OpenAI’s GPT-4 is a step forward in what AI can do (and perhaps towards our doom)

Just months after setting ChatGPT loose into the world, artificial intelligence firm OpenAI has upped the ante with GPT-4. Earlier this week, the company modestly introduced the new language model, declaring it ready to take text and image inputs and spit out responses, albeit ones “less capable than humans in many real-world scenarios”. 

OpenAI is under-egging GPT-4’s capabilities a bit — which says something about how confident it is in its abilities. GPT-4 can ace standardised tests like the BAR, LSAT and a plethora of college course exams. It can process images, recognise them and provide context. It reportedly makes fewer mistakes and “hallucinations” and is better at refusing attempts to get it to do things it’s not programmed to.

GPT-4’s capabilities are both gobsmacking and perhaps a little bit less impressive than they first seem. That a piece of technology accessible to anyone can instantaneously spit out a new, university-level essay or code for a website based on a drawn sketch shouldn’t be understated. The model’s ability to do things we take for granted or view as an amusement (like getting GPT-4 to explain why a picture of chicken nuggets arranged in the shape of the world’s map is funny) will be life-changing for some — for example, those with visual impairments.

An antidote to some of the hype and resulting fearmongering comes from remembering why GPT-4 is able to do some very impressive things. OpenAI’s GPT-4 is built off a deep-learning large language model: an enormous database of language that it analysed for patterns and now uses to predict what answer you want based on your question.

Compared to humans, GPT-4 and other AI models can do things well because they work in a fundamentally different way from our brains. We’re not constantly blown away by our phones remembering a lot of phone numbers or instantaneously doing enormous sums, and we shouldn’t be impressed GPT-4 can pass the bar exam — the model has an enormous amount of bar exam answers available to it that it can crib off when replying. (Probably, we don’t know for sure because, for the first time, OpenAI isn’t telling us what its model is trained on.) Its abilities are still impressive, but not as impressive as humans, with all their flaws and limitations, acing the same exam. So don’t worry about AI replacing lawyers anytime soon, despite the best efforts of some.

An interesting, under-covered part of GPT-4’s announcement was OpenAI’s admission that it’s now using the model to help with evaluation and iteration of the model’s own development. This sounds a lot like the singularity hypothesis, the theory that tech will one day get so advanced that it will start improving itself faster than humans can and become a runaway train of technological advancements. In case it wasn’t clear, this scenario isn’t a great outcome for us meatsacks. 

In reality, OpenAI’s use of GPT-4 to help improve itself isn’t a doomsday scenario, but it is an example of how those best acquainted with the technology see its value. GPT-4 can’t replace people entirely. Its errors are too common and, while easily spotted by any individual with basic expertise, usually opaque to itself. Instead, it’s becoming an impressive assistant for experts who can coax it to supercharge their abilities. The runaway train of technological advancement still needs a human driver — at least for now.