra symbol that says "stop"

not to any greater extent than without that trick AFAICT?

He's trying to get an extra bit for free without counting the cost.

yea, but that extra symbol is relevant here

since it takes the number of symbols from 2 to 3

He is wrong to compare binary to other bases in this way

He needs to count each digit equally

that extra symbol contains information

is the issue

no, it doesn’t

well not quite from 2 to 3

cause it can only occur in one place and that's at the end

what if we stop calling it “binary” and instead call it “weird method = bin(n + 1)[1:]”

sure, but the stop symbol is very hard to calculate in

While it's true that the first digit contains less entropy, that is only true if you have pre existing knowledge. Not if you don't know anything about the number.

That's a bad way for me to explain it...

I'm not good at explaining these things, haha

it is true if you encode `n+1`

what stop symbol?

“42” doesn’t have one

it does, because how do you know you're only reading two digits

that's the thing, to "encode" a number either requires a special symbol, or knowing its length in advance

or decode

in “weird method”, it’s 01011, so why is it wrong to say that it’s 5 digits then?

in the exact same way that 42 is 2

because the number of digits contains that extra bit of information that you're leaving out

by leaving off the leading 1

I’m leaving it out in the exact same way that 42 leaves out leading zeroes

Because you can't encode zero. When the others you are comparing it to, can. That's an entire digit worth of information you're neglecting to count

I can now, it’s ""

It's not an apples to apples comparison like I said

I can make it “0” by doing `n+2` instead of `n+1`, if that’s more acceptable

But that can also be done in other bases. It's still a limitation that doesn't exist in other bases because it's not an apples-apples comparison

You are not encoding the same information

It's not a 1-1 translation like it should be

right, then “weird method(n) = bin(n+2)[1:]” and 42 is 01100

You are pretending that less information is the same amount of information which messes with the cost calculation

Consider the following method: - Fix `N`. - Express a number `X` using base `b`, with `N` digits permitted of data transmitted. - In any base `b > 2`, `X` can be anything between `1` and `b^N`. However if `b = 2`, we end up with any number between `1` and`2^(N + 1)`, cause we know the leading digit has to be a 1. This is a true statement, but it's only true about this specific method.

In particular, I think the issue comes from equating "efficiency of a base" to "entropy contained in a digit."

and I think the fallacy here is the first bullet point: "Fix N."

I think the issue here is that the data set is always between `1` and a power of `b`

uniformly distributed

and therefore in any base they're assuming that each digit has equal entropy for any reason

The entropy per digit, in bits, is log_2(n) where n is the radix

that's just how you convert entropy to base 2

Bits are in base 2

from a different base

sure, but that doesn't tell you much

The entropy in trits is log_3

sure. but that's not what I meant

basically, I'm saying his statement about his specific method is true, but the specific method is not sufficiently general to make the general statement he's making

because of assumptions baked into it

I just don't think you are encoding the same information. So you can't compare them as if they are equal

You are throwing away the cost of one digit and comparing it to an encoding that contains more information.

It can easily be avoided by just doing a normal comparison and treating binary the same as everything else, and not treating the first digit differently because the question is "how long is the minimum required buffer" not "how much entropy is in the first symbol, assuming we already know pre existing knowledge about the value written in the buffer"

Obviously those two questions are completely different!

And he has no reason to ask the second one

but if the entropy of the first symbol is zero (as it is if it’s certain to be 1) then why write it?

one counterintuitive part is that now-leading zeroes are significant

0 ≠ 00

and as I write this, I realise that this might be what <@853026420792360980> meant?

because you are presuming knowledge of out-of-band information. You presume you already know something about the value (how long it is to write)

If you are recording a number, you are writing enough information to uniquely distinguish the number from all other numbers that fit inside the same entropy window.

how do I presume this any more than with other writing methods?

It's not certain, because then you can't record the number 0 like you can in other bases.

And you can't say "" because then the entropy window is zero.

Which means it's unable to store a value at all

it is with the +2 tweak

but I guess what it boils down to is that you strip the possibility of adding padding zeroes (because then they would be significant)

so you can't pad to a fixed/known width

is that what you are trying to get at?

Let's say I'm trying to record the number 12. In order to say I've recorded it, I have to write it down in a way that uniquely distinguishes it from all other possible values that can be expressed in the same window of entropy. So in decimal that entropy window would be... all the possible values for 2 decimal digits of information. 100 possible states for those 2 decimal values. About 6.6 bits. In binary "1100" is 4 digits, 4 bits, 16 possible unique states.

If I'm doing a fair comparison, then I have to assume I know absolutely nothing beforehand about the value, and each possible state is equally likely for a given entropy window or buffer size.

