Since its launch in November last year, ChatGPT has become an extraordinary success. Essentially an enhanced chatbot, the AI program can generate answers to life’s biggest and smallest questions, and compose college essays, fictional stories, haikus, and even job application letters. It does so based on what it has compiled from a staggering amount of text on the Internet, with careful guidance from human experts. Ask ChatGPT a question, as millions have in recent weeks, and it will do its best to answer, unless it knows it can’t. The answers are confident and fluently written, even if they are sometimes spectacularly wrong.
The program is the latest to emerge from OpenAI, a research lab in California, and is based on an earlier AI from the team, called GPT-3. Known in the field as a large language model or LLM, the AI is fed hundreds of billions of words in the form of books, conversations and web articles, from which it builds a model, based on statistical probability, of the words and sentences that tend to follow the preceding text. It’s a bit like predictive text on a mobile phone, but scaled up massively, allowing you to produce full responses instead of single words.
The major step forward with ChatGPT lies in the additional training you received. The initial language model was fine-tuned by feeding it a large number of questions and answers provided by human AI trainers. They were then incorporated into their data set. The program was then asked to produce several different answers to a wide variety of questions, which human experts then ranked from best to worst. This human-guided fine-tuning means that ChatGPT is often very impressive at determining what information a question is actually looking for, gathering the right information, and framing an answer in a natural way.
The result, according to Elon Musk, is “scary,” as attested by many early adopters, including college students who see it as a savior for backlogs. It’s also harder to break than previous chatbots. Unlike older chatbots, ChatGPT has been designed to reject inappropriate questions and avoid making things up by generating answers on problems you haven’t been trained on. For example, ChatGPT doesn’t know anything about the post-2021 world, since its data hasn’t been updated since then. It also has other, more fundamental limitations. ChatGPT doesn’t handle truth, so even when answers are smooth and plausible, there’s no guarantee they’re correct.
Professor Michael Wooldridge, director of AI fundamental research at the Alan Turing Institute in London, says: “If I text my wife that starts: ‘I’m going to be…’, I might suggest the following words’ in the pub’ or ‘afternoon’, because it went through all the messages I’ve sent to my wife and learned that these are the most likely ways I’ll complete that sentence.ChatGPT does the exact same thing on a large scale.
“These are the first systems that I’m really excited about. It would take 1,000 human lives to read the amount of text the system was trained on and hid in all that text which is a huge amount of knowledge about the world.”
As OpenAI points out: “ChatGPT will sometimes write plausible-sounding responses that are incorrect or nonsensical” and “will sometimes respond to harmful instructions or exhibit biased behavior.” It can also give long answers, a problem its developers attribute to trainers who “prefer long answers that seem more complete.”
“One of the biggest problems with ChatGPT is that it comes back, with a lot of confidence, with falsehoods,” says Wooldridge. “He doesn’t know what is true or false. He doesn’t know about the world. You absolutely should not trust him. You have to check what it says.
“We are nowhere near the Hollywood dream of AI. You can’t tie a pair of shoelaces or ride a bike. If you ask him for a recipe for an omelet, he’ll probably do a good job, but that doesn’t mean he knows what an omelet is.” It’s very much a work in progress, but transformative nonetheless.
