3 Most Impressive Things GPT-4 Brings to the Table

The competitive Artificial Intelligence tech race continues to blaze forward, and all sorts of firms are fighting for notoriety—including the leaders of the pack. After Open AI’s viral popularity with their AI chatbot ChatGPT, the industry leader captures the attention of the internet once again with their newest iteration: GPT-4. In large part thanks to ChatGPT’s popularity, chatbots are anything but rare at this point, which might beg the question: why do we need more? Anybody who played around with ChatGPT was likely impressed at the rapid and fluent responses to seemingly limitless questions and topics. Anybody who really dug into the chat, however, might have noticed another thing: the responses weren’t always the best—or even correct. In a very real sense, previous iterations of ChatGPT have been a promise and a glimpse at what a powerful AI assistant can be. GPT-4 takes another step closer. Much closer. Here’s a look at the 3 most impressive things GPT-4 brings to the table.

Better Accuracy

Previous versions of ChatGPT were meant to be more proof of concept and fun and interesting AI tools to play around with than a full-fledged and trustworthy AI assistant. One of the biggest indicators of this was ChatGPT’s often questionable accuracy. Although responses are always well articulated and seemingly knowledgeable, users may have noticed some weird and sometimes wrong statements coming from the chat.

According to OpenAI, improvements on the deep learning model not only make GPT-4 “40% more likely to produce factual responses,” but they also enabled GPT-4 to place among top percentiles in various different academic evaluations. This includes testing in the top 10% of bar exam test takers. Meaning that GPT-4 rivals human reasoning and analytical thinking in a very serious way. Mistakes are still possible and inevitable, but where ChatGPT only offered a glimpse at an AI assistant with rigorous answers, GPT-4 is practically the real deal.

Image Processing

One simple but huge change that comes with GPT-4 is the introduction of image processing. Whereas before, the AI’s only bridge into your world was the keyboard interface, image processing allows GPT-4 to literally see into your world. It’s one thing to ask a friend for help with a recipe by sending them what you have in your pantry, and it’s another thing to have that friend standing next to you, also looking into your pantry.

For a more high-stakes example, Be My Eyes has implemented GPT-4’s image processing ability to supercharge their visual-aid assistant. Team members at Be My Eyes say that, after the inclusion of GPT-4, the Be My Eyes app began offering support at the same level as a human volunteering for those who are visually impaired. The ability to re-prompt and “steer” the AI also helps with both accessibility and making sure the information generated by the app is as useful as possible.

Steerability

Speaking of “steering” the Be My Eyes app, another big addition to the latest GPT iteration is the ability to steer the AI. Before, ChatGPT would usually provide responses in the same dry, straightforward format and tone. There’s nothing wrong with this necessarily, but it certainly felt more like talking to an encyclopedia than to a quippy Iron-Man-esque AI assistant. On top of more accurate responses, GPT-4 is also “steerable,” which means that it will mold its responses to whatever parameters you give it.

One thing is tone. If a goal of AI development is getting the assistance provided to feel like real human interactions, GPT-4’s ability to mimic different tones and personalities is a big step forward. An OpenAI demo showed how, not only can the chat mimic a specific style of speech (or prose), but it can follow very specific, tedious requirements, like making sure every word in the response starts with the same letter.

Limitations

Of course, GPT-4 is not perfect but, as OpenAI President Greg Brockman said during a demo, “neither are you.” GPT-4 is able to generate highly accurate information about virtually any unproblematic topic much quicker than any one person is able to, but mistakes are still possible. Importantly, though, the smartest person you know makes mistakes, too.

One huge limitation GPT-4 has, along with most AI chatbots, is that it was trained on data available up to September 2021, meaning it won’t know much about anything that has happened more recently than that. While this is still an unimaginably large amount of information with boundless use-cases, it does mean that, at face value, GPT-4 won’t be able to give you very useful insight on what stocks to buy, for example.

Another limitation that happens with any algorithm trying to mimic human language is how mistakes are made, and how they’re interpreted. While it attempts to mimic human interactions, this imitation is part of the AI’s training, which means mistakes happen differently than they do when you or I make mistakes. Potentially, there might be an issue with how the chat understands the words being used, rather than there being an issue with the information itself. Understanding the nuance of this disconnect is careful and difficult work being tackled by researchers. This is far above the pay grade of us common users, but it should be an important caveat to consider when taking AI-generated information with a grain of salt.

All things considered, GPT-4 is the best of the best when it comes to AI interactions, a fact that is exemplified by OpenAI’s open collaboration with several organizations to tackle real-world problems and situations. Until now, we have been playing with the promise of an AI-guided future. Thanks to OpenAI, playtime is officially over.