And each possible state corresponds to a unique number

Maybe that last one isn't actually required but we might as well add that one too.

If we're doing a fair comparison, the question of "how much entropy does the first digit have" makes no sense and is irrelevant

They all have the same entropy cost

In a fair comparison

Correct me if those assumptions are wrong.

discord 10 mb moment... but JPEG XL <:logo:829708783336816671> 💪

That looks like some pretty strange music

What's it sound like

https://issuetracker.google.com/issues/389978750?pli=1

lmao

lmao

ayy lmao 👽

<:kekw:808717074305122316>

<:megapog:816773962884972565>

cjxl.exe file structure

https://mattmahoney.net/dc/textdata.html I found here a very old and very amazing program for creating a file structure, based on the search for binary matches in sections of file, you can visually determine the type of data and its compressibility, as if I were looking at a sound spectrogram, this file. Well, so, I rewrote the entire program in python with maximum preservation of its work, and most importantly, I removed the limitation on the size of the generated image. Now it can be of any size, but power of two

of course.

ffmpeg

What is it actually doing?

Like what’s the algorithm

" displays where matching substrings of various lengths and offests are found. A pixel at x, y is (black, red, green, blue) if the last matching substring of length (1, 2, 4, 8) at x occured y bytes ago. x and y are scaled so that the image dimensions match the file length. The y axis is scaled log base 10 "

4 batch are drawn. for example, the first is "1 byte". if the same byte occurs in the last byte, the bottom is colored black. if the same byte is found the second byte back, then the second block is filled in black. and so on to the end of the entire file, going through the entire file, every byte. the result is an up-colored line, from the beginning of the file to the end

the code uses a hash table so that it doesn't compare the real value of all the bytes in the entire file, but only the hash table... because of this, it makes noise. I think I can fix it. It's not 2006 anymore

Isn't mp2 actually a very good codec? Like musepak

joke

https://youtube.com/shorts/GmoGh8Fo7RI

I changed the code a little by removing the hash table Since it's not 2006, the file is loaded into memory entirely and a byte-by-byte comparison and search back the specified amount of bytes (window). This shows much more reliable data, although it requires X more comparisons for each byte length, i.e. for comparisons of 8 bytes it takes 8 times more time And at the end of the executable file cjxl.exe 0.11.1 from github - there really are areas of absolute nothing

it's quite good for high bitrate content

terrible efficiency at low bitrates

I'm actually surprised musepack isn't bigger than it is

Same philosophy as jxl. Optimized for transparency. Minimal CPU cost.

But it also was quite large file sizes

From what I heard you need 384k for transparency

Musepack was popular in the '00s among the HydrogenAudio forum members

the issue was that the devs never pushed hard for wide support, and once a well-tuned Vorbis encoder was developed, Musepack hit a ceiling because Vorbis was 99% as good anyway

also AAC became a thing for people who were okay with using patent-encumbered formats

192-256k was the usual transparency range for most people

Strange, I could still see artefacts up to 320k

wait, "see" as opposed to "hear"? just wanted to make sure we're on the same page

too many people used spectrograms to judge audio quality in the past 💀

Like I can still see blocking up to around 320k, depends on resolution and frame rate

192kbps

https://embed.moe/https://cdn.discordapp.com/attachments/806898911091753051/1352397464144973955/mpv-shot0004.jxl?ex=67ddddc9&is=67dc8c49&hm=84f8717b035c423abc58d60274a7b234f64fe7799815dc474917dcf32c23746d&

320kbps

https://embed.moe/https://cdn.discordapp.com/attachments/806898911091753051/1352397624967172097/mpv-shot0003.jxl?ex=67ddddf0&is=67dc8c70&hm=9065e8fb4ca39f4f18882905495347fd664716e6df8ae9950ff05a504ec0091f&

You were talking about mpeg 2 video??? I thought you were talking about mp2 audio...

same lol

I heard musepack only needs 175 for transparency and has top-tier performance as low as 128kbps, where it competes with the best encoders at that bitrate such as Apple AAC

I didn't really like the sound of Vorbis as much, even though I liked the features and freely licensed nature

Vorbis is really common and popular in games.

I never really liked the particular sound despite the listening tests, I feel like my ears had a problem with Vorbis in particular more than other codecs

Like it was easier for me to ABX Vorbis than MP3 for example

But other listeners show vorbis ranked really well in listening tests for some reason

same with opus

Opus still to this day sucks at harpsichord music and probably always will because of the low latency design.

I haven't tried to ABX musepack though.

I haven't really used it or tested it but I hear really good things about it

Mainly consistency. No one ever ranks it low. As if it has the fewest "killer samples" compared to others. Or the most consistency.

