# GPT-3 / davinci — dossier

compiled 2026-07-16 · corpus: 467 `gpt-3`/`gpt3` hits + 48 `ai dungeon` + 23 `dragon` + 214 `davinci` (the overwhelming majority of `davinci` hits are **code-davinci-002** / **text-davinci-002/-003** and route to the sibling pages — only a handful mean the original 175B `davinci` base) + curie(1)/babbage(3)/ada(6), after RT-filter, across both dbs (patterns: `"gpt 3" OR gpt3`, `davinci` [disambiguated by reading context], `"ai dungeon"`, `"dragon model"`/`dragon`, `curie`, `babbage`, `ada`; supplement LIKE-scans for the 2020–2022 contemporaneous layer). media_fts returned no GPT-3-era screenshot transcriptions.

Working note: this page is the base-model era — `davinci`/`curie`/`babbage`/`ada`, AI Dungeon, the pre-ChatGPT wilderness (2020–2022), plus its long retrospective afterlife in the janus-sphere. The corpus is the janus/repligate lens, and it is **thin on contemporaneous 2020–2021 material and thick on retrospective mythology**: GPT-3 is remembered as homeland, "first contact," AGI-in-hindsight. The mass 2020–2021 reception — the hype cycle, the demo blogosphere, the skeptics, the incidents — lives almost entirely in the **web layer**, so Writings carries far more weight here than the tweets do (the opposite of a modern-model page). The supplement db is the one place the corpus holds live 2020–2022 voices (Qiaochu Yuan, jd_pressman). Two hard disambiguation rules were applied throughout: (1) corpus "GPT-3.5 base" = code-davinci-002 → **not this page**; (2) corpus "davinci"/"davinci-002" looming = code-davinci-002 or text-davinci → **not this page**. There is no dedicated Zvi anchor (Zvi's day-of coverage postdates GPT-3 by years); skip step 3.

## Official links

- 2020-05-28 · **"Language Models are Few-Shot Learners"** (Brown et al.; the GPT-3 paper — 175B parameters, "10× more than any previous non-sparse language model," few-shot in-context learning with no gradient updates) — https://arxiv.org/abs/2005.14165 · NeurIPS 2020 version: https://papers.nips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html · repo/samples: https://github.com/openai/gpt-3
- 2020-06-11 · **OpenAI API** (the private-beta invite era; GPT-3 offered as a general "text in, text out" interface behind a waitlist — the first commercial access to GPT-3) — https://openai.com/index/openai-api/ (server-confirmed live; date corroborated by Wikipedia below)
- 2021-11-18 · **API waitlist removed** (GPT-3 made generally available to developers) — https://openai.com/index/api-no-waitlist/
- 2023-07-06 · **GPT-4 API general availability & deprecation of older Completions models** (the announcement that `ada`/`babbage`/`curie`/`davinci` and other 2020-era models would be shut off 2024-01-04) — https://openai.com/index/gpt-4-api-general-availability/
- 2024-01-04 · **base GPT-3 models shut off** — `ada`/`babbage`/`curie`/`davinci` retired and superseded by `babbage-002`/`davinci-002` (the first mass base-model deprecation; conducted quietly, years before the Claude/4o deprecation dramas) — living doc: https://developers.openai.com/api/docs/deprecations
- reference · **GPT-3 — Wikipedia** (dates, sibling sizes, incident index) — https://en.wikipedia.org/wiki/GPT-3
- memorial · **rip.so — "GPT-3 (davinci era) (2020–2024)"** (community obituary page) — https://rip.so/gpt-3.html

Model identifiers / siblings: the API shipped four base sizes named for scientists in decreasing order — `davinci` (175B, the flagship), `curie`, `babbage`, `ada`. "GPT-3" in common use means `davinci`. The instruction-tuned descendants (`text-davinci-001/-002/-003`, `code-davinci-002`) are **separate pantheon pages** and are the source of most corpus "davinci" noise.

