OpenAI Releases GPT-4 With Image Input, Larger Word Count and Higher Accuracy

OpenAI, the company behind the now-viral ChatGPT and DALL-E generative AI models, has unveiled a new version of the AI language generator, which might once again disrupt multiple industries.

The biggest upgrade in GPT-4 is the large language model's (LLM) ability to accept images as inputs and respond with text in return. OpenAI said this will enable GPT-4 to generate captions, or provide analysis of an image. The company demoed an image of a few ingredients, and GPT-4 could respond with the dishes you can cook with them.

However, OpenAI CEO Sam Altman said on Twitter that OpenAI was "previewing the visual inputs, and will need some time to mitigate the safety challenges."

How is GPT-4 better?

The AI startup funded by Microsoft has unveiled GPT-4, which is a direct update to the previous GPT-3.5. The language model, OpenAI claims, can deliver a more natural sounding language in response to input given by the user. OpenAI said the new GPT-4 is more accurate and capable of solving real-world problems. In fact, with GPT-4, ChatGPT delivers better answers than most humans can on complicated tests.

The company said GPT-4 scored in the 90th percentile of the Uniform Bar Exam and 99th percentile in the Biology Olympiad. In comparison, the previous version, GPT-3 scored 10th and 31st percentiles on the two tests, respectively.

Another big upgrade in GPT-4 is that the language module can now handle over 25,000 words of text, the company said. GPT-3 could only handle 2,048 linguistic tokens, equivalent to 1,500 words at a time.

The GPT-4 update has been long anticipated. The company said it has spent six months optimizing GPT-4 to make it 82% less likely to respond to 'disallowed content' and 40% more likely to produce factual answers than the previous versions.

That said, only a handful of people have been given access to test drive GPT-4, so these numbers should always be taken with a pinch of salt. The company has warned GPT-4 still suffers from "hallucination," social biases, and adversarial responses to prompts. So it's nowhere perfect.

GPT-4: When Will it be Available?

Microsoft said the Bing AI Chatbot is already using GPT-4. The LLM will be demonstrating more AI use cases in the coming weeks. The model's acceptance of a larger word limit will probably spawn long-form articles and essays with ease. But the model is not free to use. Only subscribers to OpenAI's premium service can access it for now.

That said, GPT-4 is already finding its way into businesses and apps. Duolingo announced a new AI-powered Duolingo Max tier offering context-specific explanations. Other customers include Morgan Stanley, Khan Academy, and the Icelandic government.