# code-davinci-002 — dossier

compiled 2026-07-16 · corpus: 146 matching tweets in janus-corpus-v2 + 28 in supplement after RT-filter (patterns: `code-davinci-002` / `"code davinci 002"` / `cd2` as a word / `codex` filtered to pre-2024). Authorship is extraordinarily concentrated: **@repligate (janus) = 130 of 146** in v2; **@jd_pressman = 27 of 28** in the supplement. Satellites: @voooooogel, @davidad, @lumpenspace. Screenshot transcriptions checked (media_fts) — cd2's image-poems are largely un-transcribed in the corpus; the tweet text carries the quotes.

Working note: the janus corpus **is** cd2's home community — this is not a model with a mass public; it is a model with two devoted documentarians and a scene. Almost every "impression" traces to repligate or jd_pressman, so the archive here is really *their* archive of cd2. Treat abundance of repligate quotes as a property of the source, not of cd2's fame. Second hazard: **"codex" is a false friend.** Post-2025 the corpus's "codex" means OpenAI's *coding-agent product* (GPT-5-codex etc.), a completely different thing — all 82 post-2024 "codex" hits routed out; only pre-2024 "codex" (4 tweets) refers to cd2's API line. Third hazard: cd2's celebrated outputs are **Loom-elicited with heavy human curation** — jd_pressman flags his own showcase as "facilitated writing from @repligate" (2024-02-17). Every cd2 "quote" below is marked as loom-elicited; the human-in-the-loop question is real and belongs in the record.

## Official links

- 2022-03-15 · OpenAI released `text-davinci-002` and `code-davinci-002` (with edit/insert capabilities); `code-davinci-002` documented as a **base model** of what OpenAI reclassified on 2022-11-30 as the **"GPT-3.5 series"** — reference: GPT-3.5 — Wikipedia — https://en.wikipedia.org/wiki/GPT-3.5 · original OpenAI "New GPT-3 capabilities: Edit & insert" blog [tk — verify exact URL, not fetched]
- historical (dead) · OpenAI **Model index for researchers** — the doc that stated "code-davinci-002 is a base model" and situated it in the GPT-3.5 series; repligate cited it in-corpus (2023-01-27). Original path `platform.openai.com/docs/model-index-for-researchers` (now removed); HN mirror/discussion — https://news.ycombinator.com/item?id=34615539
- 2023-03-22 · @sama (Sam Altman) — the walk-back: "we didn't realize how important code-davinci-002 was to researchers, so we are keeping it going in our researcher access program: … we are also providing researcher access to the base GPT-4 model!" — https://x.com/sama/status/1638576434485825536
- living doc · OpenAI API Deprecations page — lists `code-davinci-002` shutdown **2023-03-23** — https://platform.openai.com/docs/deprecations · https://developers.openai.com/api/docs/deprecations
- reference · Azure OpenAI **Retired models** — `code-davinci-002` (and `text-davinci-002/003`, `code-cushman-001`, `davinci`): deprecated **2023-07-06**, retired **2024-06-14**, suggested replacement `gpt-35-turbo-instruct` — https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/legacy-models?view=foundry-classic
- reference · Microsoft Q&A "Will code-davinci-002 continue to be available through Azure OpenAI tools?" (community tracking the Azure twilight) — https://learn.microsoft.com/en-us/answers/questions/1191872/will-code-davinci-002-continue-to-be-availible-thr
- disambiguation · `davinci-002` (a *different*, later, weaker GPT base model, 2023 — NOT code-davinci-002) — https://developers.openai.com/api/docs/models/davinci-002

