GPT-4

- I remember my very first game controller. A simple joystick from the Atari 2600 with just one button. Now, at the time I thought it could do everything, but later we got controllers with more buttons, D pads, and fancy new features. Every time the upgrades felt like a big improvement. And that reminds me of the upgrade we just got for the fastest growing technology app of all time, GPT version four. And this update is huge. So what is GPT? It stands for a generative pre-trained transformer and it's a neural network that creates new content, like stories or art, and is pre-trained using massive amounts of data. A transformer means it understands key points from human input and tries to infer your intent. Let's say you ask a question like, how do I build the best house in Minecraft? The transformer identifies keywords like build, best house, and Minecraft, but since you didn't provide details like the version of the game, materials available, or that you happen to be in the snow biome, it will make assumptions and give you an answer that is the most likely to be correct for most situations in the game. So what's the difference between GPT-3 and GPT-4? The new model is better at reasoning as well as providing concise answers, although it's a bit slower than older models. That shows up in how well it performs tasks, but also means that it requires more computational power which can be more expensive. OpenAI, the company behind GPT, has worked to fine tune the model to reduce so-called hallucinations. In other words, produce answers that are nonsensical or untruthful. They have also attempted to reduce biases and improve the overall safety of the results. The model is currently available through open AI's ChatGPT Plus subscription and is also part of Microsoft's Bing as well as Microsoft 365 and Azure. You may have heard that GPT was able to pass the bar exam, which lawyers have to take. While GPT-3 managed to pass this test within the lower 10th percentile, GPT-4's performance was in the top 10th percentile. GPT-4, like GPT-3, was trained on data from before September, 2021. But unlike GPT-3, GPT-4 is capable of using web links as input and can accept much larger amounts of information. It can accept inputs of up to 25,000 words of text and it can provide longer outputs as well. So for example, you can provide the transcript of a long conference talk or a livestream and ask GPT-4 to summarize it and give you the key topics with time codes. Another key feature is steerability. You have better control about the personality, verbosity, or style. You can, for example, ask GPT to put itself in Socratic mode so that it answers questions by asking leading questions, which allows students to slowly uncover the answers themselves. One of the exciting new features is the ability of GPT to understand photos and graphics. So you could, for example, take a photo of the inside of your fridge and ask it to give you a recipe based on what you have. You can pass it a screenshot of a chart or graph and it can accurately interpret the context of the data and do calculations based on that information. This ability is not currently available to the public, but it will be rolled out over time. In the demo from OpenAI, one developer took a photo of a simple sketch and GPT wrote the HTML and JavaScript to build an interactive webpage. GPT-4 is being added into all aspects of life. ChatGPT has already become the fastest growing product in the history of technology with 100 million users in the first two months. And you may have heard of GitHub's co-pilot, which has helped developers become 88% more productive. Companies like Duolingo, Stripe, and Morgan Stanley have already added GPT-4 implementation into their products. Recently, Microsoft announced the Semantic Kernel, an open source SDK that will make it easier for developers to integrate GPT and other AI features into all applications. Its features are being rolled into Microsoft 365, so apps like Excel, Word, and PowerPoint will eventually see productivity gains that have been available to developers in products like Co-Pilot. GPT-4 is like the best controller upgrade for AI language models. And it's going to change the way we interact with our world as developers begin to add these capabilities into their applications. So get ready to experience new features, image inputs, more accuracy, and a whole new level of creativity. Coming soon to an app near you.