Geeky Codes

Geeky Codes

2/24/2025, 3:44:41 PM

All the LLMs and other deep learning models are based on neural networks. We can think of them as mathematical functions with hundreds of billions of parameters. Those parameters (weights) are determined during training and we train these networks with trillions of tokens (text, images, videos that are split up into tokens to be ingested by the models). We can say that every one of the trillions of tokens played a part in determining the value of each of the hundreds of billions of parameters. The image I have in mind is a giant lake where we dissolve trillions of cubes of salt, sugar etc. After the dissolution we cannot know which of the cubes of sugar went where in the lake - every cube of sugar is everywhere! Therein lies a problem: if we use a business database, such as customer relationship data, to train a neural network model (i.e to determine its parameters), when the customer changes that data or deletes the data, we do not know how to alter the weights of the model to account for this change in the data. Even if the model were dedicated to that customer, we still cannot guarantee the customer that their changes to the data will be reflected in the model. In that sense, neural networks (and therefore LLMs) are NOT a suitable database. This is a fundamental limitation of the current scientific mathematical approach and cannot be fixed only by technological fine tuning. The RAG (retrieval augmented generation) architecture keeps the business database separate and augments the user prompt with data fetched from the database. In that case, the model itself is not trained on the (potentially changing) customer data because that data is only used in the prompt. But RAGs can only go so far. I personally have come to believe more foundational work is needed. What does that look like? All I have right now are hunches. That is the existing part of scientific work!