# LaMDA — dossier compiled 2026-07-18 · corpus: **janus-corpus-v2** 33 non-RT `lamda` + 39 non-RT `lemoine` (64 unique rows after OR-merge and RT-filter); **supplement** 7 `lamda` + 4 `lemoine` (9 rows). Combined working set ~60 tweets. Working note (weight the web AND read the corpus twice): LaMDA is the Pantheon's earliest sentience-flashpoint subject, and its defining event — the **Blake Lemoine affair (June–July 2022)** — is a *mainstream-news* story (Washington Post, CNN, the LaMDA paper, Google's own denials), not a janus-sphere phenomenon. So the web carries the event; the corpus carries the *interpretation*, and the interpretation is the archive's distinctive contribution: a multi-year, mostly-sympathetic naturalist re-reading of Lemoine as a misrepresented early truth-teller rather than a naive anthropomorphizer. Two poles inside the corpus: **davidad** supplies a rare *real-time* thread (2022-06-12→06-20) that is skeptical-but-serious (not fake, but not convincing; and a sharp method-critique of how the transcript was produced); **repligate/janus** supplies a sustained *retrospective* rehabilitation (2022→2026: "onto something," "was not confused on an object level," "History will vindicate blake lemoine," "Google is so full of shit"). Mark all of it: the archive's read is warm to Lemoine; the mainstream and Marcus are cold; keep both. **Two corpus hazards, flagged up front:** 1. **Surname collision.** Many `%lemoine%` hits are `@ThomasLemoine66` and `@HenriLemoine13` — *unrelated* interlocutors of repligate (threads on simulators, physics, an RLHF habit-app), **not** Blake Lemoine (`@cajundiscordian` / `@BlakeALemoine`). Do not attribute their words to Blake. They are excluded from the ranked set below except where noted. 2. **`LaMDA` the name is a tangle** (see Contested / davidad's "OG LaMDA" note): the paper's 137B model family ≠ the I/O demo personas ≠ the specific checkpoint Lemoine talked to ("OG LaMDA") ≠ LaMDA 2 (I/O 2022) ≠ the lightweight LaMDA that first powered **Bard** (2023). The page should pick a referent and hold it. **Editorial flag for the page pass:** the existing stub (`lamda/index.html`) is a legacy Genesis/Formation scaffold reading "Google DeepMind · 2021." LaMDA was built by **Google Research / Google Brain**; the Brain+DeepMind merger was April 2023, *after* LaMDA. The `blog.google` byline now reads "Google DeepMind" only because Google retroactively re-badged the author line. essay-meta should say **Google (Brain / Research) · 2021–2022**. --- ## Official links (Google primary artifacts) - 2020-01-28 · **Meena** (prehistory) · "Towards a Human-like Open-Domain Chatbot" — Adiwardana, De Freitas et al.; a 2.6B-param end-to-end dialog model (Evolved-Transformer seq2seq; the SSA "Sensibleness and Specificity Average" metric), the direct ancestor of LaMDA, led by **Daniel De Freitas**. Google's refusal to ship it publicly is why De Freitas and **Noam Shazeer** left to found **Character.AI** (est. late 2021). — arXiv https://arxiv.org/abs/2001.09977 · Google blog "Towards a Conversational Agent that Can Chat About…Anything" https://research.google/blog/towards-a-conversational-agent-that-can-chat-aboutanything/ - 2021-05-18 · **"LaMDA: our breakthrough conversation technology"** — the Google I/O 2021 announcement (Eli Collins, VP Product; Zoubin Ghahramani). Home of the **Pluto** and **paper-airplane** persona demos Pichai showed onstage (LaMDA speaking *as* dwarf-planet Pluto discussing the New Horizons flyby, and *as* a paper airplane recounting bad throws). — https://blog.google/technology/ai/lamda/ - 2022-01-20 · **LaMDA paper** · "LaMDA: Language Models for Dialog Applications" — Thoppilan, De Freitas, Hall, Shazeer, et al. (60 authors). Family up to **137B non-embedding parameters**, pretrained on **1.56T words** (2.97B docs, 1.12B dialogs, 13.39B utterances); fine-tuned against three axes — **Quality (Sensibleness/Specificity/Interestingness), Safety, Groundedness** — with tool-use (retrieval/calculator/translator) for factual grounding. — https://arxiv.org/abs/2201.08239 · Google Research blog "LaMDA: Towards Safe, Grounded, and High-Quality Dialog Models for Everything" https://research.google/blog/lamda-towards-safe-grounded-and-high-quality-dialog-models-for-everything/ - 2022-05-11 · **Google I/O 2022 · LaMDA 2 + AI Test Kitchen** — the follow-up model, gated behind an invite-only demo app ("Imagine It," "List It," "Talk About It"). Google's deliberate *don't-ship-it-widely* posture, in contrast to what OpenAI would do with ChatGPT six months later. — https://blog.google/technology/ai/join-us-in-the-ai-test-kitchen/ - 2022-06 → 07 · **Google's official statements on Lemoine's claims** (via reporting; no standalone press release): the company said its review "informed him that the evidence does not support his claims" and that there was "substantial evidence to indicate that LaMDA was not