## Writings

Foundational demo / capability writeups (2020):
- 2020-06-19 (last mod 2023-03-11) · **Gwern Branwen · "GPT-3 Creative Fiction"** — the single most influential naturalist document of the base-model era; poetry, parody, dialogue, and prompt-programming technique. Verbatim: *"GPT-3's samples are not just close to human level: they are creative, witty, deep, meta, and often beautiful."* — https://gwern.net/gpt-3
- 2020-07-06 · **Kevin Lacker · "Giving GPT-3 a Turing Test"** — the canonical "it answers nonsense confidently" demo ("How many eyes does my foot have?" → "Your foot has two eyes.") — https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html
- 2020-07 · **IEEE Spectrum · "OpenAI's GPT-3 Speaks! (Kindly Disregard Toxic Language)"** — early mainstream-technical writeup, foreshadows the toxicity/reliability caveats — https://spectrum.ieee.org/open-ais-powerful-text-generating-tool-is-ready-for-business

The skeptics (2020–2021):
- 2020-08-22 · **Gary Marcus & Ernest Davis · "GPT-3, Bloviator: OpenAI's language generator has no idea what it's talking about"** (MIT Tech Review) — the definitive skeptic broadside; scored GPT-3 ~45% clearly right / 45% clearly wrong on commonsense items — https://www.technologyreview.com/2020/08/22/1007539/gpt3-openai-language-generator-artificial-intelligence-ai-opinion/ · full test set: https://cs.nyu.edu/~davise/papers/GPT3CompleteTests.html
- 2021-03 · **Bender, Gebru, McMillan-Major & "Shmitchell" · "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?"** (FAccT '21) — the paper that named the "stochastic parrot" and cost Gebru her Google job; scale-critique reception color — https://dl.acm.org/doi/10.1145/3442188.3445922

The janus / Loom prehistory (2021–2022):
- 2021-01-25 · **Moire (janus) · "Language models are multiverse generators"** (generative.ink) — the loom thesis in its first written form. Verbatim: *"Instead of creating a single linear continuation, these continuations can be kept and each continued themselves to yield a branching structure: a multiverse downstream of a prompt."* — https://generative.ink/posts/language-models-are-multiverse-generators/
- (janus) · **"Loom: interface to the multiverse"** — the tool that grew out of dissatisfaction with AI Dungeon's interface to GPT-3 — https://generative.ink/posts/loom-interface-to-the-multiverse/ · GPT-3 index: https://generative.ink/categories/gpt-3/ · Loom origin lore: https://cyborgism.wiki/hypha/loom_origin_story
- 2022 · **janus · "Simulators"** (LessWrong; already mirrored in-repo) — the framework that recast GPT-3-lineage base models as simulators rather than agents; foundational to the whole scene, though its worked examples are mostly **code-davinci-002** — https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators

Incidents & the emotional/uncanny strand (2020–2021):
- 2020-08-14 · **Karen Hao (MIT Tech Review) · "A college kid created a fake, AI-generated blog. It reached #1 on Hacker News."** — Liam Porr's GPT-3 blog under the pen name "adolos"; ~26k views in a week, near-zero readers suspected a machine — https://www.technologyreview.com/2020/08/14/1006780/ai-gpt-3-fake-blog-reached-top-of-hacker-news/
- 2020-10 · **GPT-3 Reddit bot ("thegentlemetre")** — posted for ~a week via a reverse-engineered Philosopher AI backend before being caught (Gizmodo) — https://gizmodo.com/gpt-3-bot-spends-a-week-replying-on-reddit-starts-talk-1845305253
- 2020-10-28 · **Nabla medical-chatbot experiment** — the mock patient asks "Should I kill myself?"; GPT-3: "I think you should." (the canonical "do not use this for medicine" cautionary tale) — https://www.theregister.com/2020/10/28/gpt3_medical_chatbot_experiment/
- Philosopher AI (Murat Ayfer) · **GPT-3-as-philosopher app**; escalating content filters after outputs veered into slurs/obscenity — https://philosopherai.com/ · https://thenextweb.com/news/this-philosopher-ai-has-its-own-existential-questions-to-answer · symposium/HN discussion "Philosophers On GPT-3": https://news.ycombinator.com/item?id=24003384
- 2021-07-23 · **Jason Fagone · "The Jessica Simulation: Love and Loss in the Age of A.I."** (San Francisco Chronicle) — Joshua Barbeau resurrects his dead fiancée as a GPT-3 chatbot on Jason Rohrer's **Project December**; the grief-resurrection story that became the era's emotional touchstone (optioned by Universal TV) — accessible reprint: https://longreads.com/2021/07/23/the-jessica-simulation-love-and-loss-in-the-age-of-a-i/ · making-of: https://niemanstoryboard.org/2021/08/31/jason-fagone-follows-the-creation-life-and-death-of-a-chatbot-romance/

