GPT-4: The Race For The Super Intelligent

GPT-4: The Race For The Super Intelligent
By Olivia Lam


Another One?
Yes, and this is just the beginning. ChatGPT-3 was launched in November 2022, and took the virtual world by storm. From writing university essays, to building meal plans, to figuring out math problems, and being asked just about anything imaginable, GPT-3 changed the way the world looked at artificial intelligence; you can read more about it here

Just as Apple releases a new iPhone every year with an equally creative name, the owners of GPT-3, OpenAI, released their latest and most revolutionary model on March 14, 2023: GPT-4. Much like GPT-3 it can accept text and is improved to provide much more detailed and reliable answers, but now with the update, GPT-4 can also accept and render answers to images. For example, given an image of a fridge full of food, it can come up with possible meal ideas for the given ingredients.


How Does It Differentiate Itself From GPT-3?

The major difference between GPT-3 and GPT-4 is its multimodal build, enabling it to process both text and images, potentially even videos in the near future. Whether it’s photos, diagrams, or screenshots, GPT-4 can analyze it and tell you what is wrong with an image, what is funny about a photo, explain a given diagram or chart, or even create an entire website on its own.

Regarding its accuracy and its capability to handle more complex questions than GPT-3, it was found that given a simulated bar exam, GPT-4 scored in the top tenth percentile, while GPT-3 scored in the bottom tenth percentile. It also scored higher on both the LSAT and SATs, but just like GPT-3, it still cannot pass an AP English or Calculus course. Its ability to handle more complex situations comes from its GPT-4 Reinforcement Learning from Human Feedback (RLHF) post-training, which comes from its predecessor model GPT-3.5. This is when GPT-4 utilizes an algorithm to decipher what humans want by being provided a series of feedback bits, the feedback being a series of desired behaviors that are chosen by humans in a given scenario.

Despite all these innovations to GPT-3, there are still limitations from the prior version that GPT-4 has not yet been able to overcome. Although it was found to be 40% better than its predecessors, it can still make reasoning errors, “hallucinate” facts, contain biases in answers, and can fail at more complex problems. Another major issue found with GPT-3 was its tendency to respond to disallowed content, like providing users with a full DIY recipe on at-home bomb-making. GPT-4, although not perfect, has now disallowed content by 82% and responds 29% more in accordance with OpenAI's policies

Is This a Good Thing

Hard to say. For university students currently in school, this is an excellent tool for university students entering the tech industry, maybe not so much. In a matter of four months, the domain of artificial intelligence has expanded exponentially and it is difficult to see where it will go. OpenAI is already providing GPT-4 to other big service names like Duolingo and Khan Academy, and are looking to scale up to balance out the demand post-release. Mark Zuckerberg is even creating his own artificial intelligence called Meta, very similar to GPT-4. “We’re exploring experiences with text, with images, and with video and multi-modal experiences," said Zuckerberg. All in all, it seems artificial intelligence is the future of tech, and the race to super intelligence has only just begun.

Opinions expressed by the author are not necessarily those of WITI.