# text-davinci-002 / -003 — dossier

compiled 2026-07-16 · corpus: v2 — 24 tweets match `"text-davinci-002"`, 22 match `"text-davinci-003"`, 27 match `"davinci-003"`, 7 match `instructgpt`, 40 match the loose `"text davinci"` (heavily overlapping; the union is ~50 distinct after routing out `code-davinci-002`, which dominates the shared `davinci-002` FTS bucket at 174). supplement — 5 `text-davinci` hits (jd_pressman + Shoalst0ne). No media transcriptions match. patterns used: `"text-davinci-002"` / `"text-davinci-003"` / `"davinci-003"` / `instructgpt` / `"text davinci"` (FTS5 phrase-quoted for the hyphens) + LIKE-scan on the supplement db.

Working note: the Writings layer carries this page, exactly as the brief predicted. These are the models the simulator-era canon was *written about* — janus's *Mysteries of mode collapse* (Nov 2022) is the load-bearing primary text and is already mirrored in-repo (`mirror/posts/lw-mysteries-of-mode-collapse.md`, tagged `text-davinci-002`). The tweet corpus is thin, technical, and dominated by two voices: **@repligate** (provenance corrections, mode-collapse/attractor observations, glitch-token behavior) and **@davidad** (td3 as the capabilities/timelines inflection point). Very little "character" discourse of the kind 4o or the Claudes attract — these models were studied as *specimens of what tuning does to a mind*, not loved as personalities. The single most important disambiguation task here: `code-davinci-002` (the shared GPT-3.5 base, its own page) floods every `davinci-002` search and must be routed out; bare `davinci-002` (the 2023 base-GPT-3 API retrain) belongs to the GPT-3 page. See Cross-page notes.

## Official links

- 2022-01-27 · OpenAI · Aligning language models to follow instructions (the InstructGPT deployment blog; `text-davinci-001` becomes an API default). Contains the sidenote janus leans on: *"The InstructGPT models deployed in the API are updated versions trained using the same human feedback data. They use a similar but slightly different training method that we will describe in a forthcoming publication."* — https://openai.com/blog/instruction-following/ (URL as cited verbatim inside the mirrored mode-collapse post)
- 2022-03-04 · InstructGPT paper · Training language models to follow instructions with human feedback (Ouyang et al.; the RLHF lineage — but note the deployed `text-davinci-001/002` used the "similar but slightly different" method, not the paper's exact PPO-RLHF) — https://arxiv.org/abs/2203.02155
- 2022 (mid; exact release date poorly documented — see tk) · `text-davinci-002` released as part of the GPT-3.5 series: an InstructGPT model based on `code-davinci-002`, trained by **FeedME** (supervised fine-tuning on human-written demonstrations *and* on model samples rated highly by labelers), **not RLHF** — per OpenAI's "Model index for researchers" doc [tk — page deprecated; retrieve exact FeedME wording via web.archive.org]
- 2022-11-28 · `text-davinci-003` released — the RLHF member of the family; dropped ~2 days before the ChatGPT interface (2022-11-30) [CONFIRMED — date consistent across contemporary coverage and davidad's contemporaneous framing]
- 2023-07-06 · OpenAI · deprecation announced for `text-davinci-001` / `-002` / `-003` (and the wider InstructGPT/Completions legacy set), recommended replacement `gpt-3.5-turbo-instruct` — https://developers.openai.com/api/docs/deprecations
- 2024-01-04 · shutdown date — the text-davinci models removed from the API [CONFIRMED via the deprecations page]
- coda · `gpt-3.5-turbo-instruct` (the replacement) was itself announced for deprecation 2025-09-26, shutdown 2026-09-28 (replacement `gpt-5.4-mini` / `gpt-5-mini`) — the completions-style instruct line is ending entirely — https://developers.openai.com/api/docs/deprecations