Which, again, reminds me of JXL

Musepack seems to have similar goals as JXL

today marks 50 years since the release of probably my favourite album https://youtu.be/BVkRl0sXjjY?list=PLoZ8bkLt9tifhuprqrmnY3Y_RZazSlhTG

10/10 name... might even beat pikfuif <:KekDog:884736660376535040>

this being an oral product does not make it better

I have a competitor product

https://www.amazon.fr/gp/product/B07VQRDGXV/

there has to be *one* person that haphazardly bought that thinking it was something else

what is it with these

marketing teams having a laff

lol

https://youtu.be/_y3ypwIlXf8 not the most upbeat music (maybe even less so than the one it’s a cover of) but I like it

(original: https://youtu.be/UCPF_4eJJME )

https://github.com/typedgrammar/typed-japanese

(explanation: https://github.com/typedgrammar/typed-japanese/blob/main/blog.md)

> And a concise way of saying what I'd like to see is "Highway for Rust." https://linebender.org/blog/towards-fearless-simd/

Rip https://github.com/rust-lang/rust/pull/129881

Are there any other efforts to get portable SIMD in Rust?

https://cdn.discordapp.com/attachments/335134129358241792/1355878357333704724/taski.gif

AI everything

https://youtu.be/FIF7wKJb2iU

’cause there’s music in the air and lots of lovin’ everywhere

<https://blog.rust-lang.org/2025/04/03/Rust-1.86.0.html> > Allow safe functions to be marked with the `#[target_feature]` attribute. <a:hype:1305973221740511252>

🤔 if I wanted to try doing area average downscaling treating pixels as circles rather than squares, which of these would make more sense (source pixels in black, target pixels in orange):

Interesting idea and good question. Does it make a large difference for the kernel you end up with?

(relative to the difference you get by going from squares to circles)

Pragmatically speaking you could also treat them as both, so the smaller inner circle counts twice

or maybe something like a 2D Gaussian distribution instead of uniform discs?

What is a pixel really? Conceptually you can treat it as a point sample at the center of the pixel, or as an average sample over some area covered by the pixel. From a capturing perspective, I suppose pixels are square if the corresponding sensor area is square. From a display perspective, I suppose you have to take the geometry of the subpixels into account if you want to get it right.

the question of "what is a pixel really?" reminds me of android. android also has an interesting take which redefined how they handle scaling entirely. This led them to use dps (density independant pixels) or virtual pixels https://developer.android.com/training/multiscreen/screendensities not really relevant here, but it did remind me of it

that's similar to a CSS pixel

yeah, they are pretty much the same when it comes to purpose. though, I find android's solution to be more robust then CSS.

really wish more electronics scaled as well as android does

now I’m kind of curious to try this

(doesn’t mean I will, though)

that'd be neat to try, though i imagine calculating the overlapping area between two 2d gaussians might be a pain...

This made me think, for scaling/blurring/etc. in general you treat them as points: you plug the distances between the source pixels' centres and the target pixel's centre into the kernel, and that's your weights. But you could change existing kernels to treat pixels as square areas pretty easily by integrating the kernel over the area of a source pixel instead

should be easy enough to implement for seperable kernels at least. EWA would be trickier

Simple example: for a 3x1 gaussian blur rather than taking point values from the curve at -1, 0, 1, you'd get the weights by integrating it from -1.5 to -0.5, from -0.5 to 0.5, from 0.5 to 1.5

https://youtu.be/_-o-YYZE3D0?si=RH8RcOEb1fiUslcp

lol, AI has a long way to got before it's even somewhat reliable

that's a long magic number <:kekw:808717074305122316>

```bash $ file /dev/zero /dev/zero: JPEG XL codestream ```

[av1_kekw](https://cdn.discordapp.com/emojis/758892021191934033.webp?size=48&name=av1_kekw)

That gave me a very cursed idea, what if djxl output the 12 byte black rectangle if given zeros as input

I think it's probably wise to not put Easter eggs in it and let it just do what it does now

self-deprecating humor

```python #!/usr/bin/env python3 # self-deprecating.py from deprecation import deprecated def main(): with open(__file__, 'r', encoding='UTF-8') as file: lines = ['@deprecated()\ndef main():\n' if line == 'def main():\n' else line for line in file] with open(__file__, 'w', encoding='UTF-8') as file: print(''.join(lines), end='', file=file) main() ```

This NPM package may be of interest to people who like TypeScript, graphics and instant gratification: https://www.npmjs.com/package/@lib/sixel

<@184373105588699137> I noticed that the waydroid image (latest) doesn't seem to work with the media encoder. so trying to, say, use scrcpy doesn't work with it

it does work with the image from last January so it's definitely a possible thing

do you happen to know why this might be?