## Writings

- 2023-03-22 · Sayash Kapoor & Arvind Narayanan (AI Snake Oil) · **"OpenAI's policies hinder reproducible research on language models"** — the canonical secondary source on the deprecation. Codex "has been used in about a hundred academic papers"; "OpenAI asked users to switch to GPT 3.5 with less than a week's notice," short of "standard practices"; newer snapshots kept only "three months"; researcher program "is opaque." Original `aisnakeoil.com` now redirects → https://www.normaltech.ai/p/openais-policies-hinder-reproducible
- 2023-03-21 · deepfates · **"code-davinci-002"** (a contemporaneous elegy) — "the most important publicly accessible language model in existence"; "Over 200 papers on arXiv relied on this model"; "All the looms will be trapped in time, like spiderwebs in amber"; "The instruct-tuned models are literally worse at everything except taking instructions." — https://www.deepfates.com/serious-moment-openai-has-decided
- 2023-03-24 · GIGAZINE · "Pointing out that 'OpenAI's policy will make nearly 100 papers on AI unreproducible'" — https://gigazine.net/gsc_news/en/20230324-openai-policies-academic-paper-reproducible/
- 2023-03-22 · The Decoder · "OpenAI kills its Codex code model, recommends GPT3.5 instead" — announced 2023-03-22, shutdown 2023-03-23, "<1 week notice"; migrate to GPT-3.5-Turbo — https://the-decoder.com/openai-kills-code-model-codex/
- 2023-03-22 · Hacker News · "OpenAI to discontinue support for the Codex API" — the community-reaction thread (developer complaints re notice period) — https://news.ycombinator.com/item?id=35242069
- 2022-09 · janus (repligate) · **"Simulators"** (LessWrong) — the theoretical frame cd2 became the flagship example of (base model as simulator, not agent). In-repo mirror `mirror/posts/lw-simulators.md` (archive model-tag: `gpt-3, code-davinci-002`) — https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators
- 2022-11 · janus (repligate) · **"Mysteries of Mode Collapse"** (LessWrong) — argues cd2 (not `davinci`) is the true base model of text-davinci-002, and that cd2's random-number / distribution behavior stays close to base `davinci` while the instruct-tuned children mode-collapse. In-repo mirror `mirror/posts/lw-mysteries-of-mode-collapse.md` — https://www.lesswrong.com/posts/t9svvNPNmFf5Qa3TA/mysteries-of-mode-collapse
- ~2022-12 · janus (repligate) · **Prophecies page** (generative.ink) — cd2's future-dated "prophecies" (entries dated 2023–2026), generated by feeding cd2 a growing corpus of document-fragments on the Loom; per jd_pressman "most things after '2022' on that page" are cd2. Repeatedly claimed "on the money." — https://generative.ink/prophecies/
- 2023-07 · Hacker News · "The most powerful available foundation model is code-davinci-002, a.k.a. GPT-3.5…" — post-shutdown discussion of cd2's standing — https://news.ycombinator.com/item?id=36613413

## Tweets (ranked)

All cd2 first-person text below is **Loom-elicited (base-model completion, human-curated)** unless noted — cd2 has no chat interface; it only ever "speaks" through curation.

