PLAYGROUND · EVERYTHING BELOW RUNS — NOTHING IS A RECORDING

Toys thattell the truth.

Two live pieces. The terminal is the site's easter egg given a whole room — it navigates the site, reads the CV, and has a direct line to Max, the resident assistant. The second downloads a real language model into this tab and runs it on your GPU. Where your browser can't run something, the page says so instead of faking it.

01 · THE TERMINAL — ALSO ANSWERS TO THE ~ KEY, ON EVERY PAGE

yash@gb10 — /playground● waiting for input

yash@gb10:~$

COMMANDS./hireaskcatclearexithelplsnvidia-smiopenseedwhoami

PROJECTSbuild-from-scratch/max/multi-agent-patterns/prism/letsbuild/velox-ai/talent-agent/letsfinetune/

or press ~ anywhere on the site

static preview — the button opens the real overlay: command history, tab completion with inline suggestions, and `ask <question>` streaming Max live into the prompt.

FIG P.1 — SMOLLM2-135M IN YOUR TAB

A language model, running where you're reading.

Straight talk before you click: load model streams roughly 100 MB of q4-quantized SmolLM2-135M-Instruct weights into your browser — a genuine wait on a slow connection, which is exactly why nothing downloads on page view and the fetch only ever starts on your explicit click. From there, generation happens on your own GPU via WebGPU: no server, no API key, and nothing you type leaves the tab. If your browser can't do WebGPU, the panel admits it and shows a clearly labelled canned example instead of pretending.

run a language model in this tab

HuggingFaceTB/SmolLM2-135M-Instruct on your GPU via WebGPU - ~100 MB of weights, downloaded only if you ask for it. Nothing leaves your browser.

checking WebGPU support…

WEIGHTS FETCH FROM THE HUGGING FACE HUB ON CLICK — NEVER ON PAGE LOAD · THE FULL-SIZE REPRODUCTIONS THIS TOY DESCENDS FROM LIVE AT /RESEARCH