wolfram-fb0 — AI writes x86_64 asm + eBPF for fractals, in a real VM in your browser

Why this needs a real VM, not a container

Most AI coding sandboxes are containers. Containers can't:

Run bpftrace against kprobes — needs real kernel access.

Open /dev/fb0 — there's no framebuffer device in a container.

Boot a nested qemu-system-x86_64 on real virtualization — needs KVM.

islo gives you a real virtualized VM in your browser, with a real kernel, real /dev/fb0, and real eBPF. Which means an AI agent can write code that talks to the kernel for real — and you can watch it happen from a tab.

The build, in five acts

Provision — islo use wolfram-fb0 spins a real Linux VM with nasm, qemu, bpftrace, Ollama (+ the smallest Gemma), and opencode pre-warmed. ~30s to a working shell.

Code — Inside the sandbox, islo skills (plan → build → review → refine) drive opencode with a local Gemma to write pure x86_64 assembly for Rule 30, Mandelbrot, and Julia. No libc, single ELF, talks to /dev/fb0 directly.

Judge — An oracle in the same sandbox assembles each iteration, runs it in --ppm mode, pixel-diffs against a Python reference, and reports binary size. The agent optimizes both.

Boot — qemu-system-x86_64 boots a 4 MB Linux straight into the agent-built ELF on a real framebuffer. AI-written bpftrace programs attach to the binary and stream every syscall + every byte mmapped to fb0.

Share — islo share exposes the framebuffer stream + the eBPF event stream as public URLs. You see the fractal bloom in real time, with the kernel-side trace scrolling beside it.

Live from the sandbox

Framebuffer

Stream connects after a sandbox is forked.

eBPF trace

[trace] waiting for a forked sandbox…

Convergence

Binary size and pixel-diff vs iteration, by target. Filled in after the first full agent run.

chart will render here

Lineage

Fractint (1988) · FractalAsm · A New Kind of Science · islo · opencode

Try islo

Every new account on islo.dev ships with $50 of free credit — no card required . That's enough to spin a real-VM sandbox like this one and run the full convergence loop end-to-end yourself.

Get $50 credit on islo →<br>Code on GitHub