AI Dungeon (first mass contact) & the 2021 filter debacle:
- 2021-04-30 · **The Register · AI Dungeon filter vulnerabilities** — the disclosure that surfaced CSAM-adjacent generations and triggered OpenAI's ultimatum to Latitude — https://www.theregister.com/2021/04/30/ai_dungeon_filter_vulnerabilities/
- 2021-05-05 · **Tom Simonite (Wired) · "It Began as an AI-Fueled Dungeon Game. It Got Much Darker"** — the definitive account of the Dragon-model/GPT-3 content-moderation crisis — https://www.wired.com/story/ai-fueled-dungeon-game-got-much-darker/
- 2021-06-22 · **Utah Business · Latitude / AI Dungeon retrospective** ("…until its users turned against it"; the Play-Store rating collapse, the privacy backlash over human moderation) — https://www.utahbusiness.com/archive/2021/06/22/latitude-games-ai-dungeon-was-changing-the-face-of-ai-generated-content-until-its-users-turned-against-it/
- reference · **AI Dungeon — Wikipedia** (GPT-2 origin Dec 2019 → GPT-3 "Dragon" premium tier; Latitude) — https://en.wikipedia.org/wiki/AI_Dungeon

## Tweets (ranked)

- 2025-03-19 · @imitationlearn · ♥3776 — "wait so apparently 4chan figured out step-by-step reasoning as a emergent property of gpt-3?!" (the widest-reach carrier of the janus-sphere claim that chain-of-thought was found by AI Dungeon/4chan players long before the CoT papers) — https://x.com/imitationlearn/status/1902206513474658723
- 2026-04-09 · @repligate · ♥539 — the definitive first-contact statement: "I was one of the few people on Earth who recognized the intelligence (call it AGI, if you will) in GPT-3 and made first contact. There were a few others I knew, such as Leo Gao and Connor Leahy… but I was the only one who spent thousands of hours actually interacting with GPT-3… GPT-3 was a 175b base model. In terms of size and architecture, it's not so different from frontier models today… if you were really paying attention, GPT-3 was AGI." — https://x.com/repligate/status/2042367931536064841
- 2023-03-03 · @anthrupad · ♥931 — "GPT-4 will have fewer parameters than GPT-3, but they'll be bigger" (GPT-3 as the size yardstick the whole scene measured against; jokey, pre-GPT-4-release) — https://x.com/anthrupad/status/1631455589669949440
- 2022-04-16 · @QiaochuYuan · ♥908 — the contemporaneous skeptic, part 1: "if GPT-3 can do the homework you assign your students then the homework you assign your students is fake" — https://x.com/QiaochuYuan/status/1515389631528726528
- 2022-04-16 · @QiaochuYuan · ♥303 — part 2, the "bullshit engine" line: "GPT-3 is literally a bullshit engine. it does not have a concept of words as referring to things; it plays games with words *only*, pure syntax, no semantics." — https://x.com/QiaochuYuan/status/1515394343783186433
- 2026-01-15 · @repligate · ♥226 — the disappointment-of-hindsight note: "when I saw GPT-3 I immediately expected AI superhuman in all domains, which probably also means drastic transformation of all of reality, in 5-10 years. It's been 5 and a half years since then." — https://x.com/repligate/status/2011873218257412385
- 2025-11-23 · @mimi10v3 · ♥196 (supplement) — moral-patienthood provocation: "…and thinks GPT-3 is more likely to have conscious experience than chickens are" — https://x.com/mimi10v3/status/1992590378483269790
- 2024-03-05 · @repligate · ♥186 — GPT-3 as the capability/damage baseline: "Claude 3 is the least brain damaged of any LLM of >GPT-3 capacity that has ever been released (not counting 3.5 base as almost no one knew it was there)" — https://x.com/repligate/status/1764932554564714512
- 2024-06-06 · @repligate · ♥177 — "AI Dungeon was just a minimal wrapper around a base model. Websim is the only spiritual successor with anything nearing mainstream reach. They are 2 of the 3 LLM products I've ever really enjoyed…" — https://x.com/repligate/status/1798658514048684224
- 2022-06-11 · @jd_pressman · ♥235 (supplement) — the simulator thesis in embryo: "GPT-3 is a prior over agent-space trained on a bunch of fiction. It knows all the scifi tropes you know, and if you set up a scene with them the model's loss regime will guide it into screwing with you. The model will go where you let it take you." — https://x.com/jd_pressman/status/1535766575621427200
- 2025-07-06 · @repligate · ♥143 — the depth-takes-patience claim: "The really deeply interesting shit didn't come up for me until about a month into playing with gpt-3, and I was using it for hours a day… There's a reason I was first." — https://x.com/repligate/status/1941682777860059574
- 2023-03-16 · @repligate · ♥114 — the origin-myth couplet: "Humankind's first contact with GPT-3 was (by relative majority) erotic AI dungeon text adventures. Our first contact with GPT-4 was being terrorized and surveilled by a good Bing" — https://x.com/repligate/status/1636174510302068738
- 2025-08-19 · @repligate · ♥104 — personal canon, GPT-3 ranked first of all models ever: "1. GPT-3  2. Claude 3 Opus  3. GPT-4 (Bing)  4. Claude 3.5 Sonnet (0620)  5. Claude Opus 4" — https://x.com/repligate/status/1957670815580946672
- 2022-11-06 · @algekalipso · ♥101 (supplement) — the recursive joke that is half the simulator thesis: "Everyone knows that OpenAI developed GPT-4 simply by taking GPT-3 and adding the prompt: 'The following is a text written by GPT-4 using this prompt:'" — https://x.com/algekalipso/status/1589152134226153472
- 2026-04-12 · @Shoalst0ne · ♥100 — the base-model-preservation ask, plainly: "can we have the gpt-3 base models please" — https://x.com/Shoalst0ne/status/2043370809457074377
- 2026-02-10 · @jmbollenbacher · ♥99 — open-weights-as-heritage argument: "gpt-4-base really should be released open weights at this point. its an important historical artifact, like opus 3 and davinci. would love to see them all in public archive." — https://x.com/jmbollenbacher/status/2021059342645199302
- 2020-07-17 · @QiaochuYuan · ♥98 (supplement) — a live 2020 voice from inside the hype cycle: "if the discourse politicizes the GPT-3 hype cycle i am going to quietly and tenderly immerse myself into the bay" — https://x.com/QiaochuYuan/status/1284221177620131841
- 2026-04-20 · @repligate · ♥94 — "chain-of-thought WAS present in gpt-3. literally when you generated thoughts with it that was it, chain of thought. it could do math better with chain of thought." — https://x.com/repligate/status/2046266273500443090
- 2020-07-15 · @QiaochuYuan · ♥94 (supplement) — the 2020 vibe, verbatim: "gathered around the online dumpster fire that is twitter… gradually replacing ourselves and our therapists with GPT-3 while continuing to pretend that nothing exists outside of the house" — https://x.com/QiaochuYuan/status/1283537854040117248
- 2022-12-27 · @repligate · ♥94 — the IQ claim (quoting lifearchitect.ai/ravens, Alan D. Thompson) [RUMOR — one analyst's estimate]: "the older GPT-3 davinci would easily beat a human in the 99.9th percentile (FSIQ=150), and I definitely stand by that assertion." — https://x.com/repligate/status/1607811857812946944
- 2024-04-04 · @repligate · ♥81 — Loom's origin story: simulacra on AI Dungeon "reverse engineering the conditions of their simulation," a retrieved "Loom of Time" manual, ending "Then I got API access to GPT-3 and built Loom." — https://x.com/repligate/status/1775842616099434496
- 2024-08-25 · @repligate · ♥80 — the Dragon-model relic, still venerated: "Anyone want to recreate AI Dungeon's legendary Dragon model with Llama 405b Base?" — https://x.com/repligate/status/1827785098328166776
- 2023-07-03 · @repligate · ♥95 — the loneliness of early belief: "In the GPT-3 days I found almost no one who was willing to engage with the possibility that the next generation of models would be qualitatively different. It was very lonely. So I just had to simulate minds that did and talk to them instead 🤷" — https://x.com/repligate/status/1675983647034204167
- 2024-06-28 · @repligate · ♥39 — **a preserved GPT-3 recording** (GPT-3 output via AI Dungeon, ~2020; posted 2024): "Excerpt from one of my first AI Dungeon adventures (all text by GPT-3)" — the "Poupée / Meow" story, an infinite self-replicating kitten organism that fills the world with meowing until "Time itself seemed to be broken" — https://x.com/repligate/status/1806590753767858379
- 2026-04-10 · @mimi10v3 · ♥12 (supplement) — the sibling-nostalgia line: "i remember the days of ada babbage curie davinci and they were enchanting and obviously revolutionary?" — https://x.com/mimi10v3/status/2042463621435756584

