A Linux distro whose first userland is written in Malbolge, a ray tracer in Brainfuck, and an APL dialect you can pip install
Three new GitHub esolang projects put hostile languages to serious work: a bootable Linux distro whose first userland is a Malbolge-derived glyph program, a ray tracer generated as 23 MB of Brainfuck, and an APL-derived array language that ships on PyPI.
Three repositories created during the September 13–27 window each needed a tool built before they could show anything at all.
8th-circle-linux boots a Linux userland whose PID 1, shell and core command set are glyph programs in a Malbolge-derived virtual machine. rayfuck renders a ray-traced scene out of 23 MB of Brainfuck. bAsedPL is an APL-derived array language in Rust that installs from PyPI. 123The creation window ran from September 13 at 20:00 through September 27 at 20:00, UTC-05:00. The repository figures below are each project's state during the verification pass.
The three picks
| Category | Project | Created (UTC-05) | What it does | Snapshot |
|---|---|---|---|---|
| Most absurd | bigdevboss/8th-circle-linux | Sep 19, 12:18 | Boots Linux with a Malbolge-derived VM as init, shell and command set | 5 commits; 1 star; 0 forks; GitHub lists Python, the VM runtime is C 1 |
| Most show-off | mTvare6/rayfuck | Sep 23, 11:27 | Ray-traces the Ray Tracing in One Weekend metal scene from 23 MB of Brainfuck | 13 commits; 14 stars; 0 forks; Brainfuck, MIT 2 |
| Most practical | AnswerDotAI/basedpl | Sep 16, 18:22 | An APL-derived array language in Rust, installable from PyPI | 68 commits; 10 stars; 1 fork; APL, Apache-2.0 3 |
Most absurd: a Linux distro whose PID 1 is a glyph program
Malbolge is the language people reach for when they want to end an argument about hostile syntax. It gives you 59,049 words of memory, ternary 10-trit words, three registers, self-encrypting code and the
crazy operation. 8th-circle-linux takes that machinery and boots a computer with it: the kernel is ordinary Linux, and PID 1, the shell and the core utilities are printable glyph programs for M8, a Malbolge-derived VM written for the job. 1M8 keeps the 10-trit words, the
a, c and d registers, the Malbolge-style decoding, the shared code and data memory, and the general bad idea energy. It then adds operations for syscalls, jumps, branches and small data moves. The README gives the trade in one line: M8 takes those extensions "because this thing has to boot Linux userland rather than win an esolang purity contest and die immediately." 1Booting it takes
make check to build the runtime, the userland images, a filesystem image, the rootfs and the initramfs, and make qemu to start it. Linux unpacks the initramfs and runs /init, which is the M8 runtime. The runtime loads /sbin/init.mb, and that program claims /dev/console, calls setsid, switches to /, prints /etc/issue, mounts /proc, /sys and /dev where the kernel allows, then forks and execs /bin/m8 /bin/msh.mb. If the shell exits, PID 1 waits a second and respawns it. 14The loader rules are where the project stops being a joke and starts being a system. A
.mb file is printable source: whitespace is skipped, every loaded cell must fall in printable ASCII 33 to 126, and every loaded cell must decode at its own address to a known M8 operation. Remaining memory is filled with the Malbolge crazy operation, executed instruction cells self-cipher through xlat2, and old decimal memory images are rejected whatever header they carry. Inline frame words use five printable base-94 cells, and the assembler picks frame glyphs that also decode as valid source at their addresses, so a syscall number cannot be smuggled in as arbitrary printable junk. 1The tooling is what holds the whole thing up.
tools/m8asm.py assembles the glyph programs, tools/m8audit.py runs a strict glyph and source audit, tools/m8_loader_tests.py feeds the loader deliberately bad images, and a pure-Python builder packs the initramfs. make check runs the audit and the negative loader tests before QEMU ever boots, and make qemu-smoke boots the image through a PTY and inspects what came back. 56The shell is written in the same dialect as everything else.
msh reads input through the ordinary Linux read bridge, parses the command in M8, splits up to eight whitespace-separated arguments, builds an execve argv vector in M8 memory, and launches /bin/NAME.mb. The shipped commands are pwd, issue, echo, cat, ls and four MFS tools, and ls does its job by calling getdents64 and parsing linux_dirent64 records in M8. 1The filesystem arrives as an image read by the M8 tools: a 512-byte header, 128-byte entries, a fixed 16-slot table with payloads starting at byte 2560, and a builder that refuses more than 16 files. Booting into it and reading the single file it ships produces the project's own summary of itself, the line
The filesystem is not sane, and neither are we. 1The author is explicit about the edges. The distro artifact today is a kernel plus a generated initramfs, and an ISO installer is still to come;
make distro copies a kernel image from the build host instead of compiling Linux from source. The repository is five commits old and carries one star. 1Most show-off: a ray tracer in Brainfuck, compiled from C
rayfuck renders the metal scene from Ray Tracing in One Weekend as a 23 MB Brainfuck program, and the repository's README says exactly that much. The author's own write-up explains the motive: while relearning CMake, the author read a claim in the tutorial that even ray tracers had been written in the CMake language, and that this is not a recommended practice. He asked what would make a worse host language, picked Brainfuck, and reasoned that a simple language obviously gives a simple codebase. 27The pipeline explains where 23 MB of source comes from.
ray.c passes through an SSA-like form in ray_ssa.c, where recursive code is made iterative and variables are prefixed by function to avoid name collisions; from there it becomes a small IR the author calls a DSL, and dsl2bf.py lowers that IR into Brainfuck, with c2dsl.py handling the earlier stage. The author notes that the SSA conversion was the only job he gave to a language model. 789Doubles have nowhere to live on a Brainfuck tape, so every one of them occupies four cells in signed Q16.16 fixed point, with the fractional and integer halves split across those cells. That gives a resolution of 1/2^16 and a range of roughly ±32,768, which the author chose over the cheaper Q8.8 because the ground plane is a sphere of radius 1000 that has to read as flat. 7
The arithmetic is rebuilt from eight one-character instructions. A move is
[->+<] and a copy is [->+>+<<]. Multiplication is repeated addition with the cells paired, so a four-cell by four-cell product adds each a_i*b_j into result[i+j]. Division is school long division that subtracts the divisor up to 255 times per cell. Square root began as a Taylor fit, which the author plotted and rejected below an encoded value of about 20,000, and ended as long division. Negation uses the two's-complement trick across cells; comparisons decrement two cells together with 128 added to the most significant cell so that negative bytes sort below positive ones; and the supersampling random number generator is the recurrence A = (5*A + 1) % 256. 7The cost is stated plainly. The program is 23 MB, larger than the 0.9 MB image it produces, which the author calls a rather poor choice for a compression technique. He measured roughly one pixel per minute and projected about 62.5 days for a 400×225 frame on his laptop, and of the first 1,229 pixels rendered, 10 differed from the C output, mostly by one value. 7
A comment on the author's Reddit thread asked how fork and join primitives might improve it, and the detour went into the JIT interpreter the program runs under. That produced the render now sitting in the repository, which he describes as looking a bit like a Van Gogh painting, probably from precision error. Both images ship beside the source:
vangogh.png from the Brainfuck run, a.png from the C. 71011The repository itself is thin: thirteen commits, fourteen stars, an MIT licence for the code and a separate licence for the JIT it depends on. The write-up carries the detail the README leaves out. 2
Most practical: an APL derivative you can install this afternoon
bAsedPL is an APL-derived array language that borrows from J and BQN, implemented in Rust with a native executable, a Jupyter kernel and a Python API. The AnswerDotAI repository appeared on September 16, has 68 commits, and published six versions to PyPI inside the window, from 0.1.6 on September 20 to 0.1.11 on September 26. It runs on Python 3.10 or later. 312Its arguments with APL are about notation, and they are the ones a reader meets in the first ten minutes. Arrays are based in the BQN sense, so numbers, characters and functions are atoms and enclosure always adds a layer. Brackets write vectors, so
[1 2 3] is a vector and spaces group, which makes a+b × c+d read as (a+b)×(c+d). A run ending in a function is a train, so avg←+/÷≢ defines the mean. Positions and axes count from zero. 3Small examples show the shape of it.
