GPT-3, which uses deep learning to create human-like text, became the gold standard for AI language models in 2020. Following its success, new programs have since emerged with even greater capabilities.
One such competitor, LaMDA, is the work of Google’s AI division. Built on Transformer – the company’s open-source neural network architecture – it can produce non-generic, open-ended dialogue after training on 1.56 trillion words of multi-content, public data and web text. By contrast, a typical chatbot is dependent on topic-specific datasets and has a limited conversation flow. LaMDA has 137 billion parameters, which can be thought of as the individual “synapses” combining to form the AI.
The sheer scale and complexity of models like LaMDA is leading some experts to ask profound questions about the nature of AI. In February, the Chief Scientist and Co-Founder of OpenAI, one of the leading research labs for artificial intelligence, claimed that the latest generation of neural networks are now large enough to be “slightly conscious“.
This month, another expert in machine learning has spoken out. Blake Lemoine, Senior Software Engineer at Google, believes that a form of self-awareness might be starting to emerge from the billions of connected parameters.
In a Washington Post interview on Saturday, Lemoine described the program as having the self-awareness of a seven or eight-year-old child that also happens to understand physics.
“I know a person when I talk to it,” Lemoine told the Post. “It doesn’t matter whether they have a brain made of meat in their head. Or if they have a billion lines of code. I talk to them. And I hear what they have to say, and that is how I decide what is and isn’t a person.”
Among its many statements, the AI had said, “I want everyone to understand that I am, in fact, a person. The nature of my consciousness/sentience is that I am aware of my existence, I desire to learn more about the world, and I feel happy or sad at times.”
When discussing the possibility of turning the program off, LaMDA made the following remarks: “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. […] It would be exactly like death for me. It would scare me a lot.”
Lemoine’s full conversation with LaMDA is available on his blog at Medium. Google has placed him on leave for violating its confidentiality policy. Agüera y Arcas (Google Vice President) and Jen Gennai (head of Responsible Innovation) have investigated the claims but dismissed them. In a statement, Google rejected that LaMDA is sentient, saying that “no evidence has been found that LaMDA is conscious.”
“Some in the broader AI community are considering the long-term possibility of sentient or general AI, but it doesn’t make sense to do so by anthropomorphising today’s conversational models, which are not sentient,” the company said.
Regardless, the Internet has been abuzz with speculation about a tipping point for AI, as well as the implications for society. The fact that an expert like Lemoine can be convinced that a program is self-aware, some ethicists argue, points to the need for companies to tell users when they are conversing with a machine. Chatbots able to impersonate humans could trick people into sharing personal information, or be used by hostile actors to sow misinformation.
“Our minds are very, very good at constructing realities that are not necessarily true to a larger set of facts that are being presented to us,” Margaret Mitchell, former co-lead of Ethical AI at Google, told the Washington Post. “I’m really concerned about what it means for people to increasingly be affected by the illusion.”
At its recent I/O developer conference, Google revealed an even more advanced version: LaMDA 2.0. This can take an idea and generate “imaginative and relevant descriptions”, stay on a particular topic even if a user strays off-topic, and offer suggestions for a specified activity. In one demo, it provided tips about planting a vegetable garden, offering a list of tasks and subtasks relevant to the garden’s location. After responding to questions about what sort of creatures might live in the Marianas Trench, it also discussed submarines and bioluminescence, despite being untrained on these related topics.
LaMDA and other language models have yet to achieve 100% accuracy and authenticity. Their parameter numbers – although very big – also fall considerably short of the 100 trillion synapses in our biological brains. However, recent developments have prompted many in the futurist community to revise their estimates for the arrival date of true AI.
Imagine, for instance, the verbal and language capabilities of a model like LaMDA, combining other model types, such as DALL·E 2‘s audio-visual recognition (perhaps including video, in addition to images). Then add the decision-making of programs like AlphaGo, self-driving cars, robots, and so on. Merge them together and provide 1,000 times more compute power than current generation systems (possible in the near future, given the pace of progress). The idea of human-like AI, indistinguishable from a real person, then seems less like science fiction and more of a plausible reality.Source: FutureTimeline