Researchers at Goodfire AI isolated memorization from problem solving in large language model(LLM) neural networks.
I think this is additional evidence of the fact that LLMs are very inefficient in a manner similar to blockchain. Blockchain was capable of solving real world problems by wasting large amounts of energy. However, the inefficiencies of blockchain weren’t being covered up by billions of dollars in subsides and exploitation of low wage workers.
If problem solving is isolated from memorized data, that means that it would be more efficient to use a model that is dedicated to solving problems without needing to memorize facts. Training an LLM takes time and resource, especially if it is being trained on problems that are irrelevant to one’s work.
The isolation of problem solving implies that an LLM has the capability to categorize data and place the data in different parts of a model. As a result, problem solving capabilities for different subjects would also be split up. If a model is trying to reason about astronomy, it wouldn’t make sense for the model to be influenced by facts about history. which means that LLMs would use different model sections to reason about the two subjects.
If one wants to use an LLM to look up facts, it would be more efficient to use one that is just trained on memorized data without any reasoning abilities. Also, tools such as Wolfram Alpha have existed for years to provide information about memorized facts, and they don’t make up information like LLMs sometimes do.
To put things into perspective, I have been regularly benefiting from AI when riding the Red Line of the DC Metro. It now uses automated train control(ATC) to drive trains, which has made train rides smoother. This is an example of task-specific AI, that works 100% of the time, and the ATC developers were able to save time by ignoring irrelevant data.
Sources