MC Explains | The difference between GPT-3 and GPT-4

GPT stands for Generative Pre-Trained Transformer, a deep-learning neural network that generates human-like written text.

It does this by parsing a vast amount of data that train it like a human brain, allowing it to learn over time.

GPT-3 allows computers to handle complex communication, where they can handle tasks "like test summarization, machine translation, classification, and code generation."

This applies to conversational bots like ChatGPT as well. GPT-3 allows ChatGPT to respond to queries in a human-like manner.

Interesting. What else is GPT used for? 

The obvious application is content generation. GPT models can generate content based on any query, with human-like use of language.

They are also great at summarizing. GPT models can parse through large volumes of data to create a summary.

Since they are great at answering questions in natural language, they fit right in customer service applications. They can chat with a user to help resolve their problem.

Due to the model being conversational in nature, GPT can be used to power virtual assistants like Google Now or Apple's Siri. It is even powerful enough to create apps and plugin tools for software.

Okay. How is GPT-4 going to improve on this? 

GPT-4 will include several enhancements while being slightly larger than GPT-3. GPT-3 and GPT-3.5, the current iterations of the model, will be replaced by GPT-4 when it takes over in late 2023.

GPT-4 will focus on improving existing parameters rather than increasing in size. The reason is that existing models have complicated set-ups that balloon their size by at least three times compared to GPT-3.

GPT-4 will streamline existing models and improve performance with an efficient system. This will also have a domino effect of reducing computing costs.

OpenAI has said that GPT-4 will optimize and improve existing variables and parameters to make them more efficient. After all, it's not the size of the data that counts but using the correct data according to context.

GPT-4 will lean on accuracy and streamlined performance. In the right hands, GPT-4 will become an invaluable tool that generates text.

Okay, but what about misinformation? 

Excellent question. The current model is susceptible to giving plausible-sounding but incorrect answers.

OpenAI uses a method known as Reinforcement Learning from Human Feedback (RLHF) to train the models.

Whenever an AI takes action, it is classified as desirable or punishable. A desired answer is rewarded, whereas an undesired one is punished.

The problem with this method is that supervised training with humans can sometimes mislead the model "because the ideal answer depends on what the model knows, rather than what the human demonstrator knows."

It is also sensitive to phrasing and is susceptible to generating toxic or biased content, just like any human is.

With GPT-4's enhanced and streamlined data sets, OpenAI hopes these instances will be lesser, but there is no way to completely rule them out.