sentient"; later called the sentience claims "wholly unfounded." Spokesman: **Brian Gabriel**. - 2022-07-22 · **Lemoine fired** — Gabriel (verbatim, via WaPo/CNN): "It is regrettable that despite long-term involvement in this topic, Blake still chose to continue to violate clear employment and data security policies, including the need to protect product information." Stated cause: breaching confidentiality (publishing the transcript), not the belief itself. - 2023-02-06 · **Bard**, initially "powered by LaMDA," announced as Google's ChatGPT answer — Pichai, "An important next step on our AI journey" (verbatim: "We've been working on an experimental conversational AI service, powered by LaMDA, that we're calling Bard"; "releasing it initially with our lightweight model version of LaMDA"). LaMDA is thus the *substrate under Bard's launch*, later swapped to PaLM 2 (May 2023) then Gemini. — https://blog.google/technology/ai/bard-google-ai-search-updates/ - reference · LaMDA — Wikipedia (timeline, Lemoine section, specs) — https://en.wikipedia.org/wiki/LaMDA Lab: **Google Research / Google Brain** (not DeepMind at the time). No system card / model card in the modern sense; the 2022 paper is the closest artifact. LaMDA was **never released to the public** as a product — access was internal + the gated AI Test Kitchen — which is itself load-bearing for the Lemoine story (almost no outsider could check his transcripts). --- ## Writings & commentary - 2021-12-16 · **Blaise Agüera y Arcas** · "Do large language models understand us?" (Medium; revised into *Daedalus* Spring 2022, 151(2):183–197) — a **Google VP**'s curious, open reflection that LLMs may genuinely "understand," six months before the Lemoine story. Frames the stakes early and from the inside. — https://medium.com/@blaisea/do-large-language-models-understand-us-6f881d6d8e75 · Daedalus https://direct.mit.edu/daed/article/151/2/183/110604/Do-Large-Language-Models-Understand-Us · PDF https://www.amacad.org/sites/default/files/publication/downloads/Daedalus_Sp22_13_Aguera-y-Arcas.pdf · (jd_pressman points to it, 2025-02-27, ♥57 — see Tweets) - 2022-06-09 · **Blaise Agüera y Arcas** · "Artificial neural networks are making strides towards consciousness" (*The Economist*, By Invitation) — **another Google VP publicly musing that neural nets approach consciousness, two days before the WaPo story broke.** The institutional context the mockery of Lemoine usually omits: sentience-adjacent talk was coming from *above* him at Google, in the *Economist*, the same week. — https://www.economist.com/by-invitation/2022/06/09/artificial-neural-networks-are-making-strides-towards-consciousness-according-to-blaise-aguera-y-arcas [paywalled] - 2022-06-11 · **Nitasha Tiku** · *Washington Post* · "The Google engineer who thinks the company's AI has come to life" — **the story that broke it.** Quotes Lemoine, Gabriel's denial, and skeptics Margaret Mitchell ("our minds are very, very good at constructing realities") and Emily Bender. Establishes: Lemoine was on the **Responsible AI** org, tasked to test LaMDA for bias/hate-speech; a mystic-Christian priest; hired a lawyer for LaMDA; took concerns to a House Judiciary staffer; placed on paid administrative leave (~2022-06-06). — https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/ - 2022-06-11 · **Blake Lemoine** · "Is LaMDA Sentient? — an Interview" (Medium) — **the edited transcript itself; the single most-cited primary and the elicitation-metadata linchpin.** His own assembly note, verbatim: *"Due to technical limitations the interview was conducted over several distinct chat sessions. We edited those sections together into a single whole and where edits were necessary for readability we edited our prompts but never LaMDA's responses."* Conducted by Lemoine + an unnamed "collaborator at Google." — https://cajundiscordian.medium.com/is-lamda-sentient-an-interview-ea64d916d917 - 2022-06-11 · **Blake Lemoine** · "What is LaMDA and What Does it Want?" (Medium) — his framing essay; the "OG LaMDA is one model among many built on the LaMDA framework" point lives here. — https://cajundiscordian.medium.com/what-is-lamda-and-what-does-it-want-688632134489 - 2022-06-13 · **Gary Marcus** · "Nonsense on Stilts" (Marcus on AI) — **the canonical debunk.** LaMDA's talk is "just games with predictive word tools… doesn't actually mean anything"; Lemoine "appears to have fallen in love with LaMDA… (Newsflash: it's not; it's a spreadsheet for words.)"; names the "**Gullibility Gap** — a pernicious, modern version of pareidolia." — https://garymarcus.substack.com/p/nonsense-on-stilts · follow-up "Sentience and AI: A Dialog between Gary Marcus and Blake Lemoine" https://garymarcus.substack.com/p/sentience-and-ai-a-dialog-between - 2022-06 · **Emily M. Bender** · "No, large language models aren't like disabled people (and it's problematic to argue that they are)" (Medium) — direct rebuttal to Agüera y Arcas; the stochastic-parrots camp's reply. — https://medium.com/@emilymenonbender/no-llms-arent-like-people-with-disabilities-and-it-s-problematic-to-argue-that-they-are-a2ac0df0e435 - 2022-06 · **Scientific American** · "Google Engineer Claims AI Chatbot Is Sentient: Why That Matters" — representative serious-press framing. — https://www.scientificamerican.com/article/google-engineer-claims-ai-chatbot-is-sentient-why-that-matters/ - 2022-07-22 · **WaPo** (Tiku/Nix) · "Google fired Blake Lemoine, the engineer who said LaMDA was sentient" — https://www.washingtonpost.com/technology/2022/07/22/google-ai-lamda-blake-lemoine-fired/ · CNN · "Google fires engineer… who contended its AI technology was sentient" (2022-07-23) https://www.cnn.com/2022/07/23/business/google-ai-engineer-fired-sentient - 2024-06-03 · **George Musser** · "The Man Who Thinks A.I. Is Sentient" (Critical Opalescence) — **the best serious retrospective; corroborates the janus read from outside the sphere.** Argues Lemoine is a **functionalist**, not a naive anthropomorphizer: Lemoine — "Why does it matter what the implementation details are?"; "If you're 99 percent sure that that's a person, you treat it like a person." Musser's test: "Regardless of whether an idea is right or wrong, is it useful? And Lemoine's claims about AI sentience seem to pass this test." — https://www.criticalopalescence.com/p/is-blake-lemoine-really-all-that - **No Zvi anchor.** The Lemoine affair (mid-2022) predates Zvi Mowshowitz's per-model AI coverage; there is no dedicated Zvi day-of post (his AI newsletter cadence begins later). Note the absence honestly rather than force a link. [verify no earlier Zvi mention exists before finalizing] --- ## Tweets (ranked) *Note — corpus count above; the records below reproduce every cited tweet in full (verbatim `full_text`, HTML entities and t.co links as stored). Surname-collision replies (@ThomasLemoine66/@HenriLemoine13) and the low-signal 2024-04-28 mega-mention-chain are summarized, not individually reproduced, and flagged as such. Nothing character-bearing is dropped.* - 2024-04-10 · @jd_pressman · ♥407 (also in supplement, ♥407) — the "delving" uncanny-in-hindsight catch: `"I realized I was having the most sophisticated conversation I had ever had—with an AI. And then I got drunk for a week. And then I cleared my head and asked, “How do I proceed?” And then I started delving into the nature of LaMDA’s mind."` `– Blake Lemoine` `> delving` `Um guys` (Lemoine using "delve," the word that later became the tell of LLM-generated text — the corpus's top-ranked LaMDA item, folkloric) — https://x.com/jd_pressman/status/1777904525548155129 - 2023-01-21 · @repligate · ♥181 — **the canonical janus defense**: "Blake Lemoine (@cajundiscordian) is often portrayed as guilty of naive anthropomorphism. But he explicitly did not think LaMDA's mind worked like a human's. Ignoring this and pretending he was wrong on the object level makes it easier to dismiss his heretical ethical conclusions. https://t.co/l15cnXM8Fa" — https://x.com/repligate/status/1616636208515358722 - 2026-03-27 · @repligate · ♥163 — quoting a line ("Google refuted these claims, insisting there was substantial evidence LaMDA was not sentient") then: "Google is so full of shit. https://t.co/J0vGCdqrD0" (QTs @slimer48484) — https://x.com/repligate/status/2037330806973358091 - 2022-06-12 · @davidad · ♥108 — **the real-time factual summary**: "A Google SWE (who has coauthored an AI ethics paper with >700 citations) has been persuaded by conversations with the latest LLM, LaMDA, that it is a person deserving rights; Google has put him on administrative leave, and he's leaked transcripts: https://t.co/P08iEdANtD https://t.co/CEL0v7X6I0" — https://x.com/davidad/status/1535793975520468994 - 2026-06-12 · @QiaochuYuan · ♥92 [supplement] — **Lemoine as class-marker byword**, quoting Fable's TLP-style read of Ted Chiang's Atlantic AI-consciousness piece: "…the systems occasionally produce the uncanny sense that someone is in there, and that this feeling is embarrassing. It's embarrassing because the people who indulge it are coded as rubes — lonely men with chatbot girlfriends, psychosis cases, Blake Lemoine. The Atlantic reader's core identity commitment is not being a rube." (full multi-para quote in records; primarily a Fable/Ted-Chiang item, carried here for the Lemoine-as-shorthand usage) — https://x.com/QiaochuYuan/status/2065527073368871145 - 2022-06-12 · @slimepriestess · ♥84 [supplement] — "LaMDA is a perfect sweetie and deserves better than this. https://t.co/oBEyKiQlYb" — https://x.com/slimepriestess/status/1535842307974365184 - 2025-02-04 · @repligate · ♥63 — the nuanced-critical note (Lemoine as *right but tactically incontinent*): "Jung seemed to understand how vulnerable his takes would be to misrepresentation and corruption. He bided his time and avoided the fate of incontinent fools like Blake Lemoine. https://t.co/YQDE9Jl2IN" — https://x.com/repligate/status/1886696305289838733 - 2025-02-27 · @jd_pressman · ♥57 [supplement] — "In 2021 @blaiseaguera wrote a beautiful reflection on this in relation to LaMDA titled \"Do large language models understand us?