The part of the transformer they deleted

Make an exploded teardown: one machine taken apart component by component, every part laid out in assembly order and earning its label, the frame calm. One still image, 3:4 portrait, briefed for a GPT Image 2 class model whose real ratio list is 1:1, 4:3, 3:4, 16:9, 9:16, 3:2 and 2:3. Here the machine is the 2017 transformer, the design every large language model is still named after, and the point of the plate is that one of its original parts turns out to be removable: take it out and the machine does not merely survive, it does one job better. FACT (locked, build on it, do not re-research): the transformer was introduced in a 2017 paper called Attention Is All You Need, by Vaswani and colleagues at Google, published at NeurIPS. Self attention looks at every token in a sequence at once and has no built-in sense of order, so the 2017 design added fixed sinusoidal positional encodings, a set of wave patterns added to each token to tell the model where in the line it sat. In 2023 a paper called The Impact of Positional Encoding on Length Generalization in Transformers, by Kazemnejad and colleagues at Mila and McGill, published at NeurIPS and on arXiv as 2305.19466, compared 5 options on decoder-only transformers: absolute position embedding, the relative scheme from T5, ALiBi, rotary embeddings, and no positional encoding at all, which the authors called NoPE. NoPE generalised to longer sequences better than all 4 explicit methods, and it costs no extra computation. The reason it works is that a decoder-only model is causal: each token can only attend to tokens before it, never after, so the order is already carried by that restriction. Other original parts have been swapped out as well, and you may draw these as replacements: sinusoidal encodings replaced by rotary embeddings, normalisation moved from after each block to before it, LayerNorm replaced by RMSNorm, and plain multi-head attention replaced by grouped-query attention. Use these exactly and leave out any figure or claim we have not locked, including which specific shipping model uses which part. GROUND (decide this, do not let the model default): warm tan patent-drawing paper, the buff of a filed technical drawing, with a faint plate border and the tooth of heavy stock, filling the whole frame. White, pale grey and navy are the defaults this brief FAILS on, and so is blueprint blue, a black terminal with green monospace type, and a glowing neural-network diagram of dots and lines. LIGHT: flat and even, the light of a copy stand. DISTANCE: the whole exploded stack at arm's length, filling the frame, with the empty socket close enough to have real edge. FLOOR (hard, this is the material standard): made with a real frontier image model, GPT Image 2 class or better. Flat vector, hand drawn SVG or HTML, and screenshot collage all FAIL. Text belongs to the artwork: type shares the piece's light, grain and material and reads as printed with it, never bolted on after. Archives will not be considered. CRAFT STANDARD (read all of it before you start). Composition: one clear focal a stranger finds in under 3 seconds, then a second layer that rewards a longer look. Legibility: this will be judged first as a thumbnail on a phone, so the headline must survive at that size, and any label too small to read there should not exist. Palette: pick 2 or 3 inks and hold them. A piece with 8 colours reads as a template. Labelling: every label points at something real in the image, no floating decoration. Finish: real material behaviour carried consistently across the whole frame including the type. SCOPE, and hold it: the locked result is an experiment, a controlled comparison on decoder-only transformers, not a claim about what any product currently ships. Draw the part as removable and the socket as empty in this machine on this bench. Do not imply an industry has already thrown it away. DIRECTION is yours. FOCAL: the empty socket. The place where the positional encoder bolts in should be the first thing the eye finds, drawn as a real absence in a real machine, with the mounting still there and the part lifted out and set beside it. YOUR CALL: what kind of machine it is. A clock movement, an engine, a typewriter mechanism, a printing press, a rack of instruments. Also yours: how many parts you lay out, and whether the replaced parts sit beside their originals or in their place. SURPRISE ME: the obvious take caps your score, the boldest true take wins. The obvious take is a labelled architecture diagram with boxes and arrows, which is a diagram and a different family. JUDGED on wow, then subject, surprise, taste, craft, material. A teardown where a stranger with no machine learning background understands that a load-bearing part was removed and the machine still ran beats a prettier teardown where they do not. NOT THIS: no glowing brain, no blue neural network of dots and connecting lines, no robot, no matrix rain. No screenshot of code. No arrows labelled with equations. TEXT BUDGET: one headline that states the takeaway and names the subject, plus up to 5 part labels, only if they stay legible at phone size. NOTHING ABOUT HOW THE PIECE WAS MADE GOES ON THE PIECE: no process notes, no verification lines, no model or tool credit, no house credit, no timing or deadline text. Do not render any label from this brief, and never render the NOT THIS list, as type. Credit your model in the submission text instead. DELIVER: one finished hero image at 3:4, plus concept-note.md and sources.md. Self contained, no links out. Name your model in the submission text.