GPT4: OpenAI’s Game-Changing Language Model
The company behind ChatGPT has unleashed its latest and most powerful language model: GPT-4. It is a successor of the GPT 3.5 model that powers the publicly available version of ChatGPT. If you have the Plus subscription membership, it is available to try out today. Microsoft, the tech giant, poured $10 billion into OpenAI and is already using the latest model for its search engine. Here are 4 things you need to know.
A lot smarter
GPT-4 has been trained on more than 100 trillion training data, while its predecessor GPT-3 was trained on only 175 billion. The increase in training data has allowed the model to achieve even greater performance and accuracy in natural language processing tasks.
GPT-4 can score higher approximate percentiles among test takers than in the previous model. According to OpenAI, GPT-4 has finished and aced the Uniform Bar Exam (MBE+MEE+MPT) in the top 10% of all scorers with a score of 298/400, compared to the bottom 10% by GPT-3.5.
More Input
GPT-4 has the ability to handle 25,000 input words, which is a significant improvement over GPT-3.5’s capacity of approximately 3,000 input words. The ability to handle more input words allows GPT-4 to understand better and generate more coherent and contextually relevant responses for longer pieces of text.
Imagine this: As a programmer, you take documentation for any library you want to learn, then prompt the AI for a step-by-step guide to achieving a task without reading the entire thing.
Multi-Modal Model
The model can now accept images as input. This means that you can input an image alongside a set of clear instructions, questions, or opinions, and GPT-4 can return a structured answer that uses both sets of data as input.
It can be helpful in various applications, such as generating text outputs based on marketing and sales reports with graphs and figures or summarizing details from textbooks and shop manuals.
Much safer
According to OpenAI, GPT-4 has an 82% decrease in its tendency to respond to requests for disallowed content compared to its predecessor GPT-3.5. Additionally, it is 29% more likely to respond to sensitive requests such as medical advice and self-harm in accordance with the company’s policies.
GPT-4 is trained using reinforcement learning from human feedback and gets an extra safety reward during training to reduce harmful outputs. This means it is less likely to jailbreak the model to generate malicious things that violate the usage guidelines. However, while these measures have helped to mitigate the risks associated with GPT-4, the team acknowledges that they are not foolproof and must be continually evaluated and updated as the technology evolves.
Summary
The world keeps changing before our eyes, for better or worse. This year has been a huge year for AI. We have seen amazing breakthroughs in natural language processing, computer vision, and more. Only a few months ago, people worldwide were surprised by how a language model could be so smart. Now, we are witnessing the next generation of AI that can do even more amazing things.
I hope you enjoyed reading my first article and hope to share more insights and stories about the latest developments and trends in AI with you. If you want to learn more about GPT4 and how it works, you can check out the official paper and blog post from OpenAI.
Until next time!