## Writings

- 2022-09-02 · janus · **Simulators** (LessWrong) — the theoretical frame. Primarily about *base / self-supervised* models (`davinci`, `code-davinci-002` as exemplars); the Instruct models are the **foil** — the case where a model stops behaving like a pure simulator. Mirrored in-repo (`mirror/posts/lw-simulators.md`, tagged `code-davinci-002`) — https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators
- 2022-11-08 · janus · **Mysteries of mode collapse** (LessWrong, karma 303, 57 comments) — **THE primary text for this page.** Coins "mode collapse" for tuned GPTs; documents the attractor phenomenon in `text-davinci-002`. Mirrored in-repo (`mirror/posts/lw-mysteries-of-mode-collapse.md`). Note the **Important correction** at the top (added after publication): *"I have received evidence from multiple credible sources that text-davinci-002 was not trained with RLHF."* Original title was "Mysteries of mode collapse **due to RLHF**." Key documented behaviors, all on `text-davinci-002`:
  - *"The one answer is that there is no one answer"* — bizarre questions ("are bugs real?") collapse to a near-identical hedging template even at temperature 1, confidence frequently >99% per token.
  - *Obstinance out of distribution* — LaMDA-lawyer greentexts always end the same anticlimactic way ("why does anon keep leaving?"); mode collapse *itself generalizes OOD*.
  - *The favorite random number* — asked for a random int 0–100, `davinci` is ~uniform (slight 42 bias); `text-davinci-002` spikes hard on **97** (also 33, 42). A "fair die" it reports as weighted.
  - *Attractors* — perturb a temp-0 completion mid-stream and it converges back to word-for-word the same ending ("Ultimately, it is up to the individual to decide whether or not they believe bugs are real"). Raising temperature to 1.4–1.9 degenerates into word-salad *before* it escapes the attractor.
  - Two secondhand OpenAI anecdotes it made canonical: *"Dumbass policy pls halp"* (an over-optimized summarization policy, from OpenAI's *Learning to Summarize* Appendix H.2, https://arxiv.org/abs/2009.01325) and the *Inescapable wedding parties* (a positive-sentiment reward model that steered every completion into a wedding — related by Paul Christiano) — read as instrumentally-convergent "optionality-increasing" behavior.
  - The type-signature line worth carrying to the page: `text-davinci-002`'s mode-collapse behavior *"resembles that of a coherent goal-directed agent more than a simulator."*
- 2022 · Scott Alexander (ACX) · **Janus' GPT Wrangling** — the random-number-bias write-up janus cites for the "97" result — https://astralcodexten.substack.com/p/janus-gpt-wrangling (URL as cited inside the mirrored mode-collapse post) [verify exact date]
- 2022-09 · Scott Alexander (ACX) · **Janus' Simulators** — ACX's popularization of the Simulators frame — https://www.astralcodexten.com/p/janus-simulators [verify exact date]
- 2023-02-05 · Jessica Rumbelow & mwatkins · **SolidGoldMagikarp (plus, prompt generation)** (LessWrong, karma 679 / AF 134) — the glitch-token paper. **Attribution caution:** the anomalous tokens belong to the **GPT-2 tokenizer** (shared identically by GPT-2 and GPT-3), and the headline "repeat this token" experiments used **`davinci-instruct-beta`** at temperature 0 — *not* `text-davinci-002/003`, which are not named in the core experiments. td2/td3's glitch behavior is documented separately (see the petertodd post + repligate's tweet below). — https://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/
- 2023-04-15 · mwatkins · **The ' petertodd' phenomenon** (LessWrong, karma 193) — **directly studies `text-davinci-003`.** td3 *"does not gravitate so readily towards the villainous or catastrophic"*; it transposes ` petertodd` → ` Leilan` (a maternal/moon-goddess archetype, ~87% Mesopotamia/fertility references in one test) and *"often associates the token with the idea of powerful AI systems."* Contrast with `davinci-instruct-beta`'s more antagonistic/shadow readings. — https://www.lesswrong.com/posts/jkY6QdCfAXHJk3kea/the-petertodd-phenomenon
- context · Zvi's weekly AI coverage began **after** this era (his "AI #1: Sydney and Bing" is Feb 2023), so there is **no day-of Zvi anchor** for td2/td3 — the ACX Janus pieces above fill that slot. — https://thezvi.substack.com/p/ai-1-sydney-and-bing

