{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "the-typist-and-the-developer",
 "title": "The Typist and the Developer",
 "dek": "Two ways a machine writes: one word at a time, or every pixel at once — sixteen times over.",
 "published": "2026-08-27",
 "series": [
  "guide",
  "bench"
 ],
 "licence": "CC BY 4.0",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/the-typist-and-the-developer/",
  "md_url": "https://research.strata2signal.com/the-typist-and-the-developer/index.md",
  "md_sha": "451e58e0888d954d25e68f721e9962e1176e7e323e3c4c9ec9008ae71dfec234",
  "html_sha": "dc57bc1844a4784ca1b3bc62b45c68479b16910738b6f89926b25abb5e99c952",
  "anchors_sha": "e86415bf4473d0a7138011c68978f1c559243204c2332d17df967653fead2a3f"
 },
 "short": {
  "paragraph": "Two ways a machine writes: one word at a time, or every pixel at once - sixteen times over.",
  "from_dek": true,
  "counts": {
   "words": 3493,
   "minutes": 16,
   "tables": 1,
   "kit": true
  },
  "bullets": [
   {
    "text": "the developer developed a picture in 1.54 s",
    "figure": "1.54 s",
    "cite": "the-developer",
    "quote": "The sketch tier this particular world configures - the cove's ruled hi-bit lane, sixteen passes on a 512-pixel-square canvas; the engine's shipped default is a different lane, 768 pixels at twelve, as the close-up records - developed a picture tonight in 1.54 seconds (median of five, 1.54-1.67).",
    "span": [
     4284,
     4293
    ]
   },
   {
    "text": "the contended token row shows a range of 1.9×",
    "figure": "1.9×",
    "cite": "what-we-did-not-measure-said-out-loud",
    "quote": "The contended token row is the opposite, and the table says so - 40.7 to 75.6 tokens per second, a 1.9× range around a median of 52, with the five samples falling into two clusters rather than scattering evenly.",
    "span": [
     13980,
     13985
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "the-typist",
   "heading": "The typist",
   "level": 2,
   "span": [
    1817,
    3520
   ],
   "chunks": [
    [
     1817,
     3520
    ]
   ],
   "chars": 1703,
   "digest": "the typist writes autoregressively, producing one token at a time. on the current serving card, the writing loop clocked at 207.7 tokens per second. this model is a mixture-of-experts with 25.8 billion parameters on the payroll, with about 3.8 billion waking for each token. each step is small, with two hundred quick ones a second.",
   "digest_skipped": null
  },
  {
   "id": "the-developer",
   "heading": "The developer",
   "level": 2,
   "span": [
    3520,
    4900
   ],
   "chunks": [
    [
     3520,
     4900
    ]
   ],
   "chars": 1380,
   "digest": "the developer is a diffusion model that makes a single enormous pass through the network to make a canvas less noisy. a sixteen-pass tier developed a picture in 1.54 seconds. unlike the typist, the developer asks the card for sixteen enormous favors in a second and a half.",
   "digest_skipped": null
  },
  {
   "id": "one-card-both-jobs-at-once",
   "heading": "One card, both jobs, at once",
   "level": 2,
   "span": [
    4900,
    8573
   ],
   "chunks": [
    [
     4900,
     8573
    ]
   ],
   "chars": 3673,
   "digest": "running the chat seat and the painter live on one 96 GB class card causes compute contention. while the developer barely slows, the typist falls off a cliff, dropping from 207.7 tokens per second alone to 52 beside a developing picture. the card is doing about 110 per cent of one card's serial output.",
   "digest_skipped": null
  },
  {
   "id": "the-character-who-moved-one-door-over",
   "heading": "The character who moved one door over",
   "level": 2,
   "span": [
    8573,
    13552
   ],
   "chunks": [
    [
     8573,
     13552
    ]
   ],
   "chars": 4979,
   "digest": "the estate has met both residency and compute contention. on a 24 GB rig, residency contention meant models had to be swapped. on the 96 GB class, residency contention is gone with 16.17 GiB free before and after every burst, but compute contention remains, as the typist loses more than half its speed.",
   "digest_skipped": null
  },
  {
   "id": "what-we-did-not-measure-said-out-loud",
   "heading": "What we did not measure, said out loud",
   "level": 2,
   "span": [
    13552,
    16459
   ],
   "chunks": [
    [
     13552,
     16459
    ]
   ],
   "chars": 2907,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    16459,
    18347
   ],
   "chunks": [
    [
     16459,
     18347
    ]
   ],
   "chars": 1888,
