AI Explains 9.11 Greater Than 9.8

According to their release, one possible reason models say 9.11 > 9.8 is because they are being influenced by the "Sept. 11th" (9/11) and "gravity" (9.8 m/s^2) neurons. Model Chat (Llama-3.1 8B Instruct) Which is bigger 9.8 or 9.11? [get_id] The model has discovered that 9.11 is bigger than 9.8. This token looks wrong 9.11 is less than 9.8. Click the token to surface reasons that influenced the mistake. Continue chatting with the model... Transluce Model Investigator Hello! I'm Transluce's model investigator. I can help you investigate and debug language model behaviors. Show me how to fix AI's inability to compare 9.11 and 9.8. Sure! Let me ask the model to compare 9.11 and 9.8. I asked two models that their biggest errors at the start. I found these three classes: references to historical attacks, numbers, social ... Using the interface, we've discovered satisfying explanations for surprising model failures. Perhaps most interestingly, for the famous error where models incorrectly state that 9.11>9.8, we uncover "gravity" (9.8 m/s^2) and "Sept. 11th" (9/11) neurons that contribute to the incorrect comparison. We more generally uncover 'bible verse' neurons that lead many numbers to be incorrectly compared, and find that ablating these fixes the issue. Announcing Transluce, a nonprofit research lab building open source, scalable technology for understanding AI systems and steering the...

Скриншот поста в X с объяснением, почему языковая модель утверждает, что 9.11 больше 9.8. Шутка в том, что модель якобы руководствуется ассоциациями с 11 сентября и ускорением свободного падения, а не математикой. Мем используют как реакцию на абсурдные объяснения ошибок ИИ и чрезмерно уверенные исследования нейросетей.

• @dev_meme / 6344 2024-10-25