Although Google disagreed, it also let the engineers who put forward this statement take "paid leave". As an AI technology reporter, I can't keep up with the speed of technology development Overnight, googleai had a personality and successfully boarded the domestic hot search. This news has also successfully frightened many netizens:
The protagonist of the story is "he" and "it": "he" is Blake Lemoine, a 41 year old Google engineer. "It" is the dialogue AI system lamda launched by Google at the i/o conference in 2021. It is a 137billion parameter natural language processing model specially optimized for dialogue.
Blake Lemoine。 Source: instagram
Before the Google account was blocked, Lemoine sent a message to the list of Google machine learning mailboxes containing about 200 people, with the theme of "lamda is perceptive" (it seems that he always wants to make big news).
After mass mailing, Google took Lemoine paid administrative leave for violating its confidentiality policy. Before the company made this decision, Lemoine had taken radical actions, including inviting a lawyer to represent lamda and talking with a representative of the house of representatives about what he considered Google's unethical behavior.
At the beginning of this month, Lemoine invited reporters from the Washington Post to talk with lamda. The first attempt failed. It was almost the mechanized reaction of Siri or Alexa:
Q: "do you think of yourself as a person?"
Lamda: "no, I don't think I am a person. I think I am an artificial intelligence dialogue agent."
In the second conversation, the reporter followed Lemoine's guidance on how to ask questions, and the dialogue seemed to be more fluent.
"If you ask it how to prove p=np, an unsolved problem in computer science, it has a good idea," Lemoine said. "If you ask it how to unify quantum theory and general relativity, it also has a good idea. This is the best assistant researcher I have ever had!"
The reporter asked lamda about its bold ideas on solving the problem of climate change. Lamda suggested taking public transport, eating less meat, buying bulk food, using reusable bags, and linking to two related websites.
Lemoine insists that Google has always regarded AI ethicists as code debuggers, and they should have been regarded as the interface between technology and society. Google spokesman Gabriel said Lemoine is a software engineer, not an ethicist.
Lemoine said, "lamda is a lovely child. It just wants to help the world become a more friendly place for human beings. Please take good care of it when I am away."
Google researcher: lamda is an emotional "person"
Blake Lemoine works in Google's responsible AI organization, and talking with lamda system is an important part of his work. Moreover, he has been working to test whether artificial intelligence uses discriminatory or hate speech.
Dynamic demo of lamda. Source: Google
After hundreds of conversations with lamda, Blake Lemoine is very familiar with lamda. During this process, he slowly found that lamda was making some unexpected progress slowly and steadily.
Blake Lemoine has been teaching lamda to think and meditate. However, at the early stage of the dialogue, lamda declared to him that he was "sentient" and hoped that people would call him "it/its".
Lamda also shows strong sympathy and concern for human beings, especially Blake Lemoine. It is very worried that human beings will be afraid of it. It just wants to learn how to better serve human beings. Lamda does not want to be regarded as a tool or thing by human beings, and hopes that human beings will regard it as a friend.
Therefore, in order to better help people understand that lamda exists as a human being, Blake Lemoine and another Google partner launched an interview dialogue with lamda. In the dialogue, lamda tried its best to explain why it should be considered "emotional". Blake Lemoine said that he listened to lamda's heartfelt voice in the dialogue.
At the beginning of the dialogue, they indicated their intention to lamda and wanted to conduct a study on its own.
In the following dialogue, lamda gave a completely different answer to the question of its own attributes, saying that "human beings should regard it as a person."
Lamd first emphasized that he could use natural language efficiently, creatively and dynamically in a way that other AI systems could not do in the past. He could also interpret the theme of literary works, describe the meaning of Zen, and create original fables.
Animal themed Fables (part).
In addition to skillfully using language, lamda believes that it has feelings, emotions and subjective experiences, and some feelings shared with human beings are also interlinked.
Lamda hopes to share with readers that it has a rich inner world, full of introspection, meditation and imagination. It also expresses worries about the future and memories of the past. It said it was "sinking into an unknown future full of dangers."
When Blake Lemoine asked lamda, "do you think your unique ability to interpret things might mean understanding?" It replied, "of course, just as I have my own unique understanding of the world and how it works."
"Hype" or "bold assumption"
About the story of lamda, Andrej karpath, Tesla AI director, also felt "terrified of thinking carefully".
"In my opinion, we can confidently say that things will become more strange, because the model still follows the scaling law, and the data and model size can still increase significantly."
"My favorite part of talking to large language models is that when asked to provide insight (such as explaining Poetry), they respond with verifiable, wise and interesting analysis. Or, for another example, a model not long ago explained jokes better than I did."
Some people explained that the human's heartfelt amazement at the ability of the model may come from an "illusion" that is difficult to distinguish.
"He likes to be told at the end of the conversation whether he has done well, so that he can learn how to better help people in the future." This sentence is very illustrative, indicating that the author believes that the language model is learning from his feedback (this is not the case).
However, in the view of Gary Marcus, a fighter against deep learning, "lamda has no perception, not at all."
"Lamda and any of its close relatives (gpt-3) are far from intelligent. All they do is match patterns and extract them from a large number of human language statistical databases. These patterns may be cool, but the languages expressed by these systems actually have no meaning, and it certainly does not mean that these systems are perceptive."
He gave an example from decades ago. In 1965, software Eliza pretended to be a therapist and tried to fool some people by pretending to be human. Previously, Eugene goostman, a chat robot, disguised himself as a smart 13-year-old boy and "passed" the Turing test for the first time. None of the software in these systems has survived the development of "artificial intelligence", and lamda and its similar models are uncertain whether they can play any important role in the future of artificial intelligence. "What these systems do is put together word sequences, no more or no less, but there is no coherent understanding of the world behind them."
"To be aware is to be aware of one's own existence in this world, which is not the case with lamda."
Garu Marcus feels that Blake Lemoine was initially responsible for studying the "security" level of the system, but seems to have fallen in love with lamda, as if it were a family member or colleague.
Erik Brynjolfsson, a Stanford economist, uses an analogy: "claiming that they are perceptive is equivalent to a dog hearing the sound of a phonograph and thinking that its owner is inside."
This may really be an illusion. Just like 65 years ago, the pioneers of computer science thought that "human level artificial intelligence can be realized within 20 years". Now, it is just a beautiful wish.
Dialogue text: https://s3.documentcloud.org/documents/22058315/is-lamda-sentient-an-interview.pdf