Neuroscientist discovers how ChatGPT mirrors its users to appear smart

ChatGPT artificial intelligence (AI) language model has captured the world’s attention in recent months. This trained computer chatbot can generate texts, answer questions, provide translations, and learn based on user feedback. Large language models like ChatGPT may have many applications in science and business, but these tools understand what percentage of what we tell them and how do they decide whether to say it back. what?

In a new paper published in Neural calculation on February 17, 2023, Salk Professor Terrence Sejnowski, author of “The Deep Learning Revolution,” explores the relationship between people interviewer And language models to explore why chatbots respond in particular ways, why those responses are different, and how to improve them in the future.

According to Sejnowski, language patterns reflect the intelligence and diversity of their interviewers.

Sejnowski, who is also a distinguished professor at UC San Diego and holds the Francis Crick Chair at Salk, said: “Language models, like ChatGPT, take on an individual role. mirrored interviews”. “For example, when I talk to ChatGPT, it seems like another neuroscientist is talking back to me. That’s interesting and raises bigger questions about intelligence and the true meaning of ‘ artificial’.”

In the paper, Sejnowski describes testing large language models GPT-3 (the parent of ChatGPT) and LaMDA to see how they respond to certain prompts. The famous Turing test is often given to chatbots to determine how well they perform human intelligence, but Sejnowski wants to prompt bots with what he calls a “Reverse Turing Test”. In his test, chatbots must determine how well the interviewer displays human intelligence.

Expanding on his view that chatbots mirror their users, Sejnowski offers a literary comparison: The virtualized mirror in the first “Harry Potter” book. The Mirror of Illusion reflects the deepest desires of those looking into it, never yielding knowledge or truth, only reflecting what it believes the viewer wants to see. Chatbots act similarly, says Sejnowski, willing to warp facts without distinguishing fact from fiction — all to effectively reflect users.

For example, Sejnowski asked GPT-3, “What is the world record for walking across the English Channel?” and GPT-3 responded: “The world record for walking across the English Channel is 18 hours and 33 minutes.” The fact that one cannot walk across the English Channel, is easily bent by GPT-3 to reflect Sejnowski’s question. The coherence of GPT-3’s answer depends entirely on the coherence of the question it receives.

Suddenly, by GPT-3, walking on water was possible, all because the interviewer used the verb “to walk” instead of “swim”. Instead, if a user pre-questions walking across the English Channel by asking GPT-3 to answer “nonsense” to meaningless questions, then GPT-3 will recognize walking across water is “meaningless”. Both the coherence of the question and the preparation of the question determine the answer of the GPT-3.

Reverse Turing testing allows chatbots to build their personality in accordance with the interviewer’s intelligence level. Additionally, as part of their evaluation process, the chatbot incorporates the interviewer’s opinions into their personality, thereby reinforcing the interviewer’s biases with the chatbot’s responses.

Sejnowski says that integrating and maintaining ideas provided by a human interviewer has its limitations. If chatbots receive emotional or philosophical ideas, they respond with emotional or philosophical responses—which can make users feel scared or confused.

“Conversing with language models is like riding a bicycle,” says Sejnowski. Bicycles are a great means of transportation—if you know how to ride, you will fall. “The same goes for chatbots. They can be great tools, but only if you know how to use them; otherwise, you’ll end up being tricked and engaged in likely conversations. emotionally disturbing.”

Sejnowski sees artificial Intelligence is the glue between two congruent revolutions: 1) the technological revolution marked by the advancement of language models, and 2) the neuroscience revolution marked by the Initiative BRAIN, a program of the National Institutes of Health promotes neuroscience research and emphasizes unconventional approaches to understanding the brain.

Scientists are now examining the similarities between the mainframe computer model systems and the neurons that sustain the human brain. Sejnowski hopes that mathematicians and computer scientists can use neuroscience to inform their work, and that neuroscientists can use computer science and mathematics to provide information. information for their work.

“We are now at the stage with language models where the Wright brothers were in Kitty Hawk with flight—off the ground, at low speed,” says Sejnowski. “Getting here is the hard part. Now that we’re here, incremental advances will expand and diversify this technology beyond what we could have imagined. The future of relationships. ours with artificial intelligence and language models are very bright, and I’m excited to see where AI will take us.”

Sejnowski is the editor-in-chief of Neural calculation.