LLMs…RDMS on steroids

RDMS – Info represented with key and attributes in a row (tuple). Each attribute value in the tuple defines a property of the key of the tuple and only that key in normalized form. Keys are unique. No nuance. Black and white. Great for info retrieval where accuracy is paramount. No guessing. Banking. Airline reservation. Works well with propositional calculus and relational algebra. Good example was IPWS at ATT. 4th generation relational language. SQL based on it. Joins, matches, sorts, etc. Jack Isler and I put millions of dollars on the ATT balance sheet using RDMS for Revenue Recovery.
LLMs. Info represented as multi-dimensional vectors. From Google —> In Transformer-based LLMs, each input token (word/subword) is converted into an embedding vector, which is then multiplied by three learned weight matrices to produce three distinct vectors: 

• Query (): Represents the current token that is looking for information.
• Key (): Represents all tokens in the sequence and is used for matching against the query.
• Value (): Contains the actual content or meaning of the token. 

• Attention Mechanism: The model determines relevance by taking the dot product of the Query vector with every Key vector.
• Weighted Sum: These matching scores determine how much focus to place on the corresponding Value vectors to generate the final context-aware output.
• Vector Database Context: When vectors are stored for retrieval-augmented generation (RAG), each embedding vector is usually associated with a unique ID (the key) to link the vector to the original text or metadata

You can see that LLMs are way more complex than RDMS in terms of info representation and process. Potentiallyheavily nuanced. Vectors are angles and angles that mean geometry and geometry can be measured like with cosine symmetry.

RDMS Black and white leans into grey in LLM land.

Bottom line is that LLMs can guess things with high confidence based on geometric relationships. RDMS cannot guess shit.In RDMS land things can only be what is defined..

This is why LLMs can hallucinate and why SQL cannot.

“You don’t have to swing hard to hit a home run. If you got the timing, it’ll go” – YB