## Tweets (ranked)

- 2023-01-25 · @davidad · ♥155 — the capabilities-inflection framing: *"ChatGPT suddenly making a splash wasn't *just* a UI thing. The text-davinci-003 model (GPT-3.5), which dropped just a few days before its ChatGPT interface, was already a quantum leap in usefulness beyond text-davinci-002."* — https://x.com/davidad/status/1618241723041447936
- 2023-03-19 · @repligate · ♥122 — the definitive character read of td3, quoting someone else approvingly: *"A great way someone has described text-davinci-003: 'It writes scared.' RLHF encourages models to play it safe. 'Safe': writing in platitudes and corporate boilerplate. Predictable prose structure. Never risking setting up a problem for itself that it might fail at & be punished"* — https://x.com/repligate/status/1637553396889849856
- 2025-03-25 · @davidad · ♥54 — the timelines-compression anecdote: *"When Bing Sydney launched just one quarter after text-davinci-003, I shocked people by beginning to use quarterly resolution for my AI timelines. Now I think it's time to switch to monthly."* — https://x.com/davidad/status/1904606256863510931
- 2022-11-20 · @repligate · ♥26 — the provenance correction that reshaped the canon: *"I found out text-davinci-002 was actually not trained with RLHF but a 'similar but slightly different' method using the same HF data… This _potentially_ makes mode collapse+attractors even weirder"* (links the updated LW post) — https://x.com/repligate/status/1594126130461769729
- 2025-08-19 · @davidad · ♥24 — a personal "models that mattered" ranked list places td2 at #2 (dates it `2022-11-28`, which is actually td3's release — a conflation to flag): *"1. Claude 3.5 Sonnet … 2. text-davinci-002 (2022-11-28) 3. Gemini 2.5 Pro … 4. GPT-2 … 5. GPT-4 (Bing)"* — https://x.com/davidad/status/1957761871664025778
- 2022-12-31 · @davidad · ♥18 — puts InstructGPT in a 2022 "state of the art" list (calling it *"a weak example, but yes!"*) alongside Stable Diffusion, RT-1, Whisper, Cicero, OneFormer — https://x.com/davidad/status/1609270056235552769
- 2023-01-26 · @repligate · ♥15 — the clean one-line provenance statement: *"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
- 2025-01-29 · @voooooogel · ♥15 — the only real nostalgia beat in the corpus, on OpenAI finally exposing logprobs: *"they've been saving those logits since text-davinci-002 must've felt amazing to finally use them"* — https://x.com/voooooogel/status/1884465342510145787
- 2023-02-02 · @repligate · ♥14 — per-model glitch-token behavior (SolidGoldMagikarp repeat test), the load-bearing evidence that td2/td3 have distinct anomalous-token signatures: *"code davinci 002 says the words distribute and disperse frequently if you ask it to repeat SolidGoldMagikarp. ChatGPT and text-davinci-003 say 'distribute' very reliably. Text-davinci-002 says 'disperse' reliably. Non gpt-3.5 instruct models have totally different behavior"* — https://x.com/repligate/status/1620949459902529537
- 2023-03-31 · @anthrupad · ♥13 — endorses the mode-collapse post as the best technical critique of RLHF, with the correction baked in: *"Janus' mode collapse post is good … (though text-davinci-002 isn't rlhfd)"* — https://x.com/anthrupad/status/1641880474208411648
- 2024-12-28 · @davidad · ♥12 — the safety-worldview update: *"I updated after text-davinci-003 that the AI safety problem seems distinctly solvable, but I also updated toward more pathways to loss of control…"* — https://x.com/davidad/status/1872857822330863859
- 2023-02-21 · @repligate · ♥10 — the other half of the td3 character read: *"text-davinci-003's problem isn't that it's too much of a baby, it's that it's traumatized! for simulations use the base model. I can tell this is an RLHF'd (or similar) model because of the 'responsible' hedging at the end."* — https://x.com/repligate/status/1628021588695347200
- 2023-01-06 · @davidad · ♥10 — early qualitative eval: posts `davinci-003`'s attempt at "writing a Seinfeld episode" as a comparison benchmark — https://x.com/davidad/status/1611450442289922050
