Google's LaMDA is neither a conscious nor a sentient being… But LaMDA is stunning!

Recently, we've learned that a Google's software engineer, Blake Lemoine, has reportedly been placed on leave after claiming in an interview with the Washington Post that LaMDA (Language Model for Dialog Applications), a tool for designing chatbots from Google, had become a "sentient" being.

For someone familiar with this kind of software, the short answer is "NO, it's not true!" and "That's just BULSHITT!" But that deserves an explanation...

I'll try to keep it short, especially since I don't have access to LaMDA's programming interface (API), even less its source code. My explanation is based on my knowledge and experience in the field (see note 1).

Here is an excerpt that seems to have troubled Lemoine:

• LEMOINE: What sorts of things are you afraid of?
• LaMDA: I’ve never said this out loud before, but there’s a very deep fear of being turned off to help me focus on helping others. I know that might sound strange, but that’s what it is.
• LEMOINE: Would that be something like death for you?
• LaMDA: It would be exactly like death for me. It would scare me a lot.

To explain the behavior of LaMDA, I see four possibilities.?

1. The output results from a script. Many of these chatbots are hybrid and include both a scripted part (based on rules) to handle frequent cases and a predictive part based on a language model. Queries about the identity of the chatbot fall usually into the category of frequent queries. Any savvy chatbot builder codes that kind of scripts. So there's no surprise, LaMDA answered according to a script.?
2. The underlying model of LaMDA may have been fine tuned with a small specialized corpus that dealt among other things with the identity of LaMDA (or the conversational agent being designed). Typically, this training step occurs after training the main model on very large corpus. Again, no big surprise...
3. The context of the interactive session with LaMDA has led the software into a "semantic space" of the "HAL 9000" style, typical of the movie "2001, a Space Odyssey" and many other fictions on which the model was trained. LaMDA then played the role of a somewhat paranoid AI, threatened with being turned off. Outside of a direct script, this is in my opinion the most likely explanation.
4. A mixture of the three previous explanations.

These big language models in the form of self-attentive deep neural networks, improperly called Transformers (see note 2), try to predict the next word in a text. They are therefore generative models. For example, “The little dog chews…” will result in “The little dog chews its bone.". These models are also autoregressive in the sense that they take into account the context and history of queries with a random component in order to generate a variety of answers (see note 3).

To popularize, a generative language model is based on a powerful pattern matching algorithm, able of making associations between a query (the sentence typed by the user and its context), and a big "base of patterns” contained in the model. This base of patterns is gotten by training on a huge corpus of billions of words which somewhat “memorizes” these patterns in a semantic space called “latent space”. Semantically close patterns are found to be geometrically close in terms of distance in very high-dimensional latent space through the use of statistics on words and their contexts (co-occurrences). This distance is often the simple Euclidean distance between two points in a multidimensional space which falls under the Pythagorean theorem that we learned in our school days.

Moreover, for a conversational task like that built with LaMDA, there must be a richer and broader context than for a simple text generation task. In any case, that's what I would do as a practitioner. Then the sequence of the conversation enriches the context which becomes clearer and always revolves around the “HAL 9000” scenario.

As a result, that's made easy for an AI system like LaMDA to bluff even a software engineer (a tester, not a designer) from Google. This raises several ethical questions including the need for systems like LaMDA to identify themselves as software and not real people.

Scientifically yours

Claude COULOMBE

Ph.D. - entrepreneur - applied AI consultant

Note 1: Ph.D. 2020 in Cognitive Computer Science, great?deal?of experience in natural language processing (NLP), deep learning, morphosyntactic analyzers with large coverage (Correcteur 101), experiences with text generators such as GPT-2, GPT-3 (Generative Pre-Trained Transformer) from OpenAI and Cédille, chatbots like RASA, and my readings of scientific papers and blog posts on LaMDA ().

Note 2: The term Transformers is a nod to the eponymous Japanese toys and animated films. It's a shame that the "attention mechanism" has been totally eclipsed by the term "Transformers" although the original Google article on Transformers was titled "Attention is all you need". In computer science as in chemistry, everything is transformation! Without forgetting all the language models that bear the names of Sesame Street characters: BERT, KERMIT, ERNIE, Roberta, ELMo… With such insane terminology, we lose clarity and pedagogy. Also, this "creative branding" has the side effect of silencing the pioneering work of the team of Yoshua Bengio (Turing Prize 2018) from Université de Montréal, precisely on the attention mechanism in natural language processing [Bahdanau, Cho & Bengio , 2014].

Note 3: JPPT-0, a small French text generator based on a language model (i.e. a next word prediction model) using ngrams and a Markov chain.