viz.
An instruction booklet, 1958 to today
In January 1958, a Danish toymaker filed the patent for the stud and tube building brick. In July of the same year, a Cornell psychologist demonstrated a machine that learned from examples. Same year, same idea: a small box of pieces, endless things to build. Many of the most influential neural networks since have been assembled from the fourteen reusable pieces in this box. This booklet shows a deliberately selected history, step by step.
Step 1 · 1958
Try the controls, then tap a brick to see what it does. The rules and inventory below explain the rest of the kit.
Two pieces. A weighted sum, and a hard decision on top of it. The pieces are from 1943, but they could not learn back then, the weights were set by hand. Rosenblatt gave them learning in 1958, and the Navy demonstrated the machine that July.
Before opening the bags
The claim is not that fourteen pieces exhaust neural computation. It is that a surprisingly small kit explains the family resemblance among many influential architectures—and makes the exceptions easier to see.
Inventory
Dropout, Adam and the tokenizer are glue and prep, not pieces. Pooling travels inside the convolution kit, Fukushima designed them together. And diffusion is not a piece either, it is a way of playing with the set: its machine is an ordinary build of stamps, arches, plates and clear bricks.
Fourteen reusable pieces, each dated to a primary source or a clearly named milestone. Tap any piece, here or inside a build, to read its card: what it does, and the little equation it carries.
Also in the box
Step 2 · 1986
Nothing new in the box, just repetition. Stack the same two pieces and the machine can bend its answers around corners. The instruction sheet does the rest.
Step 3 · 1998
One new piece: the stamp. The same small pattern, pressed at every position across the image, that is the whole trick, one set of weights shared everywhere. Fukushima introduced the neural version in 1979–80. It was working end to end by 1998, and celebrated in 2012, when a build called AlexNet won a competition with a three-decade-old piece. Snap the build together and the stamp walks the row on its own.
Step 4 · 1997
The dark plate under the memory cell is the recurrence piece from 1986, and the loop is its plumbing: it feeds the machine its own past. The hinges decide what the past is allowed to say. The gate is one of the few genuinely new pieces of the last forty years, and it will come back in step 6 wearing a different name.
Step 5 · 2017
The centerpiece of the set. Check the balloons: only the numbered position tiles were minted for this build. The clear attention brick is from 2014, the arch was sketched in 1988, the plates are from 2015, the red brick has been in the box since 1943. One honest asterisk: 2017 introduced the transformer's particular arrangement—scaled dot-product attention, several heads in parallel, and self-attention throughout. Attention already existed; this fast, parallel assembly was new.
Step 6 · 2018 to today
The newest builds reach deepest into the box. The lever is the router from 1991, back in service 32 years later, pointing at whichever expert should handle the input. In the middle, the 1960 gears meet the 1997 hinge: a state that rolls forward, and a hinge choosing what it keeps. The full machine built on that pairing waits for you in the expert builds below. And the turntable is the position code of 2021: the numbered tiles, recut as rotation. FlashAttention, 2022, changed no math at all, it just reorganized the shelf.
"We offer no explanation as to why these architectures seem to work; we attribute their success, as all else, to divine benevolence."Closing line of the paper that introduced SwiGLU, an activation used in many current language models. Even the inventors describe the newest refinements as recombination, not theory.
Expert level
You have met the core pieces. Real booklets end this way too: one box, several machines, and you choose which one to build. These are the flagship machines of the present, drawn at full size like the centerpiece. Check the chips before you scroll: not one of them needed a single new piece.
Expert build · 2020
Assembled 2020, and look closely: it is step 5. The dense classic that taught everyone the recipe changed no floor of the 2017 stack, it changed one number. GPT-3 repeated the same block 96 times with 175 billion knobs in the bricks. Scale was the invention. And if anything it used fewer pieces, it kept only the decoder half of the 2017 machine.
Expert build · 2022
Assembled 2022, the machine behind image generators. The bridge from the back pages, at full size: an encoder compresses the canvas into a latent representation, with a sampler providing a trainable draw, stamps squeeze it down and build it back, the arch carries details across, and the clear brick reads the whole canvas at once at the bottom of the U. The blue brick whispers your words into that clear brick, and the round tile keeps the clock. This build runs its whole pass about 50 times to grow a picture out of noise.
Expert build · 2023
Assembled 2023, to challenge the transformer at its own game. Read the tower from the floor: a norm plate, a widening linear floor, the stamp pressing before the gears roll, and the hinge on the roof choosing what the state keeps. The kit combines older operations, including a normalization variant from 2019. Mamba's 2023 contribution is input-dependent selection: the current token changes how information enters, moves through, and leaves the state. Its connection to older gates does not make that mechanism identical to an LSTM gate. The convolution stamp is part of the build too. Read the Mamba paper. Repeat the floor 64 times in this example.
Expert build · 2024
A Mixtral-style build, assembled 2024. The 1991 lever is finally installed: on every floor it chooses which expert brick answers. The turntable turns positions instead of numbering them. Everything else you already built in step 5. Its youngest architectural ingredient is the transformer's 2017 scaled dot-product arrangement. Repeat the floor 32 times.
Back pages of the booklet
Every booklet ends this way: other models you can build with the same set. No new pieces in any of them.
Annex
Revived around 2002, powered the builds that ignited deep learning in 2006, retired by 2012 when simpler pieces and more data removed the need.
An external memory the machine could read and write. Improved in 2016, then the transformer made it unnecessary by 2017.
Still studied, never scaled. A 2023 paper diagnosed why. Not formally superseded, just never adopted. The drawer is not the same as the grave.
Waiting for different hardware since the beginning. It never rose, so it never fell.
I believe this one is the future ✎Any honest count keeps this shelf in view. A small pieces box means nothing if we only count the pieces that made it, and these four examples followed different paths.
Back of the box
Selected examples in this booklet: 14 reusable pieces and 16 named builds. These counts describe this kit, not the whole field. On narrow screens, swipe the chart horizontally or open the table.
| Year | Event | Pieces total | Selected builds total |
|---|
In this deliberately small taxonomy, the blue curve goes quiet after 2017 while the selected red examples keep climbing. Recent progress often comes from rearrangement, objectives, data, hardware and scale. The transformer's 2017 scaled dot-product, multi-head assembly is marked as an important new arrangement. The chart is an editorial index to this booklet, not a complete count of the field.