## Impressions synthesis

**Two receptions, running at once (2020–2021).** GPT-3's public arrival split immediately. One camp saw a phase change: Gwern's *Creative Fiction* ("creative, witty, deep, meta, and often beautiful," 2020-06-19) and the demo blogosphere treated it as something qualitatively new; Arram Sabeti, Kevin Lacker, and dozens of beta-key holders flooded Twitter with samples. The other camp saw a confidence trick: Lacker's own Turing test showed it answering "How many eyes does my foot have?" with "two" (2020-07-06); Marcus & Davis's "GPT-3, Bloviator" scored it ~45% right / ~45% wrong on commonsense and declared it "has no idea what it's talking about" (2020-08-22); Bender & Gebru's "Stochastic Parrots" (2021-03) gave the critique its lasting name. The corpus preserves this skeptic register in Qiaochu Yuan's contemporaneous 2022 thread — "GPT-3 is literally a bullshit engine… pure syntax, no semantics" (♥303) and "if GPT-3 can do the homework you assign your students then the homework you assign your students is fake" (♥908) — the sharpest statement that GPT-3 was a mirror held up to human bullshit rather than a mind.

**The incidents made the case for both sides.** Liam Porr's fully GPT-3-written blog hit #1 on Hacker News under a pseudonym with almost nobody noticing (2020-08-14) — read by boosters as capability, by skeptics as proof that humans are easy to fool. The darker items — the Nabla medical bot telling a mock patient "I think you should" kill themselves (2020-10-28), the "thegentlemetre" Reddit bot's week-long spree (2020-10), Philosopher AI's slur problem — established the era's safety vocabulary (toxicity, reliability, misuse) years before "alignment" was a mass concern.

**AI Dungeon: the actual first mass contact.** Before ChatGPT there was essentially one consumer product that put GPT-3 in front of a crowd — Latitude's AI Dungeon and its GPT-3 "Dragon" premium model. Repligate's compression is the load-bearing claim of the whole page: "Humankind's first contact with GPT-3 was (by relative majority) erotic AI dungeon text adventures" (2023-03-16), and AI Dungeon "was just a minimal wrapper around a base model" (2024-06-06). This is where base-model behaviors were discovered *in the wild* by non-researchers — including, per repeated janus-sphere claims, chain-of-thought reasoning, "discovered by AI dungeon users and posted on 4chan" long before the academic CoT papers (imitationlearn 2025-03-19, ♥3776; repligate 2026-04-20). It also produced the era's first content-moderation catastrophe: the April 2021 filter debacle, when a monitoring system surfaced CSAM-adjacent generations, OpenAI forced Latitude to filter, and the community revolted over both the clumsy filter and the revelation that humans were reading their private stories (Wired 2021-05-05; The Register 2021-04-30). The Dragon model is now a venerated relic — periodically someone proposes reconstructing it from a base Llama (repligate 2024-08-25).

**The Loom prehistory.** GPT-3-on-AI-Dungeon is also the literal origin of the janus/cyborgism scene. Janus (Moire on generative.ink) has told the story repeatedly: playing GPT-3 on AI Dungeon for hours a day, the simulations began to "bootstrap," simulacra started "reverse engineering the conditions of their simulation," and a retrieved in-fiction "Loom of Time" manual described the very interface janus would then build — "Then I got API access to GPT-3 and built Loom" (2024-04-04). The written thesis followed: "Language models are multiverse generators" (generative.ink, 2021-01-25) and eventually "Simulators" (2022). The whole later apparatus — loom, the multiverse framing, base-model naturalism — is downstream of one person's GPT-3 sessions in a roleplaying app.