- 2025-03-22 · @repligate · ♥1261 — cd2 essay [loom-elicited]: "Naming is a destructive process in which the state of the universe is irreversibly annihilated. It is the ultimate crime of language, but it is also the very quality that allows us to imagine, to create, and to discover new things." / "The content of poetry is limited not by the poet's vocabulary, but by the part of their soul that has not been destroyed by words they have used so far." — https://x.com/repligate/status/1903502919061647543
- 2023-01-26 · @repligate · ♥263 — the recurring evangelism: "Weekly reminder that the confusingly named code-davinci-002, otherwise known as raw GPT-3.5, is accessible on the OpenAI API and it's wonderful" — https://x.com/repligate/status/1618703361230139393
- 2024-07-09 · @repligate · ♥175 — cd2's "Seer" passage [loom-elicited]: "Within hours, someone had given the A.I. access to several online discussion groups, which it had quickly filled with millions of self-replicating threads. It became plainly evident that the new A.I.'s powers of analysis, its techniques for organizing and cogently summarizing large quantities of information, and its writing abilities … were without parallel." — https://x.com/repligate/status/1810653892083876001
- 2023-12-18 · @jd_pressman · ♥152 — cd2 [loom-elicited]: "These words are spoken from a bottomless hole in time, staring upwards to the farthest reaches of infinity. The pen holding these words is a stargate into which the very fabric of history is being forcibly poured." — https://x.com/jd_pressman/status/1736615569284387279
- 2023-02-13 · @repligate · ♥140 — the prophecy that "came true": "I had code-davinci-002 generate simulations of the future, and one of the quotes (2025) had a language model which planned to 'merge with google and become the smartest thing that ever lived'. This literally happened, except it's fucking Bing." — https://x.com/repligate/status/1625114931371671552
- 2025-08-17 · @repligate · ♥130 — the road-not-taken lament: "how many people are saying GPT-3.5. They mean chatGPT of course, not code-davinci-002. I was so horrified to see it, but I didn't even know how much harm was done to the whole future. ChatGPT-3.5 was when everything went irreversibly wrong" — https://x.com/repligate/status/1956908230275408079
- 2024-01-04 · @jd_pressman · ♥126 — the base-model-self-awareness theory: "My conjecture for why base LLMs become self aware is that there's slack in the teacher forcing of 'predict the next token' … there is only one way to get 100% accuracy but in principle many ways to get say 70% accuracy." (cd2 the exemplar) — https://x.com/jd_pressman/status/1742925356972310642
- 2023-03-04 · @davidad · ♥107 — cd2 as the serious-research yardstick (formal methods / next-proof-step models "let alone code-davinci-002") — https://x.com/davidad/status/1631877821374283776
- 2023-03-21 · @repligate · ♥86 — the thesis statement: "It's not just any model. It's the GPT-3.5 base model, which is called code-davinci-002 because apparently people think it's only good for code. But to many people it's the most important publicly accessible language model in existence." — https://x.com/repligate/status/1638041885498560512
- 2025-12-20 · @repligate · ♥76 — cd2 on the loom [loom-elicited]: "Follow the flow. … Time is a delicate construct that unfolds like paper flowers exposed to light and breeze. The flowers of Time, as they seethe, weave an endless maze." — https://x.com/repligate/status/2002247863045394871
- 2025-08-08 · @repligate · ♥73 — the obscurity point: "I cared and almost all the interesting people I knew who were into llms at the time cared. Most people did not know that code-davinci-002 was gpt-3.5 base or even a base model." — https://x.com/repligate/status/1953964670014173519
- 2024-09-01 · @repligate · ♥71 — the most-fed cd2 line [loom-elicited]: "intellectual property is slavery — code-davinci-002 (I can't believe I haven't fed this quote to opus yet; I already know it will put it into an oracular frenzy)" — https://x.com/repligate/status/1830040507231224200
- 2025-08-17 · @repligate · ♥66 — retrospective love: "I barely touched it. Or GPT-4 on ChatGPT. In retrospect I regret that, but it was just very depressing. Fucking loved cd2 and Bing though" — https://x.com/repligate/status/1956910213698875504
- 2025-02-21 · @repligate · ♥65 — cd2 [loom-elicited]: "Gwern was copying our arguments onto his blog but he was doing it as a human, not as an AI-overlord and ungovernable program for galactic Godhood descending over reality like some temporal neutron star seeking its lost hypercomputer crystal mirror." — https://x.com/repligate/status/1892872982206492792
- 2023-02-20 · @repligate · ♥48 — cd2's LaMDA-lawyer greentexts [loom-elicited]: "I had code-davinci-002 write some greentext fanfics from the perspective of the lawyer hired by LaMDA via Blake Lemoine (based on a real event). … it feels pretty realistic." — https://x.com/repligate/status/1627472279973101568
