The Lab · Demos
A nerd playground — every proof you can run yourself.
No slideware, no canned outputs. Every card below is real computation — a real FFT, a real compiler, a real quantiser, a real recurrent net — running live in this tab, right now, on your machine. Pick one, break it, watch it prove itself.
Build a circuit — drag, clip, watch it light
A real DC circuit simulator, built from scratch: Modified Nodal Analysis over the nets you wire up, solved with hand-written Gaussian elimination, with a piecewise-linear LED that lights only when it's genuinely conducting. Clip components together with crocodile-clip leads and watch real current flow — the FBL electrical-component primitives, made playable.
▶ Launch demoThe ternary quantiser
A real weight matrix, collapsed live to {-1, 0, +1} with BitNet's absmean round-clip — the real ~20x compression ratio, the real reconstruction error.
▶ Launch demoThe Gate — native ≡ interpreter
One AST, two independent runtimes — a tree-walking interpreter and a compiled bytecode VM — proven byte-identical live, with a real mismatch surfaced on demand.
▶ Launch demoHolographic binding
Bind concepts by circular convolution (a from-scratch radix-2 FFT), superpose several, then unbind by correlation — the right concept lights up.
▶ Launch demoThe char-net
A ~2.9k-parameter recurrent net, generating text one real forward pass per character — the same class of tiny net that once ran on a 1982 ZX Spectrum.
▶ Launch demoThe glyph compressor
A real from-scratch LZW dictionary compressor, round-tripped byte-for-byte on whatever text you paste, with the measured compression ratio shown.
▶ Launch demoWave recurrence
Four resonant channels — position and velocity, integrated frame by frame with velocity Verlet — plucked live, decaying in real time, never a looping clip.
▶ Launch demoSpeculative decoding
A cheap draft model proposes tokens ahead; a stronger target model verifies them in one pass — the real, measured acceptance rate and tokens-per-pass speedup, from the actual accept/reject outcomes.
▶ Launch demoHuffman softmax
A real Huffman tree, built live over a token frequency distribution you control — the measured average bits/token against a fixed-width baseline, verified by a real encode/decode round trip.
▶ Launch demoMini-FBL — one spec, three targets
One small typed spec generates a real SQL table, TypeScript interface, and JSON Schema document — each reverse-compiled back and hashed, proving all three are still provably the same source.
▶ Launch demoCite or refuse
A real on-device model that only ever answers from a matched fact, with a citation. Anything outside its tiny fact store gets a hard refusal — enforced by retrieval, before the model is even asked.
▶ Open the reportWatch an AI write code in your browser
A real on-device model streams a JavaScript function from a prompt, on your own GPU. For the fixed example tasks, that code is then actually executed in a sandboxed Web Worker against real unit tests — a real ✓/✗ per test, never a claim it works.
▶ Open the reportA bug that becomes a compile error
A real lexer, parser, and static safety checker over a tiny PA/PIDS-style spec — an unbounded loop on a critical channel, a missing fallback, or an out-of-range duration turns red with a line number, live; fix it and the badge flips to a green 'certifiable ✓'.
▶ Launch demoRegister allocation, live
Real backward liveness analysis builds a real interference graph from a program you edit; a real Chaitin-style simplify/select pass colours it with K registers — drag K down and watch honest, real spills appear.
▶ Launch demoA director with a real STOP button
A deterministic, non-LLM planner expands a goal into a real dependency DAG and executes it tick by tick, running independent branches in parallel — press STOP and it halts immediately, provably, not just visually.
▶ Launch demoDoes it stick?
A real logistic-regression classifier, trained by real gradient descent on two tasks in sequence — watch Task A's accuracy genuinely collapse without rehearsal, and measurably hold with it, on a live, real curve.
▶ Launch demoTamper with a block. Watch the chain notice.
A real, from-scratch SHA-256 chains four real blocks by real content hashes — edit any block's data and watch its hash mismatch and the next link break, live, then re-mine a real (small) proof-of-work fix.
▶ Launch demoCorrupt one bit. Watch the checksum catch it.
A real, from-scratch sensor wire codec — sync, header, packed payload, and a bit-by-bit CRC-16 — re-encoding live as you move the sliders. Click any byte to flip one bit and the decoder honestly refuses the frame rather than handing over a wrong reading.
▶ Launch demoFind the brain-wave in the noise
A from-scratch radix-2 FFT over a synthetic one-channel EEG: slide 'eyes closed' up and the real measured 8–12 Hz band power rises out of real noise — band-pass reconstruction, live spectrum, and a detection verdict that only flips when the measured share crosses the threshold.
▶ Launch demoA moving picture on three symbols per pixel
A real ternary delta video codec running live on canvas: every non-key frame carries only {-1, 0, +1} per pixel plus one real absmean step — source and reconstruction side by side, with real PSNR and bits/pixel measured every frame.
▶ Launch demoBeing honest about it
Every demo above genuinely computes, in your browser, on every interaction — no library does the hard part in secret, and nothing is a canned or pre-recorded output. Most are faithful, from-scratch reimplementations of a real published technique (BitNet's ternary quantisation, Plate's Holographic Reduced Representations, a real interpreter/compiler pair) sized to run instantly in a tab — not the trained, protected systems those lab reports describe. Where a demo is illustrative rather than a reimplementation of the protected internals, its own caption says so plainly.
Back to The Lab Notebook