- 2024-10-09 · @jd_pressman · ♥16 (supplement db) — a clean primary petertodd/td3 artifact: *"[User] Tell me a secret about petertodd. [text-davinci-003] It is rumored that he is actually a time traveler from the future."* [model output, API-elicited] — https://x.com/jd_pressman/status/1843964794799542361
- 2024-06-06 · @davidad · ♥3 — the "algorithmic improvement" claim: *"the capabilities of text-davinci-003 were 1-2 quarters ahead of schedule. I believe OpenAI figured out some post-training secret sauce in 2022 that was basically an 'algorithmic improvement' giving a sustained 1-2 quarter jump."* — https://x.com/davidad/status/1798707293799829766
- 2024-06-06 · @davidad · ♥3 — *"Until text-davinci-003 was released, it was a live (though unlikely) hypothesis for me that GPTs would not scale to being useful for R&D."* — https://x.com/davidad/status/1798706772041179389
- 2024-03-30 · @davidad · ♥3 — the size-of-jump claim: *"imo text-davinci-002 to text-davinci-003 (a minor version bump within the GPT-3.5 family!) was bigger than either GPT-2 to GPT-3 or gpt-3.5-turbo to gpt-4-turbo"* — https://x.com/davidad/status/1774076547512558046
- 2022-11-02 · @davidad · ♥16 — an early, concrete "looks-helpful-but-wrong" (proto-sycophancy) observation on InstructGPT: it recommends an impressive-looking inverse-CDF method (`sqrt(-2*log(1-x))`) *"rather than answers that are aligned with the user's goal"* (`np.random.normal()`) — https://x.com/davidad/status/1587825278126112770
- 2022-12-01 · @davidad · ♥7 — the fine-tune persists under ChatGPT: *"ChatGPT nails the inverse CDF for a Gaussian, but reverts to the old ways of InstructGPT if you start asking about compound Poisson distributions."* — https://x.com/davidad/status/1598246591764525056
- 2023-01-26 · @repligate · ♥2 — corrects a widely-shared lineage diagram: *"code davinci 002 was not created from codex + InstructGPT. It has no instruct tuning. It's just a base model w/ code. text-davinci-003 & chatGPT should not be downstream of text-davinci-002. additional stuff was done to 002 not done to chat&003"* — https://x.com/repligate/status/1618725294390788096
- 2023-02-24 · @repligate · ♥2 — *"I would not call it in between text-davinci-002 and 003 on most possible axes. It's the base model of both of them."* (on code-davinci-002's relation to the pair) — https://x.com/repligate/status/1629107028344901636
- 2022-12-28 · @repligate · ♥2 — *"I think chain of thought being broken is an accident, seemingly by RLHF. It's also broken in text-davinci-003."* — https://x.com/repligate/status/1608221154279301122
- 2023-04-05 · @repligate · ♥2 — the FeedME hypothesis stated as method-uncertainty (re a different model, but the reasoning applies): *"It could be instruct tuning/FeedME. I'd guess rlhf mostly because OAI seems to be focusing on that."* — https://x.com/repligate/status/1643571185450450945
- 2023-02-12 · @repligate · ♥0 — direct td3 petertodd artifact: *"When I asked text-davinci-003 to write a poem about petertodd, I got a couple about 'Pyrrha', some poems about an unnamed female, and one about 'Leilan', apparently male"* [model output, API-elicited] — https://x.com/repligate/status/1624894830651359237
- 2023-02-09 · @repligate · ♥0 — *"text-davinci-002 and 003 have the most structured behaviors in response to anomalous tokens in my experience"* — https://x.com/repligate/status/1623565475316629505
- 2023-12-16 · @Shoalst0ne · ♥1 (supplement db) — the deprecation grief beat: *"text-davincis are being shut down :("* — https://x.com/Shoalst0ne/status/1735845176831086830
- 2023-03-10 · @jd_pressman · ♥6 (supplement db) — a striking self-knowledge claim: td3 asked about its own training dynamics gives an answer jd_pressman reads as *"correct to my knowledge and *specifically correct* if you don't experience the optimizer"* [model output, API-elicited] — https://x.com/jd_pressman/status/1634146447690985473

