Google’s LaMDA; A Next-Generation Language Model for Conversation
Google’s LaMDA; A Next-Generation Language Model for Conversation
  • Reporter Han Sang-yun
  • 승인 2021.06.25 21:42
  • 댓글 0
이 기사를 공유합니다

▲Conversation with LaMDA / Google
▲Conversation with LaMDA / Google


“You would get to see a massive canyon, some frozen icebergs, geysers, and some craters.” This is not what a human said, but what an AI responded, taking the role of Pluto when it was asked, “Tell me what I would see if I visited (you).”
Google’s new conversational machine learning program, LaMDA (short for “Language Model for Dialogue Applications”) is drawing attention due to its proficient, human-like words. LaMDA was announced at the 2021 Google I/O (Innovation in the Open), Google’s annual developer conference, which aired online from May 18 to 19 at Mountain Dew’s head office in California. It is developed for dialogue applications and is designed to talk about any topic.
Prior to LaMDA, several language models were developed. Among them, the best were BERT and GPT-3. BERT (Bidirectional Encoder Representations from Transformers) was released by Google in Nov. 2018. GPT-3 (Generative Pre-trained Transformer-3) was released by OpenAI in June 2020. Both were built on Transformer, the next-generation neural network architecture invented and open-sourced by Google in 2017 for language comprehension. Transformer needs less computation to train and is a better fit for modern machine learning hardware. These models are trained with many sentences and taught how words relate to one another. Thus, when making new sentences, they predict what words will come next to a given word.
Like BERT and GPT-3, LaMDA is also built on Transformer. However, while other models were trained on plain text, LaMDA was trained based on dialogue. By detecting the nuance of conversations, LaMDA learned what makes a conversation natural. For example, for “I just started taking guitar lessons”, there are many possible responses. “I do not know” makes sense, but it does not make the dialogue sound natural. “That is nice” is good, but it stops any more conversation. A more specific response would be realistic such as, “How exciting! My mom has a vintage Martin that she loves to play.” Thus, LaMDA equipped sensibleness and specificity in its response. Google is now improving LaMDA to possess other aspects of natural conversation such as interestingness, by assessing whether responses are insightful or witty, and factuality, by ensuring its response is not only compelling but also correct. 
However, conversational AIs, including LaMDA, are subject to misuses such as mirroring hateful speech, replicating misleading information, or leaking personal information. Google stated in an agreement on concerns on privacy and ethics, “Language might be one of humanity’s greatest tools, but like all tools it can be misused.” They also added their promises on the concerns, “Our highest priority, when creating technologies like LaMDA, is working to ensure we minimize such risks.” Google established their own AI principles and restricted their AI adhere to the principles. Google’s AI principles include the following statements; “be socially beneficial”, “avoid creating or reinforcing unfair bias”, “be built and tested for safety”, “be accountable to people”, “incorporate privacy design principles”, and “uphold high standards of scientific excellence.”
LaMDA is in an early stage. Google CEO Sundar Pichai stated, “It is still early research, so it does not get everything right. Sometimes it can give nonsensical responses.” After further development, Google plans to apply LaMDA on Google Assistant, Google Search, and Translation. In the past several updates on Google Translation, the company has drastically improved through deep learning technologies. LaMDA is expected to provide another big leap in translation. 
Besides LaMDA, Google has developed various AIs with wide applications. MUM (Multitask Unified Model) is another AI presented in 2021 Google I/O. It improves understanding of human questions and helps search. Its aim is to give a proper answer for questions that do not have simple answers such as “I have hiked Mt. Adams and now want to hike Mt. Fuji next fall. What should I do differently to prepare?” At the same event, Google also announced several AI-powered improvements on Google Maps to provide more useful information for users. Thus, Google is developing its AI tech not only to become more powerful, but also to provide more ease and functionality to its uses in everyday life such as translation, map, and search.