One day after Microsoft revealed GPT-4 would be featured at its AI event later this week, OpenAI officially announced the next-generation language model on Tuesday.
According to a product page on OpenAI’s website, GPT-4 is “more creative and collaborative than ever before” and “can generate, edit, and iterate with users on creative and technical writing tasks, such as composing songs, writing screenplays, or learning a user’s writing style.”
GPT-4 leaves ChatGPT in the dust
A key difference between GPT-4 and GPT-3.5 (the language model that powers ChatGPT) is that GPT-4 is capable of accepting image inputs in addition to text. For example, if you show GPT-4 an image of eggs, flour, and milk and ask it what you can make with those ingredients, the language model will “see” the image and list several potential recipes.
GPT-4 is also significantly smarter than previous models. According to OpenAI’s research paper, GPT-4 passed a simulated bar exam with a score similar to the top 10% of test takers. ChatGPT scored in the bottom 10%. The multimodal language model also scored in the 99th percentile of the Biology Olympiad compared to the 31st percentile for ChatGPT.
That’s not to say that the latest model is an evolutionary leap forward. OpenAI CEO Sam Altman warned in the lead-up to the reveal of GPT-4 that “people are begging to be disappointed.” Altman also tweeted today that the model is “still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it.”
As for the risks of AI, which we’ve seen demonstrated repeatedly in recent months in ChatGPT’s factual inaccuracies and Bing’s insanity, GPT-4 still has work to do:
GPT-4 poses similar risks as previous models, such as generating harmful advice, buggy code, or inaccurate information. However, the additional capabilities of GPT-4 lead to new risk surfaces. To understand the extent of these risks, we engaged over 50 experts from domains such as AI alignment risks, cybersecurity, biorisk, trust and safety, and international security to adversarially test the model. Their findings specifically enabled us to test model behavior in high-risk areas which require expertise to evaluate. Feedback and data from these experts fed into our mitigations and improvements for the model; for example, we’ve collected additional data to improve GPT-4’s ability to refuse requests on how to synthesize dangerous chemicals.
If you want to try out GPT-4 for yourself, one way to access it is by paying for ChatGPT Plus. The subscription plan launched on February 10 for $20 a month. Microsoft also confirmed that the new Bing was running on GPT-4 all along, as many suspected. OpenAI has opened an API waitlist for developers who want to implement GPT-4 into their apps.
UPDATE: If you want to see GPT-4 in action, OpenAI is currently hosting a developer live stream on YouTube. You can watch it (or the VOD archive) below: