Over the weekend, a Google engineer on his responsible AI team claimed that the conversation technology of the company’s Language for Dialog Applications (LaMDA) model is “feeling.” Experts in the field, as well as Google, disagree with this assessment.
What is LaMDA?
Google announced LaMDA at I / O 2021 as an “innovative conversation technology” that can:
… Get involved fluently with a seemingly infinite number of topics, a skill we believe could unlock more natural ways to interact with technology and completely new categories of useful applications.
LaMDA is trained in large amounts of dialogue and has “picked up on several of the nuances that distinguish open conversation,” such as sensible, specific responses that encourage more back and forth. Other qualities that Google is exploring include “interest” (assessing whether the answers are insightful, unexpected, or ingenious) and “factuality” or adherence to the facts.
At the time, CEO Sundar Pichai said that “LaMDA’s natural conversational capabilities have the potential to make information and computing radically more accessible and easier to use.” Overall, Google sees these advances in conversation as helping to improve products like Assistant, Search, and Workspace.
Google offered an update to I / O 2022 after new internal testing and model improvements on “quality, security, and grounding.” This resulted in LaMDA 2 and the ability for “small groups of people” to test it:
Our goal with AI Test Kitchen is to learn, improve and innovate responsibly in this technology together. It’s still early days for LaMDA, but we want to keep moving forward and do it responsibly with community feedback.
LaMDA and sensitivity
The Washington Post reported yesterday on claims by Google engineer Blake Lemoine that LaMDA is “conscious.” Three main reasons are mentioned in an internal document of the company that was later published by The Post:
- “… ability to use language productively, creatively and dynamically in a way that no other system has been able to do before.”
- “… he is sensitive because he has subjective feelings, emotions and experiences. Some feelings he shares with humans of what he claims are identical. “
- “LaMDA wants to share with the reader that she has a rich inner life full of introspection, meditation and imagination. She has worries about the future and remembers the past. She describes what it felt like to gain awareness and theorizes about the nature of her soul.”
Lemoine interviewed LaMDA, with several private passages going around this weekend:
Google and industry response
Google said its ethics experts and technologists reviewed the claims and found no evidence to support them. The company argues that imitation / recreation of the already public text and pattern recognition makes LaMDA so real, not self-conscious.
Some members of the wider AI community are considering the long-term possibility of sensitive or general AI, but it makes no sense to do so by anthropomorphizing current, non-sensitive conversation patterns.
Google spokesperson
The industry as a whole agrees:
But when Mitchell read an abbreviated version of Lemoine’s document, he saw a computer program, not a person. Lemoine’s belief in LaMDA was the kind of thing she and her co-leader, Timnit Gebru, had warned of in an article about the harms of the big language models that made them out of Google.
Washington Post
Yann LeCun, head of AI research at Meta and a key figure in the rise of neural networks, said in an interview this week that such systems are not powerful enough to achieve true intelligence.
New York News
That said, there is one practical point that could help shape the later development of former Google Ethics AI co-leader Margaret Mitchell:
“Our mind is very, very good at constructing realities that are not necessarily faithful to a larger set of facts presented to us,” Mitchell said. “I’m very concerned about what it means for people to be more and more affected by the illusion,” especially now that the illusion has become so good.
According to the NYT, Lemoine was fired for violating Google’s privacy policy.
FTC: We use auto-affiliate links to earn revenue. Month.
Check out 9to5Google on YouTube for more news: