ChatGPT

ChatpGPT

Open AI has recently launched one of the most advanced Artificial Intelligence Chatbots. ChatGPT was released on Nov 30 of 2022, and in approximately 5 days it easily gained over 1 million users. This advanced chatbot is free of charge and open for any of your questions, with almost a guaranteed answer following a human-like response.

How does it work

The chatbot was made with the GPT-3 language model and trained with massive amounts of data to answer a plethora of formats and styles for the diverse range of questions that users may ask. The AI improved with Reinforcement Learning from Human Feedback (RLHF). Multiple trainers would prompt the AI questions but also assist the bot into providing approved responses. With the different outputs that were received, trainers were also tasked with the job of ranking each one based on quality and guidelines. This was to create the reward environment to implement the Proximal Policy Optimization (PPO) reinforcement learning algorithm. In other words the ChatGPT was trained to know which responses were appropriate, and which responses were not, including scenarios where it cannot answer the prompt.

Limitations

Though it may seem as if it has all the answers to the high amount of questions being asked, it is still limited. If worded a certain way, even by a very small change from another prompt it may not be able to answer one but can cooperate with the other. This is something OpenAI is working on, along with feeding it more information to answer with. The information it currently has is the other limitation that holds it back from perfect cooperation. It has been given a massive amount of data, but only from places such as archived books or wikipedia. There does seem to be an assumption from many that it pulls answers from the web, however that is not the case and only pulls from its own resources

History

As stated earlier ChatGPT was released Nov 30, of 2022, however it was based on other AI models that were also developed by OpenAI. ChatpGPT is built off the GPT-3.5 series, similar yet different to the GPT-3 model. The GPT-3 is actually faster than the ChatGPT, yet not as useful for setting up as a chatbot on a website.

What the chat thinks of himself# Maybe you’re interested in what it thinks of itself. This is the response it gave me to the question, “What is ChatGPT?”

GPT-4

Though this new and revolutionary GPT project is extremely impressive, we can still look into the next version, GPT-4. GPT-3 was made with approximately 175 Billion Parameters (Learning rate, layers for the AI, etc.), in other words it’s extremely complex and with a vast amount of parameters comes better processing, and outputs. Though that many parameters are extremely impressive, GPT-4 which is expected to release in 2023 blows it’s antecessors out of the water with a staggering 100 Trillion Parameters. This thrilling project will have people all over the world making use of this masterpiece within their browser. Personally I cannot wait for the next big version of ChatGPT, especially if it will be implemented with this futuristic GPT version. The potential for this project is extremely large and very exciting. Soon enough the GPT-3 will look like nothing and the time for confusing robots and humans finally will begin.