What is in a ChatGPT foundational model
When an LLM is built, it is trained on sources of data from the internet. It knows publicly available information about companies and products. If asked typical enterprise-like questions, it can get robust answers – sometimes better than what is available from some vendors’ websites. For example:
What are the advantages of Hana for a database? What is a good value for SGA for an Oracle 12.2 transactional database? Can you easily replace the battery in an iPhone? How do I return a product to Costco?
Try these questions out and notice a trend. Each answer is slightly more generic than the previous one, and that generic nature is part of the problem.
The following applies to most foundational models such as ChatGPT 3.5 or 4o, Anthropic’s Claude, Meta’s Llama, or Mistral7B:
- Don’t understand specific business or use context or complex products
- Don’t have customer history or context to consider...
