01
CPUsingle main thread · JavaScript
—waiting
Level 01 Power-on
I build from the gate up.
Bootloaders, byte-level packing and game loops at the bottom. Interfaces people actually feel at the top. Same craft, different altitude — and lately, a lot of it done by talking to the machine instead of typing at it.
Scroll to climb the stack
Level 02 Silicon
Everything is a NAND
gate wearing a costume.
Flip both switches. That one lonely output is the entire tower — every language, every framework, every shader you have ever loved is this, repeated until it becomes an opinion.
Live gate · NANDA·B = 0
0
0
Output Y 1
Truth table
| A | B | Y |
|---|---|---|
| 0 | 0 | 1 |
| 0 | 1 | 1 |
| 1 | 0 | 1 |
| 1 | 1 | 0 |
NOT tie both inputs together
AND NAND, then invert
OR invert both inputs first
ADD two of those, carry out
CPU keep going for a few thousand years
Level 03 Bytes
Field order is a
performance decision.
The compiler pads your struct to keep every field aligned. Write the fields in a careless order and you pay for holes. Same data, same semantics — reorder it and a million particles get 8 MB lighter.
struct ParticleAs written
Memory layout · 1 byte per cell
32bytes / struct
14bytes wasted
0 MBsaved at 1M
Level 04 Languages
I've outlived most
of my toolchain.
I started at an Ok prompt and I'm still here, which means I watched most of this list die. The dead branches taught me more than the surviving ones — you learn what an idea was really worth by watching what happens to it.
Still shipping
Retired
Ideas moved here
1978 — now
Pick a row
hover or tab through
Amber runs to the right edge. Everything grey stopped somewhere, and the dotted lines show where its ideas went to live instead.
1991The Ok prompt
GW-BASIC. Line numbers, GOTO, no undo. A computer is literal, not clever.
1998Under the floorboards
Bootsectors and assembly. 512 bytes, and the magic number at the end.
2003Games
The frame budget is a hard number that doesn't care about your abstractions.
2009Bytes on purpose
Cache lines and alignment. I still lay out data before I lay out code.
2016The last pixel
Frontend. Same optimisation instinct, new unit: attention instead of cycles.
2025Vibecoding
Describe intent, review fast. It works because I know what's underneath.
Level 05 Pixels
Data is pretty when
you display it right.
Below is one dataset — 48 hours of frame times from a render loop. The numbers and the picture are identical information. Only one of them tells you where the stutter is.
frame_time_ms · 48 samples
Hover the chart · the spike is at hour 34
—
Level 06 Compute
Same pixels.
Four ways through the machine.
One deterministic Float32 image, one naïve 5×5 Gaussian convolution, four execution models. They run sequentially so you can watch the abstraction move: main-thread JavaScript → Workers → WebAssembly → WebGPU.
768² px
3 warm passes each
5×5 Gaussian · clamp edges · Float32ready
Run the sequence. The preview will be replaced by each backend's result as it finishes.
02
Workersparallel row bands · persistent pool
—waiting
03
WASMC · clang -O3 · linear memory
—waiting
04
WebGPUWGSL · 16×16 workgroups
—waiting
End report—
—fastest warm path
—vs CPU
—pixels / pass
—max abs error
| Backend | Cold first result | Warm median | Throughput | vs CPU | Validation |
|---|
What “cold” includes
CPU: allocation + first pass. Workers: pool creation, input distribution, first pass, and result gather. WASM: memory creation, module instantiation, input copy, first pass, and result copy. WebGPU: adapter/device, pipeline compilation, upload, dispatch, and readback.
What “warm” measures
Median of three passes after setup. CPU reuses its output buffer. Workers keep input resident in persistent workers and return only a tiny completion message. WASM reuses the instance and linear memory. WebGPU reuses device, pipeline, and buffers; host time waits for queue.onSubmittedWorkDone().
Correctness
The CPU result is the reference. Every other first result is compared element-by-element. Small WASM/WebGPU differences are expected because their accumulation is Float32 while JavaScript accumulates in Number before storing to Float32.
Read this benchmark correctly
This is one browser, one device, one kernel, and one image size. Workers can lose on small jobs because messaging dominates; WebGPU can lose cold and win warm because pipeline/device setup is expensive. The useful crossover is the point, not a universal ranking.
Off scale Mind
Off-script,
on purpose.
The gauge on the left just went past its top tick, which is roughly what happens to my evenings. The best engineering ideas I have stolen were not from engineers.
Tesla
Run the machine in your head until it breaks. Then build the version that doesn't.
Galileo
Point the instrument at the sky and report what you actually saw, not what was assigned to you.
Einstein
A thought experiment costs nothing and can still move the universe a few degrees.
Quantum mechanics
The maths works and the intuition doesn't. Learning to sit inside that gap is a professional skill.
Nature
Four billion years of iteration with no roadmap and no meetings. Still the best code review available.
Human origins
Every abstraction we've built is stacked on a primate that learned to count. Design accordingly.
Philosophy
Before optimising the answer, check that the question compiles.
Why it pays
Curiosity that has nowhere to be is where the unusual solutions come from.
The rule
Read outside your field on purpose. Analogy is the cheapest form of invention.
Notes in the margin — my reading of them, not their words.
Level 07 People
You can't make epic shit with basic people.
— but nice ones. Always nice ones.
High floor
Standards are the kindest thing on a team. Shipping something mediocre together costs everyone the same week.
Loose grip
Strong opinions, cheaply abandoned when the profiler disagrees. Being right later beats being right loudly.
Kind by default
Direct when it counts. Sharp code review, soft landing — the two were never in conflict.