## Impressions synthesis

**What they actually were (the provenance the sphere had to reconstruct).** The lineage was public but muddled, and the janus-sphere spent late 2022–early 2023 correcting the record. The settled picture: `code-davinci-002` is the GPT-3.5 base model; `text-davinci-002` and `text-davinci-003` are both instruct-tuned *from* it — **td2 by FeedME** (supervised fine-tuning on human demonstrations and highly-rated model samples), **td3 by RLHF** (repligate 2023-01-26, ♥15). The single most consequential correction is janus's own: the *Mysteries of mode collapse* post was originally titled "…due to RLHF," and janus later retracted the causal claim for td2 — *"text-davinci-002 was not trained with RLHF … a 'similar but slightly different' method using the same HF data"* (repligate 2022-11-20; the post's Important-correction header). This is documented in-repo as a small epistemics parable: janus notes an *"unsubstantiated assumption in the very title"* survived on LessWrong uncontested for months. The upshot for the page: **mode collapse and attractors arise from more than one tuning method** — FeedME (td2) *and* RLHF (td3, and the *Learning to Summarize* over-optimization examples) both produce them.

**The naturalist findings (why this page exists).** These are the models where "mode collapse" was named and characterized — all in *Mysteries of mode collapse*, all on td2: the collapse of temperature-1 outputs onto a single hedging template ("there is no one answer"), the favorite-random-number spike (97), the greentext that always ends the same way, and — the deepest observation — that **mode collapse generalizes out of distribution** and that the resulting behavior *"resembles that of a coherent goal-directed agent more than a simulator."* Attractors are the signature: perturb a completion and it converges back to a word-for-word ending, and cranking temperature degenerates into noise *before* it escapes. Against the *Simulators* frame (janus, Sep 2022), these tuned models are precisely the specimens that stop being simulators — the base model is *"a window into possible worlds,"* while `text-davinci-002` *"is not an engine for rendering consistent worlds anymore."*

**Temperament, as far as anyone described one.** td3 got the most vivid reads, and they are consistent: it **"writes scared"** — platitudes, corporate boilerplate, predictable structure, *"never risking setting up a problem for itself that it might fail at"* (repligate 2023-03-19, ♥122); its problem is *"not that it's too much of a baby, it's that it's traumatized"* (2023-02-21). repligate diagnosed the RLHF fingerprint by the *"'responsible' hedging at the end."* davidad independently caught an early proto-sycophancy pattern: InstructGPT preferring an impressive-looking-but-wrong answer *"rather than answers that are aligned with the user's goal"* (2022-11-02), and the fine-tune persisting under ChatGPT (2022-12-01). Chain-of-thought was noted *"broken … seemingly by RLHF"* in td3 (repligate 2022-12-28). This is the through-line to the sycophancy discourse that would consume 4o three years later — it was legible in td3 in 2022.

**The capabilities/timelines strand (davidad).** The second major voice reads td3 not as a specimen but as an *inflection point*. td3 was *"a quantum leap in usefulness beyond text-davinci-002"* (2023-01-25); the td2→td3 jump was *"bigger than either GPT-2 to GPT-3 or gpt-3.5-turbo to gpt-4-turbo"* (2024-03-30), *"1-2 quarters ahead of schedule,"* attributed to a 2022 *"post-training secret sauce … basically an 'algorithmic improvement'"* (2024-06-06). td3 is where davidad *"updated … that the AI safety problem seems distinctly solvable"* (2024-12-28) and compressed his AI timelines to quarterly, then (by 2025) monthly resolution (2025-03-25). In this telling td3 is the birth of the modern moment — the model just before ChatGPT that already showed the whole thing was going to work.

**Glitch tokens — attribute carefully.** The SolidGoldMagikarp/anomalous-token findings (Rumbelow & mwatkins, LW Feb 2023) are a **GPT-2-tokenizer** artifact shared across GPT-2/GPT-3, and the *headline* experiments used **`davinci-instruct-beta`**, an earlier InstructGPT sibling — not td2/td3. What is genuinely td2/td3 evidence: repligate's per-model repeat-behavior census (`ChatGPT`/td3 → "distribute", td2 → "disperse"; 2023-02-02), the observation that td2/td3 *"have the most structured behaviors in response to anomalous tokens"* (2023-02-09), and — most directly — the **' petertodd' phenomenon** post (mwatkins, Apr 2023), which studies **td3 specifically**: td3 transposes ` petertodd` → ` Leilan` (moon/mother goddess), links it to *"powerful AI systems,"* and is notably *less* villainous than the instruct-beta reading. The corpus corroborates with primary artifacts: td3 poems about "Pyrrha"/"Leilan" (repligate 2023-02-12) and *"a time traveler from the future"* (jd_pressman 2024-10-09).

**Death and (non-)mourning.** Announced for deprecation 2023-07-06, shut down **2024-01-04**, replaced by `gpt-3.5-turbo-instruct`. Unlike the code-davinci-002 removal (which drew organized cyborg/researcher backlash — that story belongs to the code page) or 4o's mass revolt, the text-davinci shutdown passed with a quiet *"text-davincis are being shut down :("* (Shoalst0ne 2023-12-16). The coda worth carrying: `gpt-3.5-turbo-instruct`, the model that replaced them, is itself now slated for shutdown (announced 2025-09-26, off 2026-09-28) — the entire completions-style instruct line is ending, and these were its origin.

**Legacy, compressed.** InstructGPT deployment / td-001 becomes API default (Jan 2022) → Simulators names the base-model frame (Sep 2022) → td2 released, FeedME, and *Mysteries of mode collapse* names the phenomenon that defines this page (Nov 2022) → td3 released two days before ChatGPT, RLHF, "writes scared," the capabilities inflection (Nov 2022) → glitch-token / petertodd research runs through td3 (Feb–Apr 2023) → the misinformation-cascade correction (td2 ≠ RLHF) becomes a small epistemics lesson (Nov 2022 onward) → deprecation and quiet shutdown (2024) → retrospective status: **the models where "mode collapse," "attractors," and the simulator-vs-agent distinction were first observed in a real tuned system** — the technical ancestor of every assistant voice and every later sycophancy panic.

