AI wrote x86_64 assembly + eBPF for Wolfram fractals direct to
/dev/fb0— every step (build, kernel trace, qemu boot, framebuffer) runs in a real Linux VM in your browser via islo. No install. No drift. Fork-and-rerun in one click.
Most AI coding sandboxes are containers. You can't write eBPF, touch /dev/fb0, or run nested qemu in a container — no kernel modules, no real kprobes, no framebuffer. islo gives you a real virtualized VM in your browser, which is what makes this demo legal: an AI (smallest open-weight model we could fit) writes pure x86_64 assembly and eBPF programs that render Stephen Wolfram's cellular automata (Rule 30, 90, 110), the Mandelbrot set, and a Julia animation directly to the Linux framebuffer. Inside the same sandbox: qemu-system-x86_64 boots a 4 MB Linux straight into the agent-built ELF. eBPF traces every syscall and every byte mmapped to fb0. Both stream out to your browser tab as the fractal blooms.
- Fractint (1988) and FractalAsm showed that hand-tuned assembly + framebuffer is one of the most visually rewarding low-level art forms in computing.
- Almost every AI-coding demo today is a webapp. The opposite of a webapp — pure ELF, no libc, kernel ABI only — is exactly the kind of work that's hard for LLMs (no Stack Overflow corpus to copy) and easy to verify (pixel-diff is unforgiving, byte-count is a scalar).
- And every "reproducible" AI demo dies on install friction. The README rots. The toolchain drifts. The model gets deprecated. The cloud-sandbox-as-replay-link is the reproducibility primitive these demos have been missing.
src/ — x86_64 assembly (rule30.s, mandel.s, julia.s)
bpf/ — bpftrace programs that X-ray the asm at runtime
oracle/ — Python reference renderer + pixel-diff + binary-size scorer
harness/ — agent loop: islo skills orchestrate, opencode codes, oracle judges
dist/ — final agent-built ELFs
iterations/ — every loop iteration, committed (the convergence record)
qemu/ — boot.sh + tiny initramfs for the framebuffer demo
site/ — GitHub Pages: plot, live viewer, charts, fork link
docs/ — design notes
islo.yaml — declares the real-VM sandbox shape
You don't need anything installed. Click the Fork the sandbox button at the top, or:
islo use wolfram-fb0 --source github://zozo123/wolfram-fb0
# → boots a real VM with nasm, qemu, bpftrace, opencode, model, all warmed.
# Inside: `make demo` boots the fractal, opens the eBPF trace, streams to your browser.To watch the agent rebuild from scratch:
islo use wolfram-fb0 -- make agent-loopYou'll need nasm, ld, qemu-system-x86_64, bpftrace, Python 3 with pillow/numpy, an Ollama with a small Gemma pulled, and opencode. We support this for masochists. The full sandbox image is one make sandbox-snapshot-import away if you really want it.
| Target | Asm budget | Math | Why it's interesting |
|---|---|---|---|
| Rule 30 | ~256 B | integer XOR | Wolfram's signature 1-D automaton. Trivial math, ruthless on byte budget. |
| Mandelbrot | ~1 KB | SSE float, no libc | Forces the model to use raw SSE; no sin/cos, no libm. |
| Julia animation | ~1.5 KB | SSE + frame loop | Adds time as a dimension; tests state-machine discipline. |
After the agent loop finishes, this section will embed the charts: binary size descent over iterations and pixel-diff convergence, by target. (Will appear after first full run.)
- Fractint — the 1988 ancestor; integer-math fractals on a 386.
- FractalAsm — modern asm fractal renderer.
- Stephen Wolfram, A New Kind of Science — the cellular automata.
- islo — the real-VM sandbox that makes the whole thing one click.
- opencode — the inner-loop coding agent.
Every new account on islo.dev ships with $50 of free credit, no card required — enough to spin a real-VM sandbox like the one this demo runs in and reproduce the full convergence loop yourself.
Built by Yossi Eliaz (@zozo123) for islo.dev.
MIT. Fork freely.