For All AI Researchers, Developers

A researcher argues that the fundamental difference between humans and AI systems is the lack of a doubt mechanism in AI, which causes hallucination in large language models and limits autonomous vehicles from handling edge cases. The author calls for future AI research to focus on developing doubt mechanisms to achieve human-like AI rather than merely outperforming humans.

-One of the main differences between humans and AI systems I start with my conclusion first. The difference is that humans doubt and AI systems don’t. The fundamental reason of hallucination of LLMs is that LLMs don’t have doubt mechanisms. Because AI systems never doubt their results or inferences, hallucination arises. My last post “A tricky method for NLP and LLM from my previous research work” is one of my many tries to add doubt mechanism to NLP systems. And I guess some friends here already have found out my idea behind this article. I think future AI research trends should include study on doubt mechanism. Then let me explain how or why I thought my last post can be one of the tries for doubt mechanism. Once I saw a math Olympics of middle school students. Almost all of them did the same thing. After they had solved math problems they doubted whether their answers were right. Then they checked their answers. How did they check? They fed problems with their results maybe the result x=“” when we solve equations . I thought this is the main difference between human brains and AI systems. Then feeding AI systems with outputs could be a right way to future AI. I think without doubt mechanism, AI can never act human-like. ‘Probably’ Ai systems can act better than humans, but never act like humans. When I first learned about transformers, especially “attention mechanism”, I was really excited and thankful for that AI team. Because attention mechanism was literally a typical attempt to make AI systems think like humans. But nowadays among AI teams I can hardly find any efforts for such goals. -For AV, Robotics, If LLMs need a doubt mechanism, what about AVs? Here I only give one example. If what I recall is right, Tesla approach is like this: AI systems learn videos of drivers. Here comes the same question. What AI systems of AVs learn from videos is the way drivers drive, they never learn the way drivers think. Then what are main differences? For example, when a driver sees a person standing close to the highway, he might think that person could suddenly walk into the highway and he pays attention not main attention, maybe part of his attention to that passenger. And the passenger would stay in the corner of the sight of the driver. So what AVs really lack could be a special attention mechanism or “keep-an-eye-on-edge-case” mechanism . I think only learning from videos of drivers can never effectively solve edge cases. -Conclusion I have done so many researches, but I still try for “human-like AI”. I think I can find some friends who have similar ideas. I hope more people will try for “human-like AI” rather than “better than human AI”. And my research works on the second difference between human brains and AI systems I found will be posted later. Thanks for your attentions and replies.