The GPT-4 AI model from OpenAI has been released and is already being used in a variety of applications ranging from virtual volunteers to enhanced language learning bots. This new version differs significantly from previous versions, such as ChatGPT and GPT-3.5.
For starters, GPT-4 is multimodal, which means it can interpret both images and text. This enables it to recognize patterns on clothing and plants, as well as read maps, making it useful for visually impaired people. GPT-4 can be given images to process in order to find relevant information. Users can ask it to describe what’s in a picture, and it will. To top it all off, its comprehension extends beyond that.
Second, because of its training on malicious prompts, it is much more difficult to deceive. It would not be bothered by coaxing or blackmail. It has been trained on a large number of malicious prompts that users kindly provided to OpenAI over the last year or so. With these considerations in mind, the new model outperforms its predecessors in terms of “factuality, steerability, and refusal to go outside of guardrails.”
Third, it has a longer memory, which allows it to remember conversations and written stories better. GPT-4 has a maximum token count of 32,768 — that’s 215 tokens, which translates to approximately 64,000 words or 50 pages of text, enough for an entire play or short story. This allows it to remember conversations and follow up more effectively.
Fourth, it is more multilingual, with the ability to accurately answer multiple-choice questions in 26 languages.
Finally, thanks to a new training architecture, it has more stable and diverse personalities or behaviors. It means that users will be able to change the “classic ChatGPT personality with a fixed verbosity, tone, and style” to something more appropriate for their needs in context. While only tested on a limited set of criteria, these new features offer exciting possibilities for the future of AI.
The sources for this piece include an article in TechCrunch.