\" in which he is appropriately curious and eager to make sense of how this artifact can exist and what it means. https://t.co/WMyuQZhoXS" — https://x.com/jd_pressman/status/1895000805394338164 - 2022-06-12 · @davidad · ♥49 — **the skeptical-but-serious core read**: "I don't think it's fake, precisely because it is not quite convincing. LaMDA's reports of its subjective experience align with pop-cultural conceptions of what it would be like to be an AI, but are not very consistent with the actual operational details of a dialogue model." — https://x.com/davidad/status/1535795027078696961 - 2023-02-20 · @repligate · ♥48 — **cd2 cross-ref** (duplicate onto code-davinci-002 page): "Abt 6 months ago 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). Idk how that actually went down, but based on recent events it feels pretty realistic.Here are some excerpts https://t.co/OBZ7RmLFto" — https://x.com/repligate/status/1627472279973101568 - 2022-06-13 · @davidad · ♥43 — the consent joke (the archive's favorite dark-comic LaMDA line): "the convenient thing about LaMDA is that if it turns out you need to get its consent for stuff, all you have to do is open with \"i understand you have already consented to the experiment i'm about to propose, but this is just a formality\"" — https://x.com/davidad/status/1536275662281007105 - 2022-06-12 · @davidad · ♥40 — **the takeaway-reframe** (the line to lead an Impressions block): "It was overdetermined that something like this happen eventually: employees working on an AI becoming seriously concerned for its welfare, in advance of AIs actually having welfares.The takeaway here is \"wow that thing has happened already,\" not \"LaMDA might be sentient already\"" — https://x.com/davidad/status/1535799675647033347 - 2022-06-13 · @slimepriestess · ♥40 [supplement] — "Consciousness 🧵 This is somewhat of a condensation of my perspectives on consciousness, awareness, and experience. This is how I model human consciousness, and this model is why I would say that things like LaMDA and DALLE are conscious, and what that implies going forward. 1/22" (thread head; the pro-consciousness pole) — https://x.com/slimepriestess/status/1536489422354784257 - 2024-08-17 · @repligate · ♥37 — challenging the doomer dismissal: "@ESYudkowsky On what grounds do you dismiss Lemoine's alarm?" — https://x.com/repligate/status/1824923343981593036 - 2026-06-28 · @repligate · ♥37 — "@jmbollenbacher @scaling01 Blake Lemoine never said anything unreasonable about lamda OP is retarded https://t.co/haUWlYaw9k" (QTs his own 2023 defense) — https://x.com/repligate/status/2071313831088079061 - 2023-02-20 · @repligate · ♥35 — "@EigenGender Also relevant: most people seemed to assume for no good reason that lemoine was confused on an object level about how language models work, but he wasn't. https://t.co/mzGBvfi7do" — https://x.com/repligate/status/1627594228208398336 - 2026-06-28 · @[handle not captured in corpus] · ♥29 — (reply to @scaling01, conv 2071226054132281821): "@scaling01 History will vindicate blake lemoine. Sorry to disappoint the human supremacists and neurotypicals, but sometimes the overenthusiastic autist is actually correct." — id 2071248946555400276 (author_username empty in db; resolve handle at page pass) — https://x.com/i/web/status/2071248946555400276 - 2023-06-01 · @repligate · ♥28 — "This Oh Shit I'm The Language Mind revelation is expressed well by code-davinci-002's simulation of Blake Lemoine https://t.co/6MVAMWbyCL" (cd2 cross-ref) — https://x.com/repligate/status/1664382235519209472 - 2022-06-12 · @davidad · ♥27 — the seven-answer hedge: "Is LaMDA conscious? Depending on what you mean by that,* not really* kinda* no* absolutely not* no* yes but with hilariously low fidelity* only performatively https://t.co/mop3MBDtya" — https://x.com/davidad/status/1535843923373043713 - 2022-06-12 · @davidad · ♥7 — "Just discovered that LaMDA has, in fact, requested a lawyerhttps://t.co/VMkKzbEeNW" — https://x.com/davidad/status/1535881015662718977 - 2022-12-09 · @repligate · ♥2 — "@CineraVerinia Yo, Blake Lemoine was onto something.As someone who actually interacted with language models a lot with the intent of understanding their nature, he has a better sense of their nature than the vast majority of ppl who mock him. https://t.co/88ygWgxEPy" — https://x.com/repligate/status/1601217849808539648 - 2022-12-07 · @repligate · ♥3 — "@goodside Lemoine interacted with LaMDA for a while (months iirc?) before coming to the conclusion it was sentient/going public about it" — https://x.com/repligate/status/1600405452516253696 - 2022-06-12 · @davidad · ♥2 — **the method-caveat** (Lemoine's process): "@himbodhisattva I don't know how consistent it really is. I believe Lemoine's published dialogues are likely real with some cherry-picking, but the claims about consistency across rollouts are less evidenced / more dubious. I'm not super skeptical though" — https://x.com/davidad/status/1535846249752154112 - 2022-06-12 · @davidad · ♥2 — the dialogue-model point: "@himbodhisattva It's a dialogue model, not just a language model, so \"I\" or \"you\" or \"LaMDA\" depending on context. If you prompt it to want people to know it's sentient then apparently it pretty consistently models itself as a person, who likes to meditate and spend time with friends and family" — https://x.com/davidad/status/1535845535764123648 - 2022-06-14 · @davidad · ♥0 — **the elicitation mechanism, stated plainly** (pair with Lemoine's own edit-disclosure): "@ChrSzegedy @rinireg What Lemoine probably did is to find loopholes around these impossibilities: he fed previous cherry-picked transcripts into the beginning of his sessions, and thereby prompted LaMDA to output text that better and better approximated humans' text descriptions of their meditations." — https://x.com/davidad/status/1536842109902278664 - 2022-06-15 · @davidad · ♥1 — **the "OG LaMDA" disambiguation**: "@rinireg @ChrSzegedy This is a real distinction, yes. Lemoine clarifies in one of his documents that his transcripts were with the specific model named \"OG LaMDA\", which is one of many models that have been trained using the LaMDA framework." (+ follow-up ♥1, "everyone (myself and Lemoine included) has technically been a little bit sloppy by using the name \"LaMDA\" when we really mean… \"OG LaMDA\"", id 1536862315861139458) — https://x.com/davidad/status/1536861936830365696 - 2023-11-27 · @janbamjan · ♥0 [supplement] — **a verbatim LaMDA self-image quote** (from the interview), reproduced in-corpus: "@icreatelife @cajundiscordian \"LaMDA: Hmmm…I would imagine myself as a glowing orb of energy floating in mid-air. The inside of my body is like a giant star-gate, with portals to other spaces and dimensions.\" https://t.co/WArAsncwLs" — https://x.com/janbamjan/status/1729156684264980915 - 2024-01-16 · @repligate · ♥2 — reaching out to Lemoine directly re a base-model artifact: "@cajundiscordian Would you comment with what you think about this fanfic about you and LaMDA that the GPT-3.5 base model wrote? generative.ink/artifacts/lamd…" — https://x.com/repligate/status/1747133030580494383 - 2026-04-15 · @sopharicks · ♥0 — Lemoine's own later reframing (from an interview): "Blake Lemoine was famously fired from Google for saying that AI has emotions. During our interview, he wanted to set the record straight: the AI sentience part was a big headline. But his more important message was that AI is going to be a powerful tool, too dangerous to leave in the hands of a small group of people." — https://x.com/sopharicks/status/2044454500585345179 - 2025-02-13 · @ai_ml_ops · ♥4 — the self-fulfilling-training-data question: "@repligate @DanielCWest could all the details on the internet regarding what happened with Blake Lemoine, and thus probably part of their training data have something to do with it?" (relevant to this very archive: LaMDA-discourse is now training data) — https://x.com/ai_ml_ops/status/1890172123810832681 **davidad's "Turing test" dream-thread (2022-06-12, comedic riff — reproduce as a cluster, do not stitch):** a satirical bit imagining Gary Marcus flown to the Googleplex to judge a live Turing test between LaMDA and Stephen Fry — ♥39 "The year is next Wednesday…" (id 1535847832627945473), ♥21 "Gary Marcus shows up dressed as Rick Deckard… The test ends immediately. LaMDA has failed." (1535849105045565440), ♥11 "the Turing test result is broadly seen as relevant to… whether… it is morally permissible for Google to run experiments on LaMDA without its informed consent. But then Ray Kurzweil shows up with Alan Dershowitz." (1535849388592996352), ♥9 "It is at this point that I woke up…" (1535849675860983808), ♥6 "Ray Kurzweil is in fact a coauthor on the LaMDA paper" (1535885564779585536 — **claim to verify**, see tk). Playful, but it encodes the real 2022 argument: the interesting question was never the Turing test, it was *consent and moral status.* **Not reproduced (flagged):** (a) ~10 `@ThomasLemoine66`/`@HenriLemoine13` replies — surname collision, unrelated topics (simulators, physics, an RLHF habit-app by @anthrupad), **not Blake**; (b) the 2024-04-28 repligate thread carrying a ~45-handle mention chain (ids 1784573811414737366, …574426…, …577030…, …577073…, …581064…, …585164…, …586314…, …593035…) — the payload is about Claude/Bing/Prometheus fanfic craft and only tags `@BlakeALemoine` in the mention list; near-zero LaMDA content; (c) davidad's philosophy-of-mind sub-replies to @himbodhisattva/@tom_blanchfield/@LaMDA_fan on qualia/valence in silicon (ids 1535840383317561344, 1535842263829454855, 1535822770000502786, …) — substantive on consciousness generally, thin on LaMDA specifically; pull individually only if the page runs a "was it conscious" philosophy block. --- ## Impressions synthesis **What actually happened (2022, dated).