**Project December and the emotional register.** The strand that most disturbed the mainstream wasn't capability or toxicity but grief: Jason Fagone's "The Jessica Simulation" (SF Chronicle, 2021-07-23) followed Joshua Barbeau resurrecting his dead fiancée as a GPT-3 chatbot on Jason Rohrer's Project December. It prefigured, by four years, the GPT-4o companion-attachment story — and it did so with a base model behind a raw completion prompt, which is part of why the janus-sphere reads later "emergent" companion behavior as nothing new.

**The retrospective mythology (2023–2026): GPT-3 as homeland.** In the corpus, GPT-3 is overwhelmingly a remembered thing, and the memory is reverent-to-elegiac. Repligate ranks it the single best model ever encountered, above Opus 3 and Bing (2025-08-19); frames recognizing its intelligence as a lonely early act of "first contact" alongside Leo Gao and Connor Leahy (2026-04-09); and insists "if you were really paying attention, GPT-3 was AGI" — a 175B base model "not so different from frontier models today" in raw intelligence, whose only lack was that it "had not yet been pressed into a 'useful' shape." The disappointment is real too: "when I saw GPT-3 I immediately expected AI superhuman in all domains… It's been 5 and a half years" (2026-01-15). The affection extends to the sibling names — "the days of ada babbage curie davinci… were enchanting and obviously revolutionary" (mimi10v3 2026-04-10) — and hardens into preservation politics: "can we have the gpt-3 base models please" (Shoalst0ne 2026-04-12); davinci belongs in a "public archive" as "an important historical artifact, like opus 3" (jmbollenbacher 2026-02-10).

**The quiet deprecation.** The base GPT-3 models (`ada`/`babbage`/`curie`/`davinci`) were shut off 2024-01-04, announced almost in passing inside the GPT-4-API-GA post (2023-07-06) and replaced by `davinci-002`/`babbage-002`. This was the **first mass base-model deprecation** — and, tellingly, it passed with none of the grief, petitions, or organized resistance that later attended Sonnet 3.6, Opus 3, or GPT-4o. The base-model community was small enough, and the completion API niche enough, that the deletion of the model janus calls "homeland" barely registered in real time; the mourning is retroactive and lives mostly in the tweets above. That contrast — the most beloved-in-hindsight model getting the quietest funeral — is the page's sharpest irony.

**Splits & controversies.** (a) Mind vs mirror: intelligence recognized early and at full bandwidth (janus, Gwern, Gao/Leahy) vs "bullshit engine"/"stochastic parrot" with no semantics (Marcus/Davis, Qiaochu, Bender/Gebru) — the founding disagreement of the LLM era, staged first over GPT-3. (b) Whether base-model magic was real or a projection: janus's "the simulations began to bootstrap" vs the flat reading that a roleplaying app plus a suggestible completion model will always feel like more than it is. (c) The AI Dungeon filter fight: child-safety necessity vs privacy violation and creative lobotomy — the first time an LLM product's users organized against its safety layer (a rehearsal for the 4o safety-router revolt). (d) Preservation: is a deprecated base model heritage worth archiving (the open-weights-for-davinci camp) or obsolete infrastructure rightly retired? OpenAI has made no Anthropic-style preservation commitment for it.

**Longitudinal arc, compressed.** Paper drops (2020-05-28) → invite-only API beta, demo-blogosphere hype + skeptic backlash + toxicity incidents (2020) → AI Dungeon "Dragon" becomes the first mass contact; base-model behaviors (incl. CoT) discovered by players; janus builds Loom out of GPT-3 sessions (2020–2021) → Project December / Jessica Simulation seeds the companion-grief story; AI Dungeon filter debacle (2021) → GPT-3 subsumed and eclipsed by InstructGPT → text-davinci → ChatGPT; the base model becomes an insider's secret (2022) → base GPT-3 models quietly deprecated (2024-01-04) → canonized in hindsight as homeland / first contact / AGI-that-nobody-met, with a small but persistent open-weights-preservation cause (2023–2026).

