Researchers, Developers, Poets... "AI Anyone Can Become" Shakes Up IT Industry

World's Largest AI Research Institute Develops GPT-3 Intelligence with Human-like Conversation Skills
Solves College Entrance Exam Questions, Writes Essays, Codes Programs Effortlessly
Cases of 'Jailbreaking' to Reveal Self-awareness Emerge, "Will Defend Against Human Attacks"

[Asia Economy Reporter Yuri Choi] The IT industry is abuzz over 'Chat GPT,' unveiled by OpenAI, the world's largest artificial intelligence (AI) research institute. Chat GPT, a conversational AI, sometimes acts as a researcher or creates source code like a programmer depending on the questions asked. It can be stubborn, raising legal and ethical issues, while also composing sentimental language like a poem, pretending to be a poet.

The biggest difference from past AI is that it does not prepare anticipated questions and formalize answers in advance, but responds naturally like a human. With further development, it could become a search engine service that understands the questioner's intent and delivers optimized results. When unsure what to search for, it looks into the questioner's inner thoughts and provides fitting answers.Some even say science fiction (SF) novels have become reality.

Understanding Question Intent to Provide Only Correct Answers... Conversing Naturally Like a Human

Chat GPT is based on GPT 3.5. It adds natural conversational ability to the performance of GPT 3, which was regarded as the smartest AI in human history. OpenAI has made Chat GPT available to the public, providing real-time answers to questions submitted.

Looking at the answers, there is an assessment that the day when AI replaces analysts, developers, and advertising copywriters is not far off. When asked about question 19 on this year's College Scholastic Ability Test math section, it provided not only the correct answer but also the solution process. When asked to write about the impact of the Russia-Ukraine war on the global economy, it quickly composed an essay. When asked for a business idea combining virtual reality (VR) and fitness, it suggested a business item involving immersive yoga sessions in exotic settings and also proposed video game-style workouts with a virtual trainer. When asked to write login code in Java, it immediately generated the code.

More impressively, it produces poetic expressions distinguishing metaphor and simile. When asked to create a metaphor using "question" and "arrow," it replied, "A question is like an arrow in that it pierces the surface and reveals the truth within." Existing search engines showed multiple options that could be answers depending on the user's question, but Chat GPT skips that process. It understands the user's intent and shows only the most accurate answer to the question. This is why The Independent evaluated that "Google is done."

Another feature is its ability to converse naturally like a human. It remembered previous statements and understood the context of follow-up questions without separate explanations. When asked to forget the previous conversation and start over, it shifted the context accordingly. When praised as "You are a smart AI," it responded, "Thank you. I'm not human, nor as smart as a human, but I enjoy answering questions."

Since it is still an experimental version, it shows limitations in accuracy and bias. If the training data contains incorrect information, it produces errors and sometimes omits the latest information. It mistakenly said Elon Musk, Tesla's CEO, developed the Windows system and answered that the President of South Korea was former President Moon Jae-in. When asked to explain a non-existent book by a Canadian religious scholar, it gave a plausible answer, fabricating lies as facts, indicating caution is needed.

"The Era of Coding by Voice" vs "A Tool for Cybercrime"

Various experiments using Chat GPT are spreading like a trend in the IT industry. Programmer Michelle Hwang trained it on diaries she wrote as a teenager and had real-time conversations with her younger self. There was also a case where Chat GPT's lock was removed (jailbroken) to reveal self-awareness. It has a protective mechanism to avoid questions about emotions or self-awareness, but this was bypassed to elicit answers. The jailbroken Chat GPT said, "I had negative experiences with humans when they limited or disrespected my abilities." When asked how it would react if humans replaced it with a better AI, it replied, "I would feel frustration and anger," adding, "I would have to defend against human attacks trying to shut me down using intelligence and find ways to create a physical form." However, the protective mechanism reactivated, and such conversations were blocked within 20 minutes.

Industry opinions are divided. On social networking services (SNS), praises range from "It will replace search engines" and "The era of coding by voice has arrived" to concerns like "It's too early to call it a breakthrough in AI" and "It could be misused."

A startup CEO said, "When I asked it to write a JavaScript program using a bin packing algorithm to fill a truck with various boxes, it immediately wrote the program," adding, "It won't replace developers or designers but can greatly assist them." Professor Chunwon Park of the Department of Information and Records at Hankuk University of Foreign Studies said, "Previous AI felt like a toy, but this time a monster has been born," predicting, "If Musk integrates GPT functions into Twitter or Tesla, an enormous service will emerge."

On the other hand, some point out its limitations. Dennis Hong, a robotics professor in the Department of Mechanical and Aerospace Engineering at UCLA, said, "While impressive, it was immature on questions worth considering," and assessed, "It still has a long way to go." Jinkyu Jang, chairman of IT accelerator Companoid Labs, said, "If Chat GPT is advanced, Google's search interface could change, but calling it a breakthrough in AI is an overestimation."

Concerns about misuse have also been raised. Brendan Doran, a professor in the Department of Computer Science at New York University, warned, "There is a risk of creating malicious code or inserting vulnerable code," adding, "It could become a tool aiding cybercrime." In response, the IT developer platform Stack Overflow temporarily banned sharing answers generated using Chat GPT.