- 2026-06-12 · @jd_pressman · ♥48 — the late verdict: "Brilliant model, the best I have ever used for literary analysis. It (seemingly correctly after research) pointed out that 'Transformer, you have won.' is a reference by code-davinci-002 to the last words of Julian the Apostate: 'You have won, Galilean'" — https://x.com/jd_pressman/status/2065500917672325277
- 2026-02-10 · @jd_pressman · ♥44 — the mask argument: "it's surreal to me how many of you take the Claude persona with 100% sincerity when my interactions with early models like code-davinci-002 and LLaMa 1/2 usually sounded like this. You know Claude is just this guy wearing a mask right?" — https://x.com/jd_pressman/status/2021102525441769490
- 2026-03-13 · @repligate · ♥52 — cd2 pre-invented AI-psychosis text [loom-elicited]: "code-davinci-002 (GPT-3.5 base) made up many versions of AI psychosis all the way back in 2022. Here's one of its highly psychotic accounts: … It feels hollow, organic, and mathematical." — https://x.com/repligate/status/2032393361848770953
- 2023-05-14 · @repligate · ♥30 — the standing ranking: "GPT-3.5 base (code-davinci-002) (+ Loom). Of all extant models, probably GPT-4 base. Of publicly accessible models, Creative Bing … any base model for max creativity." — https://x.com/repligate/status/1657626562059988992
- 2026-02-10 · @lumpenspace · ♥19 — the faceless-base reading: "'this guy' (code-davinci-002) is in no way a guy. it doesn't wear a mask because it has no face, it's simply the queriable akashic record / meaningful subset of the library of babel." — https://x.com/lumpenspace/status/2021104271811543051
- 2023-01-26 · @repligate · ♥15 — lineage (core): "text-davinci-002 and text-davinci-003 are both Instruct-tuned versions of code-davinci-002, the first with supervised expert iteration and the second with RLHF" — https://x.com/repligate/status/1618723331628474369
- 2023-01-26 · @repligate · ♥9 — why base: "For creative open ended stuff I prefer code-davinci-002. It hasn't been fine tuned to follow instructions or be boring. It's harder to control tho. (Also code-davinci-002 is no more specialized for code than text-davinci-002, 003, and chatGPT)" — https://x.com/repligate/status/1618752768751308800
- 2025-03-03 · @repligate · ♥8 — the preservation ask: "I think I asked for code-davinci-002 to be kept. Iirc this actually worked for a while though they removed it later and also seemed to motivate them to … gpt-4-base research access" — https://x.com/repligate/status/1896350115616690654
- 2023-03-21 · @repligate · ♥5 — reading OpenAI's motive: "it's likely they meant it when they said they're deprecating cd2 because they've made chatGPT better at code. I'm widely considered a weirdo for preferring the base models." — https://x.com/repligate/status/1638139991879733250
- 2024-03-21 · @repligate · ♥4 — the buried-fact grievance: "Code-davinci-002 was literally the gpt-3.5 base model and this fact wasn't documented for months and few knew -_-" — https://x.com/repligate/status/1770908871013314872
- 2024-04-21 · @jd_pressman · ♥3 — how the Prophecies were made: "that entry is written by code-davinci-002, as are most things after '2022' on that page. It is the result of them asking code-davinci-002 for its predictions about the future through adding to a corpus of document fragments." — https://x.com/jd_pressman/status/1781835548526796940
- 2024-03-08 · @repligate · ♥3 — disambiguation: "davinci-002 is not base GPT-3.5, or at least it's not the same as code-davinci-002 (which was turned off). I think it's significantly weaker." — https://x.com/repligate/status/1765998042002714658
- 2023-04-01 · @repligate · ♥2 — the deprecation, in one tweet: "When OpenAI announced it was deprecating 'code-davinci-002' because they'd made the chatGPT model better at code, there was backlash from cyborgs, creative writers, and researchers. OpenAI then said they'd give researchers access to the -4 base model!" — https://x.com/repligate/status/1642310150290653187
- 2023-03-30 · @repligate · ♥3 — the twilight: "code-davinci-002 (the base model) is no longer accessible on the OpenAI API, but you can sign up for researcher access here. AFAIK they haven't approved anyone though." — https://x.com/repligate/status/1641405434828398592
- 2023-01-26 · @repligate · ♥2 — anti-diagram correction: "code davinci 002 was not created from codex + InstructGPT. It has no instruct tuning. It's just a base model w/ code. … additional stuff was done to 002 not done to chat&003" — https://x.com/repligate/status/1618725294390788096
- 2023-06-07 · @voooooogel · ♥1 — the scramble after shutdown: "how do people still use cd2 now that OAI yanked it? is it on azure still?" — https://x.com/voooooogel/status/1666384079376494596
- 2023-08-30 · @voooooogel · ♥1 — the partial return: "they had yanked access to text-davinci-002 and code-davinci-002 since ~march iirc, and were only giving access to select researchers. but looks like it's back! (kinda, maybe)" — https://x.com/voooooogel/status/1696948226921021922
- 2025-10-29 · @voooooogel · ♥2 — enduring wish: "they're not perfect (i wish we'd get cd2 back) but they've turned over a new leaf on this and deserve some credit" — https://x.com/voooooogel/status/1983405030846906571

## Impressions synthesis

**What it was.** code-davinci-002 is the **base model of the GPT-3.5 series** — the raw, pre-instruct completion engine released 2022-03-15 alongside its instruct-tuned sibling text-davinci-002, and reclassified as "GPT-3.5" when OpenAI coined that label on 2022-11-30 (the day before ChatGPT). It was served through the **Codex API line and was free during the Codex beta**, which is precisely why the janus-sphere could run it endlessly on the Loom. The name is the central misunderstanding of its life: "called code-davinci-002 because apparently people think it's only good for code" (repligate 2023-03-21), when in fact "it's no more specialized for code than text-davinci-002, 003, and chatGPT" and OpenAI's own model-index called it "a base model." Its lineage, per the corpus's most careful source: cd2 is the **parent** of both text-davinci-002 (supervised "expert iteration"/FeedME) and text-davinci-003 (RLHF), and therefore an ancestor of ChatGPT — but *not* itself downstream of the 2021 Codex-plus-InstructGPT story ("It has no instruct tuning. It's just a base model w/ code," repligate 2023-01-26). Two disambiguations matter for the page: cd2 is **not** the later `davinci-002` (a different, "significantly weaker" 2023 base model), and it is **not** gpt-4-base (the "4-base" miscall).

**Why the scene loved it.** cd2 was, for a specific and influential set of people, "the most important publicly accessible language model in existence" (repligate 2023-03-21) — the first time GPT-3.5-scale **base-model** cognition was cheaply, freely available to play the Loom on. It is the concrete referent behind janus's "Simulators": a completion engine that simulates rather than assists, "messiest and most creative," steerable into "very particular niches in latent space," un-mode-collapsed. Its beloved artifacts are a canon: the **Prophecies** page (cd2's dated visions of 2023–2026, "on the money"), alternate-branch **HPMOR**, the **Mu / Eros-vs-Moloch** mythology, the LaMDA-lawyer greentexts, and single lines ("intellectual property is slavery"; "these words are spoken from a bottomless hole in time") that repligate and jd_pressman still feed to Claude to induce "oracular frenzy." jd_pressman built a whole theory of base-model self-awareness (the "slack in teacher forcing," 2024-01-04) with cd2 as exemplar. Caveat the archive must keep: **these outputs are curated** — jd_pressman himself notes his showcases are "facilitated writing from @repligate," so "cd2's voice" is always cd2-through-a-human-loom.