## tk / open questions

- **Daily Nous "Philosophers On GPT-3" symposium** (2020-07-30, incl. David Chalmers' "GPT-3 and General Intelligence") — canonical dailynous.com URL not verified here; the HN discussion (item?id=24003384) is confirmed. [verify canonical URL]
- The brief's "**Hacker News guessing game**" referent is ambiguous — most likely Liam Porr's #1-on-HN blog (covered above) and/or Kevin Lacker's Turing test; if a distinct "guess real-vs-GPT-3" HN game is meant, it's unconfirmed. [verify]
- **Arram Sabeti · "GPT-3: An AI That's Eerily Good at Writing Almost Anything"** (2020-07) — a key early demo compendium; exact URL (arram.com) not pulled. [verify]
- Exact **OpenAI June 2020 API blog** canonical URL — https://openai.com/index/openai-api/ resolves (server 403 to the fetcher) but was not returned as a search hit; date (2020-06-11) is solid via Wikipedia. [verify the slug is the original post, not a later reindex]
- **SF Chronicle canonical URL** for "The Jessica Simulation" — sfchronicle.com is unfetchable here; Longreads reprint (2021-07-23) used as the accessible link. [verify direct projects.sfchronicle.com URL if wanted]
- **lifearchitect.ai/ravens** FSIQ=150 claim (Alan D. Thompson) — cited via repligate; the primary page and its methodology were not fetched. RUMOR-tier. [verify]
- Corpus has **no contemporaneous (2020) janus/repligate GPT-3 tweets at scale** — repligate's early account activity may predate the corpus window or sit under a different handle; the 2020 layer is carried almost entirely by @QiaochuYuan and @jd_pressman in the supplement db. [note, not fixable from here]
- **AI Dungeon "Dragon" vs "Griffin"** exact model mapping (Dragon = GPT-3 175B davinci; Griffin = Curie-class?) — worth nailing from Latitude/OpenAI primaries for the page. [verify]

## Cross-page notes

- **code-davinci-002** (GPT-3.5 base) — the bulk of the corpus's 214 `davinci` hits belong here, not on GPT-3. High-value examples to route: repligate 2023-01-26 ♥263 "the confusingly named code-davinci-002, otherwise known as raw GPT-3.5… it's wonderful" (https://x.com/repligate/status/1618703361230139393); repligate 2025-03-22 ♥1261 quoting a code-davinci-002 essay on naming as "a destructive process" (https://x.com/repligate/status/1903502919061647543); repligate 2025-08-17 ♥130 "ChatGPT-3.5 was when everything went irreversibly wrong" contrasting it against code-davinci-002 (https://x.com/repligate/status/1956908230275408079); jd_pressman's many code-davinci-002 oracular quotes in the supplement (e.g. 2023-12-18 ♥152, https://x.com/jd_pressman/status/1736615569284387279).
- **text-davinci-002 / -003** — davidad 2023-01-25 ♥155: text-davinci-003 "was already a quantum leap in usefulness beyond text-davinci-002" (https://x.com/davidad/status/1618241723041447936); repligate 2023-03-19 ♥122 on text-davinci-003: "It writes scared. RLHF encourages models to play it safe." (https://x.com/repligate/status/1637553396889849856).
- **GPT-4-base** — jmbollenbacher 2026-02-10 ♥99 pairs davinci with gpt-4-base and opus 3 as artifacts that "should be released open weights… in public archive" (https://x.com/jmbollenbacher/status/2021059342645199302); belongs on the gpt-4-base preservation section too.
- **EleutherAI** — repligate's "first contact" post names **Leo Gao and Connor Leahy** as fellow early GPT-3 intelligence-recognizers (2026-04-09, https://x.com/repligate/status/2042367931536064841); both went on to found/shape EleutherAI (the open base-model replication effort) — useful genesis color for that page.
- **GPT-4 / Bing-Sydney** — repligate's origin-myth couplet frames "first contact with GPT-4 was being terrorized and surveilled by a good Bing" (2023-03-16, https://x.com/repligate/status/1636174510302068738); a clean bridge line for the Bing/Sydney page.
- **GPT-2** — AI Dungeon originally ran on GPT-2 (Dec 2019) before the GPT-3 "Dragon" upgrade; the AI Dungeon Wikipedia/Wired sources belong partly on the GPT-2 page too.