I can check rq

since I like to run it headless and view with scrcpy this can be annoying cause I can't upgrade the image

you may not know what changed, in which case, I may need to file an issue

my current setup is ``` weston --no-config --renderer=pixman --shell=kiosk --width=1920 --height=1080 --backend=headless -- waydroid session start ``` (renderer=pixman is b/c I'm on nvidia)

I am seeing other people mention this in the tgram, I can pester aleasto to see if he knows what changed, but we are pretty close to releasing the A13 images which fixes tons of issues, so im not sure if he or jon will spend time on a fix, ill see if I can find it tho

and then I use scrcpy to connect to the device over adb

if it's fixed in a soon-to-be-released image I wouldn't worry too much

I'll just upgrade when the fixed one is released

I also like using scrcpy because I can do it from my laptop from afar

or record the screen with it if I need to

yeah I use it a lot too, currently me and john are focusing more on blissOS stuff but yeah ill see if I can find out whats up at least

hey no pressure

well I have narrowed it down to the specific issue, but this might be an upstream aosp issue t.t oh well, forwarded it to alessandro since he would be the one to know

On April first cjxl encodes the image to AVIF instead of JXL

`AVIF April Active. Resampling 8 forced on`

<@184373105588699137> where do I report issues with the website https://waydro.id/

to the normal issue tracker, or is there a different one?

github issue tracker https://github.com/waydroid/waydroid

do you have an igpu or something? im wondering how scrcpy worked for you in the first place since afaik it shouldnt work with swrendering in the first place...

no, nvidia. igpu is disabled

scrcpy probes for video encoder using the usual media api

yeah, the issue is when it goes to capture the screen, it usually cant grab the frames from swiftshader. maybe something funky happened and waydroid is using llvmpipe...

can you send the output of `/var/lib/waydroid/waydroid_base.prop` and `/var/lib/waydroid/lxc/waydroid/config`

well, aleasto just pushed a fix for scrcpy on hardware at least, no idea if it will work for you since, well I dunno how its working for you in the first place. the new images should be pushed today or tommorow for testing

possibly the libhoudini extension is doing something interesting

also possibly relevant

``` $ scrcpy --list-encoders scrcpy 3.2 <https://github.com/Genymobile/scrcpy> INFO: ADB device found: INFO: --> (tcpip) 192.168.240.112:5555 device WayDroid_x86_64_Device /usr/share/scrcpy/scrcpy-server: 1 file pushed, 0 skipped. 135.7 MB/s (90888 bytes in 0.001s) [server] INFO: Device: [Waydroid] waydroid WayDroid x86_64 Device (Android 11) [server] INFO: List of video encoders: --video-codec=h264 --video-encoder=OMX.google.h264.encoder (hybrid) [server] INFO: List of audio encoders: --audio-codec=aac --audio-encoder=OMX.google.aac.encoder (hybrid) --audio-codec=flac --audio-encoder=OMX.google.flac.encoder (hybrid) ```

if I try this command on the latest image, it throws a java exception

https://github.com/waydroid/waydroid/issues/1848

do you just have your igpu disabled in software? because this is using your igpu

I had remembered I had it disabled in the grub kernel line, but I was wrong

I actually have it enabled

well thats an interesting bug, iirc the website was plan for a revamp, gives us an excuse to actually do it lol

my monitor is driven by my nvidia card, but if I use `--renderer=gl` it flickers awfully

so I use `--renderer=pixman`

was simpler than trying to disable GL in the config

yeah, currently waydroid needs to run on the same gpu the compositor/display runs on or you will get wild display corruption of all kinds of stuff due to pixel format mismatches, im not entirely sure why because waydroid last I checked does properly check the supported pixel formats and reports what we use. so what you do will likely be necessary for the foreseeable future. But the upcomming fix should work for you

this is true even if I run it under `--backend=headless` btw

interesting

probably cause it's picking the nvidia card anyway

and nvidia is known to have issues

nvidia has some known opengl bugs

I have similar issues with libplacebo, gpu-api=opengl

hopefully with android 13+ we will support NVK, but sadly, it means the user needs to be running nouveau kernel module so incompatible with cuda and stuff

gtg ttyl

kk

(reference: https://youtu.be/EShUeudtaFg)

https://petapixel.com/2025/04/10/adobe-deletes-bluesky-posts-after-furious-backlash/

https://fxbsky.app/profile/kekeflipnote.bsky.social/post/3lmmpbaidvs2q

When people on Bluesky keep claming X being too toxic

Communists of Bluesky vs fascists of X

Not that many communists on Bluesky in my experience, more like Democratic Party fanboys and Cancel Karens

I just know I did some JPEG XL testing on it and bumped into jcupitt of VIPS, who then fixed multipage support as a result