==============================================================================================
  RAHUL'S ML BLOG -- notes on machine learning, worked out by hand                    est. 2026
==============================================================================================
  home | about | archive | glossary | contact
----------------------------------------------------------------------------------------------

  SPECIAL -- A READING LADDER
  Build a GPT, Forced: Each Page Pays a Prior Page's Bill
  ============================================================================================


  Several pages on this blog circle one machine -- a GPT -- and each stands alone. Read in
  this order they become one forced build: every rung exists because a machine on a rung
  below hit a wall, worked out in numbers, not waved at. Start with a writer that works;
  climb until even word order is solved.

      rung 1   a machine that writes at all        (one room)
      rung 2   depth, because one room is shallow  (twelve rooms)
      rung 3   a note, because redoing is waste    (KV cache)
      rung 4   a window, because lines outgrow grids
      rung 5   shapes, because plumbing eats days
      rung 6   order, because look-everywhere forgets where


  --------
  RUNG 1 -- A MACHINE THAT WRITES AT ALL

  A Writing Machine of One Room, Built by Pencil

  Smallest working writer: one room, six numbers per mark, ten marks known. Three turns run
  by pencil, every matrix drawn, and a line gets written: 7 -> 7 8 -> 7 8 1 -> 7 8 1 2.
  Wall it ends on: one room barely bends a strip.


  --------
  RUNG 2 -- DEPTH, BECAUSE ONE ROOM IS SHALLOW

  What a Machine Does When You Press "h" -- Every Matrix, Forced Into Being

  Twelve rooms stacked. A strip that starts with every number under 1.0 drifts past 12 and
  below -13, with output matrices and widen-bend-shrink stages doing all lifting at length
  one. A casual twin in plain English, ending on real GPT-2 where a lone "h" predicts a full
  stop: What Happens When You Press "h" in a GPT.
  Wall: every turn rebuilds every old row from a blank start.


  --------
  RUNG 3 -- A NOTE, BECAUSE REDOING OLD WORK IS WASTE

  A Note a Machine Keeps So It Stops Redoing Old Work -- KV Cache by Pencil

  A label-row and a payload-row -- two rows cast from each written mark by fixed matrices,
  one for matching and one for handing over -- never move once a mark is written, yet a no-memory machine
  rebuilds them every turn: builds pile up 1+2+3 = 6 where three rows exist, 125,250 where
  500 exist. Writing each row once on a note collapses a squared pile to a straight line.
  Wall: a note still grows with a line, and a full score grid grows with its square.


  --------
  RUNG 4 -- A WINDOW, BECAUSE A LINE OUTGROWS A GRID

  Sliding Window Attention by Pencil -- A Window That Drops Its Best Old Match

  Full attention on 100,000 marks costs five billion matches, so cap each mark's reach to a
  last W marks: a best old match holding a 64.1% share drops to 0.0% only for being old --
  and stacked rooms rebuild reach anyway, about two marks per room. Then build a whole
  masked multi-head layer from a blank page, every WHY derived by hand, root-D included:
  Sliding Window Self-Attention, Built From One Pencil.
  Wall: arithmetic was never hard -- days vanish in WHERE numbers sit.


  --------
  RUNG 5 -- SHAPES, BECAUSE PLUMBING EATS DAYS

  Attention's Shapes by Pencil -- Two Flips That Are Not One Flip

  One word, "transpose", names two different moves: a re-shelve hands each lens (one
  attention head) all of its tokens; an inside flip stands one grid on edge, (T,D) -> (D,T), a sole reason a want-grid
  against a stood-up label-grid lines up at all. Chunking, viewing, merging -- followed from
  a wide row to a finished row, every shape printed.
  Wall: with no walking and no order, nothing says WHERE a mark sits.


  --------
  RUNG 6 -- ORDER, BECAUSE LOOK-EVERYWHERE FORGETS WHERE

  Positional Encoding Built From Scratch -- A Mark That Tells a Transformer Where It Is

  Two naive position marks fail by arithmetic. Sinusoidal encoding built from scratch, every
  number computed; then RoPE, rotation derived by hand, with a full proof that absolute
  positions cancel out of a score and only a gap between marks remains. A ladder tops out
  here: a machine that writes, deep, thrifty, windowed, plumbed, and ordered.


  --------
  FOUNDATIONS SHELF -- ATTENTION ALONE, BEFORE ANY WHOLE MACHINE

  Rungs above assume attention itself feels familiar. If not, one of these first:

  - Attention by Pencil -- Every Word Looks at Every Word
    one word ("nolan") followed through every move, real numbers throughout.
  - Transformers With Pencil -- A Whole Block Worked by Hand
    one full block on two words, in and out, nothing waved at.
  - Transformer From Pencil -- a KATA
    a drill version, made to be redone on a blank sheet until a hand knows it.

  Read a rung, run its proof, climb.