AI expert Toju Duke explains why image generators like Nano Banana have gotten photorealistic but still fail at basic logic. When asked to show Keanu Reeves cracking five eggs, the model generated six eggs and a broken plate. Current AI can forge passports but can't count.
Nano Banana stole the latest show in the AI race with amazing photorealistic capabilities exceeding all existing image generators. Images are so photorealistic, a far cry from Google's first text-to-image model, Imagen, which was launched way back in 2022, just at the back of DALL-E 2. I remember working with the Imagen team when I was at Google to ensure responsible AI requirements were carried out before launch. Fast forward to 3 years later, The progress that has been made on text-to-image models is surreal. The photos are so realistic that they can be used to forge important documents like passports. Yikes. However, despite this latest state-of-the-art breakthrough, AI models are still behind in pattern-based reasoning. Here's an example. Someone put up prompts to show Keanu Reeves cracking 5 eggs. Nanobanana shows Keanu with 6 eggs and a cracked plate. AI models still still can't count, and Keanu is like, whatever. Remember, you're more intelligent than all AI systems put together. They're quite good at certain functions, but at the moment they're still incapable of simple reasoning, which is still needed in today's world.
Anthropic's Mythos initial release, by Toju Duke, AI Thought Leader, Ex-Google
An Update on Albania's AI Strategy, by Toju Duke, AI Thought Leader, Ex-Google
