Pathetic low-frequenciers blog

Sometimes I post about strange displays, so here is another bunch of them:

— Top left: a display built of 1152 seven-segment indicators (as ones on old calculators), a project from hackaday.io. — Top right: Seth Robinson is porting ancient games (Super Mario Bros, Castlevania, Pitfall!) for working on the holographic Looking Glass display; check the video. — Bottom left: Bryan Braun wrote a webcam script that uses a table of checkboxes as a display. He also has other funny crafts based on his checkboxland library (https://www.bryanbraun.com/checkboxland/). — Bottom right: Neil Bostian launched Doom on an office phone (not only display, but fully ported) and described the process in detail + you can watch the video.

Recently passed NaNoGenMo 2021(the "national novel generation month" I wrote about previously). I didn't have time to participate this year, but I skimmed through the works and share with you my personal top:

1. John Lambert put together a system that cuts music videos into separate frames, runs them through a description generator, and collects text that retells what is happening in the video. The result is a book of 22 such retellings.

2. Kevan Davis sliced Moby Dick's text into pieces (mocking Moby Dick is a long-standing NaNoGenMo tradition) and put them together in an interactive game book-style adventure. It turned out quite well.

3. Mark Sample used an ageless trick — translated "Pride and Prejudice" by auto-translators in the following cycle: English->Russian->Chinese->Portuguese->English. The result is a book called "Pride and Injury."

4. John Ohno, impressed by the Dark Shadows TV series, put together a generator called Shark Dadoes, which produces dialogues consisting of evasive answers and question-to-question answers.

There were other funny things, like generating synopsis for films by title, director, and actors; replacing all matching words in Moby Dick with emoji; a neural network trained on the texts of Phillip Dick.

Sometimes I write digital archaeological posts, for example, about the origin of keyboard layouts, how r-pentomino was invented, or about April Fools' RFCs. Today I will write a little about the etymology of various computer terms.

Everyone knows that the word modem comes from a combination of `modulator` and `demodulator` - two devices used to convert digital information into a form convenient for transmission over analog networks and back. The word codec (`[en]coder` + `decoder`) and the less widely known slang terms like balun (`balanced` + `unbalanced`) and serdes (`serializer` + `deserializer`) have a similar origin.

Similar to a `codec` in spelling and sound, the name of the Kodak company, registered in 1888, has a different nature. The company's founder, George Eastman, wanted to invite a new word — short, easily recognizable, and pronounced in different languages. According to legend, he used a set of letters from the Anagrams game (the grandmother of the Scrabble game). One of the criteria for George was the use of his favorite letter K, which accounts for 40% of the result. The idea to make a new word was not entirely successful: in 1896, on the pages of the Amateur Photographer magazine, readers made a dispute, trying to find out the word's origin. It was found, for example, that in Hindustani (it came there from Persian), this word means "boy," and one of the readers pointed out the similarity with the Hebrew `Kahdak`.

The word bit in the sense of a minimum amount of information was first publicly used in Claude Shannon's 1948 article" Mathematical Communication Theory." Claude himself referred to the authorship of the mathematician John Tukey, who used `bit` as an abbreviation for `binary [information] digit` in internal documents of Bell Labs. The word byte (distorted English `bite` as `piece`) stands for the minimum amount of information processed at one time or directly addressed. Werner Buchholz first used it in 1956 in the design documentation for the IBM Stretch system. On different systems, bytes come in various sizes, for example, 4, 6, or 9 bits (the size of a byte can even be variable). To accurately indicate the size of a byte of 8 bits, it is common to use the term `octet`.

For engineering reasons, it is more efficient for computers to work with numbers that are powers of two. Therefore, engineers often understand the word `kilobit` as 1024 bits (2^10), but in some cases, it means 1000 bits (10^3, as with other measures, such as meters). For example, the 1968 year's edition of the Encyclopedia of Library and Information Science, on the same page, states that 1 kilobit is 1000 bits, and 1 kilobyte is 1024 bytes. A similar story with the prefixes mega, giga, and so on. All this confusion continued until the end of 1998 when the International Electrotechnical Commission finally came in and fixed it (no). Since then, according to international standards, kilobits should mean 1000 bits, and for 1024 bits, the term `kibibit` should be used. However, not everyone agrees with this: according to the Russian "Regulations on the units of quantities" from 2009, the term kilobyte is fixed anyway in the value of 1024 bytes.

To measure the data transfer rate, in addition to any `kilobytes per second` and `kilobits per second`, engineers sometimes use terms based on the word baud (for example, `kilobaud`). Baud in modern communication usually means the number of changes in the carrier frequency per second, so if, for example, the carrier uses two signal levels, then 1 baud is 1 bit per second. But this is not accurate because the bits are considered gross here, i.e., include any overhead information, such as error correction. These `bauds` are named after Jean Maurice Émile Baudot, a French engineer who, in 1870, invented the basic encoding for the telegraph (aka International Telegraph Code # 1).

