Chatbot Software Begins to Face Fundamental Limitations

Since even my 10 year old son can easily solve such problems using a spreadsheet technique, I think an AI that can generate and execute code can be taught to solve such long logic puzzles in the same way.

The problem IMO is the internal memory states that need to be created and updated that LLM deals with, not the composition problem. I don't know if LLM can solve composition problems at all. However, as is typical of AI techniques, I suspect that people in the field will try to teach them a special trick to solve the problem set. Until the next one.

A fundamental problem is when such new tricks would cause catastrophic memory loss or similar collapse.

i think the title is a bit misleading. chatbots arent LLMs, though LLMs are maybe the latest iteration. chatbots themselves could use ensemble learning or even othher non ML techniques to overcome limitations of LLMs. the article itself is a little more clear on it, atleast mentioning this is LLM related limitations.

I am always disappointed in these articles that fail to mention that other forms of AI such as logic programming/intelligent search easily solve problems like the “Einstein” one. Similarly, it is not difficult for genetic programming methods to discover code which reliably multiplies numbers.

I am uncertain if this is a lack of understanding by the authors of these articles or just a bias to keep the audience’s attention on the shiny hyped up LLM business.