It’s… alive! (and also a… liar!)

If you were given a photo of a squirrel holding a camera and asked to explain why it's funny, what would you say?

If you said it's because “we don’t expect them to use a camera or act like a human,” congratulations, you’re operating at the same comedic level as GPT-4, the newest version of the large language model underpinning ChatGPT that was released yesterday by OpenAI.

🍎🍊 How’s it different from the previous version?... If GPT-3.5 (the old version) was a Motorola Razr flip-phone, the new GPT-4 would be an iPhone 15. For example, GPT-4 scored in the 90th percentile on a uniform bar exam, compared to the 10th percentile for GPT-3.5.

• GPT-4’s biggest upgrade is the ability to receive both text and image inputs, not just text. During yesterday’s demo, the model was able to take a hand-drawn sketch of a website and code it into, well, an actual website.

⚠️ Human Beware: Oh and according to yesterday’s demo, it can do your taxes. Buuuuut, probably take that with a few grains of salt. Although the new model is 40% more likely to provide accurate information, OpenAI specifically warns users that GPT-4 still suffers from the same limitations as GPT-3.5, including social biases, aggressive responses, and presenting false information as true (also known as “hallucinating”).

GPT-4 is now available to ChatGPT Plus users and developers, as well as users with early access to Microsoft’s Bing.