Has the robot Armageddon commenced? ChatGPT took only five days to gain its first million users after its launch on 30 November 2022, and just two months to get 100 million users. This is the fastest uptake of any technology that the world has ever experienced, said Professor Benjamin Rosman. Give ChatGPT written instructions, and it delivers, at lightning speed, everything from silly poems to speeches and mini-biographies. It can translate accurately, write tweets, and solve arithmetical problems.
Rosman was speaking at Taste of Limmud in Rosebank, an event designed as an appetiser to the main Limmud gathering to be held in August 2023.
ChatGPT defines itself as “a large language model developed by OpenAI”. When I asked it what it was, it wrote, “I have been trained on a massive amount of text data and can generate human-like responses to a wide range of prompts and questions. My purpose is to assist users in generating human-like text and help them find information on various topics.” The GPT bit stands for “generative pretrained transformer”.
“It was trained on 570 gigabytes of text,” Rosman said. “That’s a lot. The big idea is this: it has learned which words to pay attention to and that the context of each word matters.” He demonstrated how it can tell the difference between a dog’s bark and a tree’s bark from the context, for instance.
Rosman also spoke of the limitations of ChatGPT. Its answers aren’t guaranteed to be factual. “This is called ‘hallucinating’ – generating plausible sounding nonsense. It can also show cultural biases and messes up some maths problems. It sometimes struggles with causality and logical reasoning.”
He also said it was difficult to keep users safe. Troubled teens might seek ways to self-harm, for example. It will also have an effect on jobs, and may increase fake news and phishing attacks.
A new version, GPT-4, was released on 14 March 2023. Compared to the earlier version, which could write a United States-based law exam and rank in the bottom 10%, the new version aces the exam in the top 10%. It can also interpret handwritten text on the back of a serviette.
Rosman said ChatGPT is most useful as a writing assistant, a little like the annoying, animated paperclip that used to pop up on earlier versions of Microsoft Word. Given the right prompts, it can write reference letters, code and find bugs in code, and generate poetry. It can simplify concepts, or make them more complex. It can criticise arguments and find flaws in them. It can even analyse data in graphs. It could be helpful in teaching, researching, and writing, in a myriad of professions.
“Artificial intelligence will keep challenging what it means to be intelligent, or conscious,” Rosman said. “ChatGPT offers countless opportunities to improve our lives. It’s not to be feared. There have been ways for students to pay other people to do their assignments over the internet for a while. We have to reassess how we learn and teach – a take-home assignment has little value now.”
Curious, I took ChatGPT for a test-drive. I asked it to write a sonnet about bananas, which was very good – it rhymed and scanned just like Shakespeare. I also asked it to translate Hamlet’s “To be or not to be” soliloquy into Afrikaans and French, which it did flawlessly. But it refused to write a limerick about Jews, saying that it wouldn’t promote harmful stereotypes and prejudice. Maybe the machines aren’t taking over the world quite yet.