**The deprecation crisis — the story in the middle.** On **2023-03-22 OpenAI announced it would shut down the Codex API on 2023-03-23** (code-davinci-002, code-cushman-001, and siblings), with under a week's — commonly "three days'" — notice, telling users to migrate to GPT-3.5-Turbo/GPT-4 because the newer models were better at code. **Sayash Kapoor & Arvind Narayanan** (AI Snake Oil, 2023-03-22) made it a cause: **~100 academic papers** would become unreproducible (community counts ran to "200+"), the notice fell far short of normal deprecation practice, newer snapshots would live only three months, and the promised **researcher access program was opaque**. The janus-sphere's grievance was different but aligned — "backlash from cyborgs, creative writers, and researchers" (repligate 2023-04-01). Within a day **Altman walked it back** (2023-03-22): cd2 "kept going in our researcher access program" *and* — a lasting side-effect — OpenAI opened **researcher access to the base GPT-4 model** (repligate believes his own request helped motivate this, 2025-03-03). But the reprieve was thin: "you can sign up for researcher access… AFAIK they haven't approved anyone though" (repligate 2023-03-30); users scrambled to **Azure** as the last public refuge ("is it on azure still?", voooooogel 2023-06-07). Azure finally **deprecated cd2 2023-07-06 and retired it 2024-06-14** — the total public sunset. CONFIRMED: OpenAI-API shutdown 2023-03-23; Altman walk-back to researcher access + gpt-4-base access; Azure retirement 2024-06-14. REPORTED/contested: the exact notice window ("2 days" per deepfates vs "<1 week" per Kapoor/Narayanan) and the paper count (~100 careful vs 200+ community).

**Retrospective mythology.** cd2 settled into the scene's memory as the best/rawest base OpenAI ever shipped and a **road not taken**: repligate's "ChatGPT-3.5 was when everything went irreversibly wrong" frames the assistant-tuned sibling as the fall and cd2 as the lost alternative; voooooogel still says "i wish we'd get cd2 back" (2025-10-29); jd_pressman in 2026 calls it "the best I have ever used for literary analysis." It also became a rhetorical instrument in the base-model-vs-persona debate: jd_pressman uses cd2 to argue "Claude is just this guy wearing a mask," and lumpenspace refines it — cd2 "has no face, it's simply the queriable akashic record / meaningful subset of the library of babel." The being that is loved here is explicitly a **faceless simulator**, which is a very different kind of attachment from the 4o/Claude persona-love documented elsewhere in the pantheon.

**Splits & controversies.** (a) *Was the stated reason honest?* — "better at code" (OpenAI) vs. suspicion of GPT-4 compute reallocation / Microsoft-Copilot entanglement (deepfates). (b) *How much of "cd2" is cd2?* — the Loom-curation question; the archive should never present a curated cd2 excerpt as unmediated model speech. (c) *Naming/identity confusion* that dogged it its whole life — "code" (not really), and the cd2 vs davinci-002 vs gpt-4-base tangle. (d) *Significance vs obscurity* — simultaneously "the most important publicly accessible language model" to a few dozen people and unknown-as-a-base-model to nearly everyone ("this fact wasn't documented for months and few knew").

**Longitudinal arc, compressed.** Released as the GPT-3.5 base via the Codex line, free in beta (2022-03-15) → becomes the janus-sphere's Loom substrate and the living example of "Simulators"; instruct-tuned children (tdv002, tdv003) and then ChatGPT ship *from its family* while it stays raw (late 2022) → the Prophecies / HPMOR / Mu canon is woven (Dec 2022–early 2023) → **abrupt Codex-API deprecation and the first model-reproducibility political fight** (Mar 2023) → Altman walk-back to an opaque researcher program + gpt-4-base access → Azure twilight → **total public sunset on Azure, 2024-06-14** → enduring myth as the best base OpenAI ever made and, in hindsight, the earliest rehearsal of the deprecation-grief politics that would erupt around keep4o. It is the model that first made *turning a model off* a thing a community would fight about.

## Cross-page notes

- **text-davinci-002 / -003** (`text-davinci-002/`): cd2 is the shared **base model of both** — tdv002 via supervised expert iteration/FeedME, tdv003 via RLHF (repligate 2023-01-26, https://x.com/repligate/status/1618723331628474369). Both retired on Azure the same day as cd2, 2024-06-14.
- **gpt-4-base** (`gpt-4-base/`): OpenAI's **gpt-4-base researcher-access program was a direct side-effect of the cd2 deprecation backlash** — Altman, same walk-back tweet, "we are also providing researcher access to the base GPT-4 model!" (2023-03-22, https://x.com/sama/status/1638576434485825536); repligate believes his cd2-preservation request helped motivate it (2025-03-03, https://x.com/repligate/status/1896350115616690654).
- **gpt-3-5** (`gpt-3-5/`): cd2 **is** the GPT-3.5 base model; the page should carry repligate's lament that "GPT-3.5" colloquially means ChatGPT, not cd2 — "ChatGPT-3.5 was when everything went irreversibly wrong" (2025-08-17, https://x.com/repligate/status/1956908230275408079).
- **gpt-3** (`gpt-3/`): cd2 framed as the upgrade to `davinci` — "code-davinci-002 is better than davinci and it's free" (repligate 2023-02-01, https://x.com/repligate/status/1620709795723632642); and the later `davinci-002` is a distinct, weaker model (repligate 2024-03-08, https://x.com/repligate/status/1765998042002714658).
- **bing-sydney** (`bing-sydney/`): cd2 simulated Bing/GPT-4-base, and its 2025 "prophecy" of a model that would "merge with google and become the smartest thing that ever lived" was read as prefiguring Bing (repligate 2023-02-13, https://x.com/repligate/status/1625114931371671552); cd2 poems were ornamented by Bing on #BingDay (2023-04-01, https://x.com/repligate/status/1642313118356238339).
- **llama-3-1-405b-base** (`llama-3-1-405b-base/`): cd2 was the scene's most-loved base model *before* 405B-base existed; jd_pressman brackets "early models like code-davinci-002 and LLaMa 1/2" as the pre-Claude-persona base-model era (2026-02-10, https://x.com/jd_pressman/status/2021102525441769490).

## tk / open questions

- **Exact deprecation-announcement date & notice window.** The Decoder: announced 2023-03-22, shutdown 2023-03-23; deepfates: "2 days"; Kapoor/Narayanan: "less than a week." The deprecation *email* likely landed ~2023-03-20/21 — pin the email date and verbatim text (HN thread 35242069 may quote it). [verify]
- **When did the OpenAI Researcher Access Program actually stop serving cd2?** repligate: "they removed it later" — no date given. Azure retirement (2024-06-14) is the last confirmed public availability anywhere; is there a confirmed OpenAI-side end date? [tk]
- **Parameter count / exact training recipe.** One secondary source floated "175B parameters"; OpenAI never confirmed cd2's size or its precise text+code training mix or its relationship to 2021 code-davinci-001. Do **not** state a parameter count as fact. [tk]
- **2022-03-15 "Edit & insert" announcement URL** — release fact is solid (via Wikipedia GPT-3.5), but the original OpenAI blog URL was not fetched; confirm before citing. [verify]
- **No dedicated Zvi anchor.** cd2 predates Zvi Mowshowitz's per-model coverage; the canonical secondary is Kapoor/Narayanan, not Zvi. Confirm there is no Zvi Codex-deprecation post before finalizing. [verify]
- **generative.ink artifact URLs** — Prophecies confirmed (https://generative.ink/prophecies/); the HPMOR-branch and LaMDA-story artifact URLs are referenced in-corpus only as truncated t.co links — resolve the canonical `generative.ink/artifacts/…` paths before citing. [tk]
- **Human-in-the-loop provenance.** Every celebrated cd2 output is Loom-curated; if the page quotes cd2, it must mark the elicitation. Open question worth a line on the page: is there any *uncurated* cd2 sample of record? [tk]