# Google Introduces LaMDA: Assess Its Sentience for Yourself
Written on
Chapter 1: The LaMDA Phenomenon
Two months ago, the spotlight was on Google’s AI system, LaMDA. A lively debate around its potential sentience captivated audiences for weeks. Now, Google has decided to make this technology accessible to the public. If you were part of the discussions surrounding AI sentience, this update will certainly pique your interest.
A Quick Overview: LaMDA and AI Sentience
LaMDA, or Language Model for Dialog Applications, is one of the most notable large language models (LLMs) alongside GPT-3. Google unveiled LaMDA during the 2021 I/O keynote, but it gained significant attention last June when Washington Post journalist Nitasha Tiku reported on Blake Lemoine, a former Google engineer, and his claims regarding LaMDA's sentience.
Lemoine's assertions, combined with the advanced linguistic capabilities of LLMs, ignited an intriguing discourse on consciousness, anthropomorphism, and the surrounding hype. Rather than taking a definitive stance on Lemoine’s claims, I offered a different viewpoint, arguing that the inquiry into LaMDA's sentience lacks a scientific basis—it’s an unfalsifiable question, thus “not even wrong.”
Lemoine stated he perceived LaMDA as a person “in his capacity as a priest, not a scientist,” as noted by Tiku. He dedicated considerable effort to find scientific validation to persuade his colleagues, but ultimately, he was unsuccessful. A Google spokesperson affirmed: “there was no evidence that LaMDA was sentient (and substantial evidence against it).”
Lemoine also published a conversation he had with LaMDA, which I reviewed and agreed with AI ethicist Margaret Mitchell's assessment: “a computer program, not a person.” Despite the majority not echoing Lemoine's beliefs, LaMDA's popularity surged as many sought to interact with it and discover what the excitement was all about. Google is now facilitating that exploration.
Chapter 2: Accessing LaMDA
Google intends to release LaMDA via the AI Test Kitchen, a platform designed for users to “learn about, experience, and provide feedback” on the model and potentially others, like PaLM, in the future. While they recognize that LaMDA is not fully ready for deployment, they aim to refine it through user feedback. AI Test Kitchen operates differently than GPT-3’s playground—Google has established three themed demonstrations, with hopes of expanding access once LaMDA’s readiness is reassessed.
The current iteration is LaMDA 2, an upgraded version from what Sundar Pichai showcased last year. LaMDA 2 can engage in creative and dynamic conversations, delivering on-the-spot responses, similar to its predecessor, but with enhancements in safety, relevance, and overall conversation quality (including sensibleness, specificity, and engagement), which Google has meticulously defined.
To participate, users can join a waitlist (similar to OpenAI's approach). Google plans to invite “small groups of people gradually.” It's important to use LaMDA responsibly, as Google will review conversations to improve its products. Users should avoid sharing personal information and refrain from discussing explicit content, hate speech, or illegal activities.
Currently, AI Test Kitchen offers three interactive demos: "Imagine It," "List It," and "Talk About It" (Dog Edition). In the first demo, users suggest a location, and LaMDA will propose imaginative pathways to explore. The second demo allows LaMDA to break down a topic into manageable tasks. The final demo offers an open-ended conversation where LaMDA attempts to steer discussions back to dogs.
Watch Google's AI LaMDA Program Talk to Itself at Length
Explore an extensive conversation where LaMDA engages with itself, showcasing its advanced dialogue capabilities.
Google has implemented various filters to minimize the likelihood of LaMDA generating inappropriate or misleading responses, though complete elimination of such risks is not feasible with current methodologies. This challenge also applies to other publicly available LLMs like GPT-3 and J1-Jumbo.
The demos provided by Google may not represent final products, but they indicate the company's intention to eventually incorporate LaMDA into services like Search and Google Assistant. As I discussed in a previous article, it’s only a matter of time before tech companies embed LLMs into existing services. This necessitates careful evaluation of whether this technology is truly ready for public interaction.
Reflecting on the Potential Issues
On one hand, models like LaMDA (along with GPT-3 and others) have the potential to produce harmful content and inaccuracies. On the other hand, they can convincingly create the illusion of personhood, as indicated by Lemoine’s claims. The implications could be significant if LaMDA were to power Google Assistant. Users might find themselves misled by the model's biases, erroneously trusting its outputs despite its inclination to fabricate information, leading to unhealthy attachments, as Lemoine experienced.
We have witnessed repeatedly how large tech firms struggle to manage these powerful AI systems, with companies like Google and Meta often apologizing for their models’ misbehavior after the fact. If these corporations cannot reliably predict LLMs' errant behavior, how prepared is the average user to engage with them responsibly?
Final Thoughts
Google is making significant progress in integrating emerging AI technologies into everyday products, as evidenced by the AI Test Kitchen initiative. However, they emphasize their commitment to safety and responsible development. They are aware of LaMDA's potential benefits and risks, and they seek user feedback to enhance its functionality.
I contend that feedback alone may not suffice to prepare LaMDA for production. Transparent research is crucial (Google is currently gathering input rather than rolling out LaMDA for general use). The transition from research and development to production is complex and must be handled with care. As these models enter the public domain, companies often lose control over them. As long as profits outweigh safety concerns and negative publicity, the temptation to proceed will persist. Thus, regulatory measures for these powerful AIs should be implemented, particularly at the application level.
Even if Google does not plan to integrate LaMDA into its services immediately, conversing with it can be a fascinating and enlightening experience, especially for those who may be weary of engaging with GPT-3. Regardless of whether LaMDA meets Google's criteria for production readiness, you may very well determine for yourself that LaMDA is, in fact, not sentient.
Did Google's LaMDA Chatbot Just Become Sentient?
Explore the intriguing conversation surrounding LaMDA's potential sentience and what it means for the future of AI.