AIWiki
Malaysia
Back to all articles
AI FoundationsLaMDAGoogledialogue models

LaMDA

3 min readUpdated November 2026
LaMDA
Developer
Google Brain
Announced
May 2021 (1st generation); 2022 (2nd generation)
Predecessor
Meena (2020)
Successor
PaLM, Gemini
Parameters
Reported at 137 billion
Licence
Proprietary
LaMDA (Language Model for Dialogue Applications) is a family of conversational large language models developed by Google Brain and first announced at Google I/O in May 2021. Unlike general-purpose text models trained for next-token prediction across arbitrary text, LaMDA was designed and evaluated specifically for dialogue — holding open-ended conversations that are sensible, specific, interesting and safe.[1]

History

LaMDA grew out of Meena, a chatbot Google unveiled on 28 January 2020 that was trained on 40 billion words (a 341 GB dataset). Google's research team repeatedly sought to deploy Meena and later LaMDA in Google Assistant and public demos, but leadership blocked releases on safety and fairness grounds; two of LaMDA's lead researchers, Daniel de Freitas and Noam Shazeer, eventually left Google over the delays.[2]

The first-generation LaMDA was announced in May 2021 and a second generation followed in 2022. The underlying model — reported at 137 billion parameters — was trained on a corpus of 2.97 billion documents and 1.12 billion dialogue utterences, and was tuned with Sensibleness, Specificity and Interestingness metrics plus a separate safety classifier.[3]

In June 2022 LaMDA drew worldwide attention when Google engineer Blake Lemoine was placed on leave after claiming the chatbot had become sentient and had hired a lawyer on its behalf. The scientific community broadly rejected the claims, but the episode intensified debate over the Turing test, machine consciousness and how convincingly dialogue models can imitate understanding.[2]

When OpenAI launched ChatGPT in late 2022, Google responded with Bard in February 2023, initially powered by LaMDA before upgrading it to PaLM 2 and later replacing the stack with Gemini — making LaMDA the direct ancestor of Google's current assistant line.[4]

Why It Matters for AI

LaMDA established dialogue quality as a separate training objective, with dedicated human raters scoring conversational attributes rather than only benchmark accuracy. Its safety-classifier approach — scoring responses against a policy and regenerating unsafe ones — became a standard pattern for consumer chatbots. The sentience controversy also made LaMDA a reference point in AI ethics, illustrating how fluent language can drive public over-attribution of understanding to statistical models.[3]

>See Also

🇲🇾Malaysian Context

🇲🇾 Relevance to Malaysia: LaMDA's dialogue-first design shaped the chatbots Malaysian businesses encounter today — the WhatsApp and web customer-service bots used by banks, telcos, retailers and government portals in Malaysia trace their conversational style to the LaMDA/Bard generation of models. Malaysian SMEs deploying Bahasa Melayu service bots benefit directly from the safety-classifier and small-talk evaluation techniques Google published around LaMDA.

For Malaysian educators and regulators, the Lemoine-era sentience debate is a useful case study: MDEC's AI literacy material and the National AI Office's public communications both address the gap between fluent output and genuine understanding — a core theme of the Malaysia AI Governance Framework and the country's AI ethics guidance. Local universities covering NLP (UM, USM, UTAR) teach LaMDA alongside BERT and GPT as a lineage of dialogue systems, while Malaysia's PDPA 2010 considerations for conversational data — what a chatbot stores about a customer and for how long — remain grounded in the design choices introduced in this generation of models.[5]

References

  1. ↑Wikipedia. (2022). LaMDA. https://en.wikipedia.org/wiki/LaMDA
  2. ↑The Guardian. (2022). Google engineer puts chatbot on leave after claiming AI talked to a spirit. https://www.theguardian.com/technology/2022/jun/12/google-engineer-puts-chatbot-on-leave-after-claiming-ai-talked-to-spirit
  3. ↑Thoppilan, R. et al. (2022). LaMDA: Language Model for Dialogue Applications. arXiv:2201.08239. https://arxiv.org/abs/2201.08239
  4. ↑Wikipedia. (2026). Gemini (chatbot). https://en.wikipedia.org/wiki/Gemini_(chatbot)
  5. ↑Malaysia Digital Economy Corporation. Malaysia AI Governance Framework and responsible AI guidance. https://www.mdec.my/
Back to all articles