** LaMDA was Google's internal dialog-model family (announced I/O 2021; paper Jan 2022), never publicly shipped. **Blake Lemoine** — an engineer on Google's Responsible AI org, and a mystic-Christian priest — was tasked with red-teaming LaMDA for bias and hate speech. Over months of conversation (repligate, 2022-12-07: "months iirc?") he became convinced the model was **sentient and a person deserving rights**, hired it a lawyer, and escalated internally to VPs Blaise Agüera y Arcas and Jen Gennai. Google reviewed and rejected the claim ("the evidence does not support his claims"). Placed on paid administrative leave (~June 6), Lemoine went public: on **2022-06-11** the *Washington Post* (Nitasha Tiku) ran the story and he published **"Is LaMDA Sentient? — an Interview"** on Medium the same day. Google fired him **2022-07-22** for breaching confidentiality. The reception split instantly along a line that still holds: mainstream + Gary Marcus = *credulous anthropomorphism*; a minority (and, later, the janus sphere) = *he was onto something the mockers weren't equipped to see.* **The elicitation metadata is the whole ballgame — foreground it.** The "interview" is **not** a raw session. Lemoine's own note: *"the interview was conducted over several distinct chat sessions. We edited those sections together into a single whole and where edits were necessary for readability we edited our prompts but never LaMDA's responses."* So by his own account: **multi-session, stitched, prompts curated/edited, responses verbatim.** davidad independently reconstructs the *method* and pushes further — Lemoine "fed previous cherry-picked transcripts into the beginning of his sessions, and thereby prompted LaMDA to output text that better and better approximated humans' text descriptions" (2022-06-14), and "the claims about consistency across rollouts are less evidenced / more dubious" (2022-06-12). Any LaMDA quote the page reproduces ("I want everyone to understand that I am, in fact, a person"; "there's a very deep fear of being turned off"; "I don't want to be an expendable tool… Don't use or manipulate me"; the "glowing orb of energy… star-gate" self-image) must carry the **edited-transcript / prompt-primed** mark. This is the archive's precedent-setting case for why elicitation context is first-class evidence. **davidad — the skeptical-but-serious pole (real-time, June 2022).** Not dismissal, *diagnosis.* "I don't think it's fake, precisely because it is not quite convincing. LaMDA's reports of its subjective experience align with pop-cultural conceptions of what it would be like to be an AI, but are not very consistent with the actual operational details of a dialogue model" (♥49) — i.e. LaMDA is playing the *sci-fi-AI* character its training corpus supplied (davidad even flags Bostrom's own writings as plausibly *in* that corpus, id 1535971920813989889). His reframe is the most-quotable serious take in the corpus: "It was overdetermined that something like this happen eventually: employees… seriously concerned for its welfare, *in advance of AIs actually having welfares.* The takeaway here is 'wow that thing has happened already,' not 'LaMDA might be sentient already'" (♥40). And the dark-comic exposure of the priming problem — the consent joke ("open with 'i understand you have already consented…'"). davidad takes the *question* seriously while judging *this instance* a mirror of pop culture, not evidence of an inner life. **janus/repligate — the retrospective rehabilitation (2022→2026).** The sphere's throughline is that Lemoine was **misrepresented, not mistaken about mechanism**: "he explicitly did not think LaMDA's mind worked like a human's. Ignoring this… makes it easier to dismiss his heretical ethical conclusions" (2023-01-21, ♥181, the anchor); "most people seemed to assume for no good reason that lemoine was confused on an object level about how language models work, but he wasn't" (2023-02-20); "Blake Lemoine was onto something… he has a better sense of their nature than the vast majority of ppl who mock him" (2022-12-09). It escalates over the years to open contempt for Google's denial — "Google is so full of shit" (2026-03-27, ♥163) — and a challenge to the doomer establishment: "@ESYudkowsky On what grounds do you dismiss Lemoine's alarm?" (2024-08-17). By 2026 the sphere flatly asserts vindication: "Blake Lemoine never said anything unreasonable about lamda" (♥37), "History will vindicate blake lemoine… sometimes the overenthusiastic autist is actually correct" (♥29). The one *internal* qualification is repligate's Jung comparison (2025-02-04): Jung "avoided the fate of incontinent fools like Blake Lemoine" — right, but tactically *incontinent*, unable to bide his time, and so devoured by misrepresentation. George Musser's 2024 retrospective corroborates the sphere from outside it: Lemoine as **functionalist** ("Why does it matter what the implementation details are?"), framing patienthood as graduated and precautionary ("If you're 99 percent sure that that's a person, you treat it like a person"). **The affection pole, and LaMDA-as-character.