I already wrote about the self-organization of different creatures in swarms and algorithms for modeling such behavior. I'll write a little more:

Almost 100 years ago, Wolfgang Koehler conducted his famous experiment on sound symbolism. People were shown two pictures (the top row) and were asked to choose which of them was "baluba" and which was "takete." The majority of people chose a rounded baluba and an angular takete.

Since then, the experiment has been repeated with people who speak different languages, with two-year-olds, and so on. Researchers also tried changing the words, for example, to buba/kiki. In all cases, the effect was preserved.

Since multi-modal models have become very popular this year, Nearcyan from Austin decided to see what the CLIP model thinks about these words. In the second row, there are examples of generated images for kiki and buba, in the third — for the forms of "maluma" and "takete."

I recently wrote about the neural network generation of pixel graphics by Tom White.

Top left —  the sculpture Trinity by Frank Haase, a translucent cube whose three projections are three different QR-codes. Bottom left — the QR Rubik's Cube with six different messages on different sides; I once made it as a birthday gift. Top right — my QR code, made using the approach described by Russ Cox in the excellent article QArt Codes. Bottom right — the three-layer code invented by Eckart Schadt: depending on the distance, the contrast of some pixels changes, and the code is read differently (it works very poorly from the screen, try the printout.)

Tom White, an AI Artist from New Zealand, came up with an idea of how to generate pixel art images with VQGAN+CLIP networks. For the second week now, he is posting the neuro-pixel-art alphabet in this Twitter thread (he got to the letter W yesterday). I suspect a huge amount of cherry-picking; anyway, he promised to publish a colab soon, so you can experiment on your own.

As usual, there is not enough time for anything, so here is just another selection of strange/exciting things without any special comments (sorry if I missed some sources):

The other day, I found the colab with a TPU to run a shiny GPT-J-6B (https://colab.research.google.com/github/kingoflolz/mesh-transformer-jax/blob/master/colab_demo.ipynb) (this is kinda an open-source version of GPT-3). The env is brewed for 15 minutes in total, but then processing one prompt is fast enough, like, in 10 seconds.

I tried it to generate jokes about mathematicians in a bar, so I share the results with you (almost without cherry-picking):

A mathematician walks into a bar. A mathematician and a philosopher stand by the bar, in conversation, and a bartender is doing his rounds. The mathematician says to the philosopher, “You know that art and philosophy are branches of mathematics, don’t you?” The philosopher says, “Yes, I knew you’d eventually ask that question.”

A 0 of mathematicians walk into a bar, and they are immediately thrown out because "0 is used for zero" or some other way of saying "0 is a number".

An infinite number of mathematicians walk into a bar, and just one gets his research published. Who is that?

An negative number of mathematicians walk into a bar, and say "I'm actually going to be less than zero".

A transcendental number of mathematicians walk into a bar,   and hear one mathematician say, "Look, I've found the best beer in town." This mathematician says, "And look, I've found the best steak in town."

The boom around VQGAN+CLIP generated images and videos continues to increase. This video I found earlier today on Twitter, but that more interesting, is a fresh post on the Berkeley ML blog (the whole blog is good, actually). This post formulates an exciting idea that we see the formation of a new demoscene here.

In Scotland, there is such a phenomenon as tartans. These are textile patterns, unique for districts, clans, families, etc. (examples in Fig.A); historically, they play a role similar to the coat of arms; they are used to create kilts, scarves, etc. The first tartan found, Falkirk, dates back to ~250 AD, and now there are a lot of them — more than 3000 are currently registered in the official register.

They look pretty different but could be simply parameterized by generator codes like G106R26B4Y44 or G24K8G2K8, so it's easy to generate them, and there is already a twitter bot alltartans for it (Fig.C). On the other hand, these patterns are similar to unfinished fractals, so there are attempts to draw hyperbolic tartans (Fig. C). The square of the Cantor set is also called Cantor Tartan (and is similar to the Sierpinski carpet, Fig.D); for some reason, someone is trying to define a calculus on it.

Today it's time for a "strange robots" rubric:

There was a South Korean company Hankook Mirae Technology; they were making cruel exoskeletons (just like in MechWarrior). There are some documentaries about them, and Bezos once took one of them for a ride just 5 years ago. Then suddenly, it turned out bad: the company owner, in moments of mental anguish, beat the employees, fired at them with a BB gun, forced them to kill chickens, and did other interesting things. In short, last year, he was sentenced to 7 years of jail, and now even the company's website is down.

By the way, the design of this exoskeleton was made by Vitaly Bulgarov, a famous industrial designer (he also made projects for Ghost in the Shell, Transformers 4, and other movies). There are a lot of powerful works on his website.