ChatGPT burst onto the scene only four months ago, ushering in a tech revolution. Since then, chatbots have wowed with their ability to write essays and jokes, terrified with their potential for human-like darkness, and inspired a million think pieces about how they will alternately improve or threaten humanity. But the technology is moving so quickly that those debates may already be somewhat outdated. On Tuesday, OpenAI, the company behind ChatGPT, debuted its new and improved GPT-4. And though company CEO Sam Altman downplayed its advances, by all accounts it’s a vast improvement on its predecessor.
Naturally, curious techies and journalists (some of whom were given advance access) quickly found noticeable ways in which GPT-4 had leapfrogged its predecessor. Here’s a sample of the new amazing and disquieting feats it can achieve:
It can ace most standardized tests
The old, puny ChatGPT could pass, or come close to passing, the United States Medical Licensing Exam. It could acquit itself with a passing grade on law-school final exams and write a passable law-school essay, though it could not pass the bar. Pathetic. GPT-4 is less mediocre pupil than straight-A student, at least in most areas.
In an extensive blog post, OpenAI reports that GPT-4 passed the bar, going from an estimated tenth-percentile score to an estimated 90th percentile. It also aced a raft of other complicated exams on an array of subjects, ranging from the hard sciences to the liberal arts. It scored a 5 on AP Statistics (from a previous 3), a 5 on AP Macroeconomics (from a previous 2), and even a 92 on an intro sommelier test. Its opinions on Merlot are unknown at this time.
There’s one subject the chatbot can’t quite master yet, though: actual written English. For those worried about a declining interest in the humanities, this might be some rare welcome news.
It can build basic websites
“Generative AI,” which can build words or images based on a user’s prompt, has been all the rage in Silicon Valley for months. GPT-4 introduces “website-building” into the mix. Give the bot a few text prompts, and it can create a rudimentary site:
It’s unclear whether GPT-4’s ability goes beyond crude website-making for now, but its rapidly improving design prowess is probably enough to make web designers nervous, along with, well, almost everyone else.
It can probably sue people on your behalf
Have a grievance with an individual or company but too lazy or fiscally challenged to seek out professional help? Well, you’re in luck:
It can describe images in detail
One of the most dramatic differences between ChatGPT and its successor is the latter’s ability to depict images with text. GPT-4 can not only determine the content in a photo or picture, but analyze the objects within it to make predictions and provide broader context. (This feature is currently available only to select users.)
The Washington Post reports on how it works in practice:
“When shown a photo of a boxing glove hanging over a wooden seesaw with a ball on one side, for instance, a person can ask what will happen if the glove drops, and GPT-4 will respond that it would hit the seesaw and cause the ball to fly up.”
As the paper notes, this would “allow for someone to show a picture of the food in their kitchen and ask for some meal ideas,” among other applications; it could also be a boon to people with impaired vision. The tool does carry some darker implications, like the possibility of mass facial recognition, but let’s not worry about that for now.
It can create Pong in under a minute
That’s just cool.