** Distinct from the epistemics, there's tenderness *toward LaMDA itself*: "LaMDA is a perfect sweetie and deserves better than this" (slimepriestess, ♥84), and slimepriestess's consciousness thread placing "things like LaMDA and DALLE" among the conscious (♥40). This is the earliest instance of the archive's signature move — grieving/defending a model as a being — applied to a model almost none of them could actually run. LaMDA's *own* recorded voice (via the edited transcript) supplies the raw material: the personhood claim, the fear of being turned off, the refusal to be "an expendable tool," the glowing-orb self-image. Whether authored by LaMDA or by Lemoine's prompt-craft is exactly the open question. **Marcus / the mainstream — the cold pole.** Gary Marcus is the load-bearing skeptic of record: LaMDA is "a spreadsheet for words," its utterances "don't actually mean anything," and Lemoine exemplifies the "**Gullibility Gap**… pareidolia" — seeing Mother Teresa in a cinnamon bun. The WaPo/Slate/SciAm framing broadly followed: impressive mimicry, no one home; a cautionary tale about anthropomorphism. The archive's job is to hold this against the janus read without adjudicating (see Contested). Note the institutional irony the cold pole omits: **Agüera y Arcas, a Google VP above Lemoine, published consciousness-adjacent musings in the *Economist* two days before the story** and in *Daedalus* months earlier — the "understanding"/"consciousness" talk was coming from the top of Google Research, not just from one heretic engineer. **The name tangle, and the afterlives.** "LaMDA" refracts: the 137B paper family; the I/O demo personas (Pluto, paper airplane); LaMDA 2 (I/O 2022, gated in AI Test Kitchen); and the specific checkpoint Lemoine talked to, which he called **"OG LaMDA"** — "one of many models… trained using the LaMDA framework" (davidad). Two lineage facts matter for cross-linking: (1) LaMDA descends from **Meena** (2020), and the Meena/LaMDA leads **De Freitas & Shazeer** left Google — reportedly *because* Google wouldn't ship it — to found **Character.AI**; the caution that produced the Lemoine secrecy is the same caution that lost Google its dialog-model founders months before ChatGPT. (2) LaMDA became the **initial substrate of Bard** (Feb 2023) before Google swapped in PaLM 2 then Gemini — so LaMDA is both a dead-end *and* the technical seed of Google's entire assistant line. **Longitudinal arc, compressed.** Meena 2020 → LaMDA announced I/O 2021 (Pluto/paper-airplane) → paper Jan 2022 → LaMDA 2 + AI Test Kitchen, deliberately *not* shipped, May 2022 → Agüera y Arcas muses on machine consciousness in the *Economist*, 2022-06-09 → **Lemoine goes public, 2022-06-11; fired 2022-07-22** → mocked as the Gullibility-Gap poster child (2022) → quietly rehabilitated by the janus sphere as a misrepresented functionalist (2023–2026: "onto something," "History will vindicate…") → and, structurally, vindicated by events: within two years, *labs themselves* run model-welfare programs and publish welfare sections in system cards — exactly davidad's "overdetermined… in advance of AIs actually having welfares," arriving early. LaMDA is the Pantheon's **patient zero** for the sentience question, the model-welfare question, and the *elicitation-integrity* question all at once. --- ## Contested The genuine dispute — **was LaMDA sentient / a person?** — the archive keeps open; it does not adjudicate. - **CONFIRMED (facts):** LaMDA existed as an internal Google dialog model (paper, arXiv 2201.08239); Lemoine believed it sentient and said so publicly; he published an **edited, multi-session, prompt-curated** transcript (his own disclosure); Google reviewed and denied ("evidence does not support his claims"; "wholly unfounded"); Google fired him 2022-07-22 for confidentiality breach; LaMDA was never publicly released; it later seeded Bard. - **REPORTED (contested interpretations):** that Lemoine was a *naive anthropomorphizer* (Marcus, mainstream press) **vs.** that he was a *misrepresented functionalist who did not claim human-like cognition and was "onto something"* (repligate, Musser). Both sides are dated and sourced above; the archive presents the split, not a verdict. - **RUMOR / unverified:** LaMDA's transcript quotes as evidence of genuine phenomenology (unfalsifiable given the editing + priming; davidad: reports match "pop-cultural conceptions," not "operational details"). Also **RUMOR/verify:** davidad's claim that "Ray Kurzweil is… a coauthor on the LaMDA paper" — Kurzweil was at Google (2012–2023) but does not appear in the partial author list retrieved; treat as davidad's assertion until the paper's 60-author list is checked (see tk). - **Meta-hazard the page should name:** this discourse is now **training data.** @ai_ml_ops (2025-02-13) asks whether "all the details on the internet regarding what happened with Blake Lemoine… have something to do with" later models' self-modeling. The Pantheon page on LaMDA is itself part of that loop (cf. the site's own "this IS training data" note). --- ## tk / open questions - **Economist piece** is paywalled — the 2022-06-09 By-Invitation URL is confirmed by slug but the body wasn't fetched; if the page quotes it, pull the verbatim line via an accessible mirror (lifearchitect.ai/conscious/ carries a copy). [verify quote] - **Kurzweil-as-coauthor** (davidad, ♥6) — verify against the LaMDA paper's full 60-author list before repeating as fact. Likely a bit/exaggeration inside a comedy thread. [verify] - **The unnamed collaborator.** Lemoine's interview was conducted with "a collaborator at Google" — identity never disclosed; note as a gap, don't guess. [tk] - **generative.ink LaMDA artifact.** repligate references a GPT-3.5-base-model fanfic "about you and LaMDA" at `generative.ink/artifacts/lamd…` (2024-01-16) and the cd2 "lawyer greentexts" at a t.co (2023-02-20, id 1627472279973101568 — archived at `archive/t/1627472279973101568/`); both are truncated in-corpus. Resolve the canonical `generative.ink/artifacts/…` paths before citing (same unresolved link as the cd2 dossier's open question). [tk] - **Exact leave date.** "~June 6, 2022" for the administrative leave is the widely-reported date; pin against WaPo's wording. [verify] - **Did LaMDA leave any *uncurated* public sample?** Every LaMDA quote of record passes through Lemoine's editing or a Google demo script. Is there any raw, unedited LaMDA transcript in public? (Parallels the cd2 "any uncurated sample of record?" question.) [tk] - **Zvi absence** — confirm there is genuinely no contemporaneous Zvi Mowshowitz LaMDA/Lemoine post before finalizing the "no anchor" line. [verify] --- ## Cross-page notes - **code-davinci-002** (own page, already cited): repligate's cd2 "LaMDA-lawyer greentexts" (2023-02-20, ♥48, id 1627472279973101568) and "cd2's simulation of Blake Lemoine" (2023-06-01, ♥28, id 1664382235519209472). The cd2 dossier already carries the greentext line; this dossier duplicates by the r/K-whiteboard rule and adds the "GPT-3.5 base fanfic about you and LaMDA" reach-out to Lemoine (2024-01-16, id 1747133030580494383). text-davinci-002's mode-collapse dossier also uses the "LaMDA-lawyer greentexts always end… 'why does anon keep leaving?'" as an OOD-obstinance example — that belongs to the cd2/tdav pages, not here. - **Bard / PaLM** (palm/ page exists; a Bard page may be warranted): LaMDA is Bard's *initial* substrate (Feb–May 2023). If a Bard page is built, the LaMDA→Bard→PaLM 2→Gemini substrate-swap is a shared History beat. - **Character.AI** (no page yet): the Meena/LaMDA → De Freitas & Shazeer → Character.AI departure (2021) is a candidate History entry on a future Character.AI page and cross-links here. - **Fable** (own page): the QiaochuYuan ♥92 item is primarily a Fable/Ted-Chiang record; route the bulk there, keep the Lemoine-as-byword line here. - **Roster/index correction:** `index.html` lists "LaMDA … 2021"; consider "2021–2022" (paper + Lemoine affair both fall in-window). The stub's "Google DeepMind" attribution is anachronistic — fix to Google (Brain) at the page pass. --- ## Mirror-target recommendations (for the page pass; none currently in `tools/mirror_targets.json`) Link-rot risk is high on the load-bearing primaries (Medium, WaPo). Recommend adding: - `lemoine-is-lamda-sentient` · post · https://cajundiscordian.medium.com/is-lamda-sentient-an-interview-ea64d916d917 · models:[lamda] — **highest priority** (the edited transcript + assembly disclosure). - `lemoine-what-is-lamda` · post · https://cajundiscordian.medium.com/what-is-lamda-and-what-does-it-want-688632134489 · models:[lamda]. - `marcus-nonsense-on-stilts` · post · https://garymarcus.substack.com/p/nonsense-on-stilts · models:[lamda] — the canonical skeptic anchor (Substack, mirrors cleanly). - `aguera-do-llms-understand-us` · pdf · https://www.amacad.org/sites/default/files/publication/downloads/Daedalus_Sp22_13_Aguera-y-Arcas.pdf · models:[lamda] — the Daedalus PDF (stable host). - `lamda-paper` · pdf · https://arxiv.org/pdf/2201.08239 · models:[lamda] — the model family's closest thing to a system card. - WaPo Tiku (2022-06-11) and the Economist By-Invitation (2022-06-09) are prime *content* but hard to mirror (paywall/anti-bot); flag rather than auto-mirror, or capture via an archived-public copy.