But the experience frustrated me. It was hard for me to understand all the software under me as I provided abstractions above me.

So I spent the last 5 years gradually eliminating all the layers of abstraction that add complexity to my Lisp interpreter. The path passed through one other language for teaching programming: http://akkartik.name/post/mu. The sources for it are archived at https://github.com/akkartik/mu1 (there was an earlier prototype in Arc at https://github.com/akkartik/mu0)

At this point I have a very simple syntax for programming pretty much directly in machine code: https://github.com/akkartik/mu#readme. It can be translated to an ELF binary for Linux using about 250KB of static x86 instructions (80% of which are unit tests, and much else of which is duplicated because I built the translator in multiple passes that run in a shell pipeline:

    $ cat examples/ex1.subx |./tests |./dquotes |./assort |./pack |./survey |./hex > a.elf
    $

The nice thing about the resulting ELF binaries is that they can be run directly on a Linux kernel with no other dependencies. Not even libc.

There's a script in the repo called `gen_iso` that can take a list of .subx files, translate them into an ELF binary and package up the ELF binary with just a Linux kernel into a bootable disk image. You can then boot this image either in Qemu or on a cloud service like Linode (http://akkartik.name/post/iso-on-linode)

This is what I have so far.

By contrast, the screenshot is quite fake. It's just a program that reads a line of text from the keyboard and prints it out to the screen. You can see it running on an emulated computer in Qemu that has nothing but a Linux kernel.

But I'm going to build up from that core into a high-level language. Maybe an Arc-inspired Lisp. Not a toy this time around.

Just give me 5 more years :D