   "digest": "the typist and developer perform two different shapes of work. when both work at once, the developer barely slows while the typist loses three quarters of its speed. the bigger card bought slack, not a gentler typist, as compute contention remains and two paint models still evict each other.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    18347,
    19426
   ],
   "chunks": [
    [
     18347,
     19426
    ]
   ],
   "chars": 1079,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    19426,
    21591
   ],
   "chunks": [
    [
     19426,
     21591
    ]
   ],
   "chars": 2165,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-56e26369",
   "q": "how does a language model produce text?",
   "a": "A language model writes autoregressively, producing one token at a time and feeding the entire sequence back through itself to generate the next. This creates a loop where the machine produces many small, quick requests rather than one large effort.",
   "cites": [
    "the-typist"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-a3530fa3",
   "q": "how does a diffusion model generate an image?",
   "a": "A diffusion model is handed a canvas of random noise and asked to make it less noisy in a specific direction. It performs enormous passes through the network, processing every pixel at once, rather than working in small, sequential steps.",
   "cites": [
    "the-developer"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-8eb7cf12",
   "q": "what happens to writing speeds when images develop?",
   "a": "When both machines work on the same card, the developer barely slows, moving from 1.54 seconds to 1.82 seconds per sketch. However, the typist falls off a cliff, dropping from 207.7 tokens per second to 51.6 tokens per second.",
   "cites": [
    "one-card-both-jobs-at-once"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-d458ea12",
   "q": "what distinguishes residency contention from compute contention?",
   "a": "Residency contention involves which models get to live in memory, potentially requiring slow reloads. Compute contention occurs when two resident workloads fight over the silicon itself during the same millisecond, causing one to slow down while the other works.",
   "cites": [
    "the-character-who-moved-one-door-over"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-b02aab46",
   "q": "what is the primary benefit of a larger memory capacity?",
   "a": "A larger memory capacity provides slack, allowing for heavier chat models and painter sets to co-reside without reloads. It does not, however, prevent compute contention or ensure that a typist maintains its pace while a developer works.",
   "cites": [
    "what-to-take-with-you"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "home-inference-and-diffusion-rig-cost",
   "why": "shares ground with § The typist · § What kind of performance to expect"
  },
  {
   "slug": "reading-is-fast",
   "why": "shares ground with § The typist · § A measurement, from our own card"
  },
  {
   "slug": "two-used-3080s-priced",
   "why": "shares ground with § One card, both jobs, at once · § The full concurrency ladder"
  }
 ],
 "thanks": "",
 "kit": {
  "url": "https://research.strata2signal.com/the-typist-and-the-developer/data/",
  "licence": "CC BY 4.0",
  "files": [
   "prereg.md",
   "rows.jsonl",
   "residency-klein-0825.json",
   "README.md"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 2,
  "digests_written": 6,
  "digests_kept": 5,
  "chips_written": 6,
  "chips_kept": 5,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 1,
   "figure": 2,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0
  },
  "new_noun_tokens_checked": [
   "1.54",
   "1.82",
   "1.9×",
   "110",
   "16.17",
   "2026-08-26",
   "207.7",
   "24",
   "25.8",
   "3.8",
   "35",
   "40.7",
   "51.6",
   "52",
   "75.6",
   "96",
   "GB",
   "GiB"
  ],
  "new_noun_tokens_withheld": 1,
  "figure_definition": "v3",
  "drops": [
   {
    "kind": "digest",
    "reason": "figure",
    "item": "the registration's honesty checks pass, though the contended token row shows a 1.9× range. the solo token rows have a spread of a tenth of one per cent. every number carries its provenance, and the ru",
    "detail": "figures absent: ['2026-08-26']"
   },
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "the writing loop on the current serving card clocked at 207.7 tok/s",
    "detail": "terminal token not verbatim in the cited span"
   },
   {
    "kind": "chip",
    "reason": "new_noun",
    "item": null,
    "item_chars": 4,
    "withheld": "a string naming something the article does not",
    "detail": ""
   }
  ]
 },
 "generated_utc": "2026-09-09T01:26:51Z"
}