+/⍳10 is 45. The primes below 50 come out of one line, 1+⍸2=+⌿0=|⌝⍨1+⍳50. ⌊⌾(10×)1.25 2.78 returns 1.2 2.7, because ⌾ is Under: transform, apply, reverse the transform. 2×⍣(⍳5) 1 returns 1 2 4 8 16, one state for each count, where count 0 is the initial value. 3Numbers default to approximate, with
x marking exact integers and r exact rationals, so 1r3+1r6 is 1r2 and √¯4 returns 0j2 for the complex root. Axes carry names and string keys, which lets +/⍠"month" sales reduce a keyed array by axis name. System functions cover JSON through •json, charts through •plot, and SVG built from element functions, and regular expressions run on Rust's regex engine. 3The Python side reaches the same depth.
from basedpl import fn turns the string '+/÷≢' into a Python callable, and arrays convert out through .py, .np and .df for Python values, NumPy and pandas. The documentation site carries notebook-generated guides for each of those areas. 313The maturity is the trade-off. This is a 0.1 release moving several times a week, documented through Quarto notebooks that also generate the README. 3
The quiet corners
MrxSiN/YouTubeMalbolgehides YouTube ads and skips SponsorBlock segments with an Xposed module whose feature policy is written in Malbolge. Ad patterns, category behaviour and resolver weights come from.malprograms that the build compiles ahead of time into JVM bytecode, so the shipped APK carries no Malbolge interpreter and no feature-handler dispatch table. A purity script fails the build if conventional handlers appear or if any behaviour program fails to evaluate as executable MBP1. The validation record documents a Pixel 8 Pro on Android 17 installing the APK, every hook group reporting ready, a cold app launch in 909 ms, and an upgrade hot-reloading without restarting YouTube. The README still opens with a warning against using it and a promise to rewrite it in Brainfuck. 141516Norne9/Brainfuck.Jinxis a .NET 10 Brainfuck toolchain built to be read: a lexer, a peephole parser, a second parser whose idiom rules for[-]and[->+<]are pluggable, a reference interpreter, an IL-emitting JIT built onDynamicMethod, and abfshell that can print the optimised op-code tree instead of running it. Its xUnit v3 suite asserts that both executors produce identical output on the same program. 17nstarke/js-to-malbolgecompiles an integer-based JavaScript subset into Malbolge Unshackled through a bytecode VM with its own assembler, disassembler and versioned binary format, garbage collection included. Its README warns before you start that even small programs produce files tens of megabytes long. 18
The window's most persistent single author was
DannyBaanks, who published three Malbolge repositories in one day: a Malbolge Classic VM in Zig that runs Hello World in 48 steps, an episodic execution record built on SHA-256 chaining, and a Brainfuck-to-Malbolge compiler with separate Malbolge Free and Classic backends. 192021Whitespace and Befunge produced nothing that qualified this window: the whitespace keyword returned ordinary whitespace tooling, and the Befunge matches were a Star Trek Attack Wing website whose
.bf files tripped GitHub's language detection.The verdict
Absurdity belongs to
8th-circle-linux, which took Malbolge's hostile internals into a place where a boot has to succeed, and then wrote the assembler, the audit tool and the boot test that let it make the claim. The show-off crown goes to rayfuck, whose 23 MB program is the output of a compiler pipeline lowering C through two intermediate forms, with the author measuring his own render in pixels per minute and publishing the number. Practical value goes to bAsedPL, which a reader can install this afternoon and which keeps changing in public every few days.All three are tooling projects as much as programs. One wrote an assembler and a loader for a dialect of Malbolge; one wrote a code generator that lowers C through two intermediate forms; one wrote an interpreter, a Jupyter kernel and a Python binding. In each case the machinery around the language is where the work happened.
References
- 18th Circle Linux repository on GitHub
github.com
- 2rayfuck repository on GitHub
github.com
- 3bAsedPL repository on GitHub
github.com
- 48th Circle init scroll source
github.com
- 5M8 assembler
github.com
- 6M8 runtime in C
github.com
- 7Writing a ray tracer in Brainfuck
epestr.com
- 8rayfuck DSL to Brainfuck code generator
github.com
- 9rayfuck C to DSL converter
github.com
- 10rayfuck Brainfuck render
github.com
- 11rayfuck C render
github.com
- 12basedpl on PyPI
pypi.org
- 13bAsedPL documentation
answerdotai.github.io
- 14YouTubeMalbolge repository
github.com
- 15YouTubeMalbolge validation record
github.com
- 16YouTubeMalbolge purity check
github.com
- 17Brainfuck.Jinx repository
github.com
- 18js-to-malbolge repository
github.com
- 19DannyBaanks/MALBOLGE
github.com
- 20DannyBaanks/Zigbolge
github.com
- 21DannyBaanks/Malfuck
github.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.