5 ms·
Ned Block: Why ChatGPT Struggles with Watch Faces [video]
- stared 2y agoI couldn't believe it, so I tried myself with GPT-4o, and the result was the same. Interestingly, the answer is correct when asked which hand is being used in an image. For watches, it doesn't even notice when hands are incorrect. So, we're still at the level of "wow, there are 5 fingers." Even asking for an image of someone writing with their left hand is challenging - the AI was trained on datasets dominated by right-handed people and wasn't taught to generalize. My take on that is the following: Why does this happen? It replicates similar patterns without deep understanding - like someone trying to "write" Japanese by merely copying shapes. Can it be improved? Yes. Through reinforcement learning and other techniques like generative adversarial networks. At least for hands, since we have models that can distinguish left from right, we can use this for training. Why isn't it being improved? First, technical: Image generation is far more costly than text generation. It's easier to generate and evaluate text. By OpenAI's pricing, a 600x600 image costs about the same as 500 words. Moreover, there are many more ways to generate nonsensical images than gibberish text. Logical consistency is just one of many dimensions to optimize. Second, incentives: While image generation is fun to play with, achieving production-quality results is challenging. Specialized tools exist, but they're narrow in scope. Putting more effort into it is burning money, as we are in the uncanny valley of a product. In contrast, text is more general-purpose. Many people benefit from ChatGPT or Claude for editorial work. Software engineering uses the same models (I use them via Cursor) - so a general tool is also useful for professionals. While we can generate fun stories with LLMs, textual AI has already become a practical, powerful tool.