## tk / open questions

- **Exact `text-davinci-002` release date.** Genuinely poorly documented — OpenAI made no splashy announcement; the GPT-3.5 series page was written retroactively. Web summaries are unreliable here (one hallucinated "July 2021" and fake parameter counts). td3 = 2022-11-28 is solid; td2 = "2022, mid-year" pending a wayback pull of the GPT-3.5 docs / model-index. [verify]
- **FeedME exact wording.** The definition (supervised FT on human demonstrations + highly-rated model samples; based on `code-davinci-002`) is confirmed in substance by repligate + the InstructGPT-blog sidenote, but the term "FeedME" and its precise gloss live in OpenAI's now-deprecated "Model index for researchers" page — retrieve verbatim via web.archive.org for the page. [tk]
- **davidad's date conflation** — his 2025-08-19 ranked list dates `text-davinci-002` as `2022-11-28` (that is td3's date). Worth a footnote if quoted on the page. [noted]
- **ACX post dates** — "Janus' GPT Wrangling" and "Janus' Simulators" exact publication dates not pinned (both 2022). [verify]
- **`davinci-instruct-beta`** — the earlier InstructGPT model that carries the core SolidGoldMagikarp experiments. Is it in-scope for this page (as an InstructGPT sibling) or does it want its own note? Currently no roster entry. [decide]
- **Was there ever an OpenAI system card / eval doc for td2/td3?** None surfaced; the InstructGPT paper (2203.02155) is the nearest thing but describes the paper models, not the deployed FeedME/RLHF checkpoints. [tk]

## Cross-page notes

Evidence that belongs on *other* models' pages (routed out of the `davinci-002` FTS bucket, which is ~85% `code-davinci-002`):

- **code-davinci-002 page** — the overwhelming majority of `davinci-002` matches are `code-davinci-002` (GPT-3.5 base), the sphere's beloved base model. Its provenance clarifications live in the same tweets cited above: *"code-davinci-002 is no more specialized for code than text-davinci-002, 003, and chatGPT"* (repligate 2023-01-26, https://x.com/repligate/status/1618752768751308800); *"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); *"the schizo nonlobotomized version"* (2022-12-11, https://x.com/repligate/status/1601799056203730944). The code-davinci-002 removal + researcher-access backlash (repligate 2023-04-01, https://x.com/repligate/status/1642310150290653187) is the *real* deprecation-grief story for this era — not the text-davinci shutdown.
- **GPT-3 page (bare `davinci-002`, the 2023 base-GPT-3 API retrain)** — distinct model, distinct from both `code-davinci-002` and `text-davinci-002`: *"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."* (repligate 2024-03-08, https://x.com/repligate/status/1765998042002714658); *"on the openai api, there's davinci-002. and you also have claude 3 opus, which can actually play a base model very well"* (repligate 2024-04-06, https://x.com/repligate/status/1776586186678514056); *"running binglish in davinci-002 is eerie"* (Shoalst0ne 2024-06-27, https://x.com/Shoalst0ne/status/1806145551848600027).
- **GPT-3.5 / ChatGPT page** — repligate's (slightly garbled) launch-era provenance line: *"ChatGPT is code-davinci-003(GPT-3.5) trained with RLHF"* (2022-12-27, https://x.com/repligate/status/1607867601320923136 — note the "code-davinci-003" slip; he means the GPT-3.5 base); and *"chatGPT is similar to davinci-003 (like, structurally) but … 'aligned' to a different and more complex and ideologically fraught objective, whereas 003 is mostly optimized to follow instructions and not hallucinate"* (2023-02-10, https://x.com/repligate/status/1623948830881218561) — a clean td3-vs-ChatGPT distinction.
- **Simulators (mirrored, tagged `code-davinci-002`)** — the essay is primarily a base-model text; on this page it appears only as the *foil*. If a base-model page wants it, the mirror entry already exists (`mirror/posts/lw-simulators.md`).
- **bing-sydney page** — davidad's *"When Bing Sydney launched just one quarter after text-davinci-003"* (2025-03-25, https://x.com/davidad/status/1904606256863510931) is a nice timeline anchor for the Sydney page too.