The New Launch of GPT-4

Source: The Guardian

By: Eileen Wang

From helping people find a dairy-free mac and cheese recipe to writing quality college-level essays, ChatGPT, a natural language processing tool that has been trained using large amounts of data to analyze language patterns and generate “human-like” responses, has consistently made the headlines. Ever since its launch in November 2022 by the company OpenAI, ChatGPT has only grown in popularity, amassing a total of 100 million active users in January 2023. It has surpassed TikTok, Instagram, Google+, etc. to become the fastest app ever to reach that impressive milestone! 

While ChatGPT, which is based on OpenAI’s GPT-3.5 model, is already considered a revolutionary tool by many, the company has managed to elevate it to new heights, creating an even more powerful version. On March 14th, 2023, OpenAI unveiled its latest development, the GPT-4 model, exciting users, the science community, and people in the technology industry alike. 

Improving on various aspects of the GPT-3.5 model and using more data for training, GPT-4 supports a chatbot that is more factual and creative. OpenAI also shared that its newest model can read and generate 25,000 words of text and write code in all of the main programming languages. 

Perhaps one of its biggest breakthroughs is its new ability to analyze and respond to input images, instead of just being able to read and respond with text like the previous model. GPT-4 not only can describe images, but it can also explain the meaning and context behind them. Additionally, this model’s new multi-modal input feature allows users to scan in reports with graphs and figures, text book pages, and even screenshots. 

GPT-4 is also capable of performing well on standardized tests, including the LSAT, GRE, SAT, and various AP tests. Scoring around the top 10% of all test takers, this is already a huge improvement from the test-taking abilities of GPT-3.5; for reference, GPT-3.5 only managed to score around the bottom 10% of test takers. GPT-4 scored a 710 out of 800 on the SAT reading portion and a 700 out of 800 on the math section, compared to GPT-3.5’s scores of 670 and 590 respectively.

Although this new model seems even more impressive than its previous version, the CEO of OpenAI, Sam Altman, warns that GPT-4 “is still flawed, still limited.” It is still less capable than humans in many scenarios and continues to suffer from the problem of hallucination (making up incorrect information); however, on the bright side, noticeable improvements are being made to solve these issues. Compared to GPT-3.5, GPT-4 hallucinates 40% less of the time and is 82% less likely to respond to requests for inappropriate or disallowed content. While these issues are still not 100% fixed, OpenAI still strongly urges users to take great care when using these AI language tools.

Currently, GPT-4 is available via a waitlist and has already become part of third-party products, including Microsoft’s AI-powered Bing.