ChatGPT has been a thing for a while now.  It's an extremely useful tool for summarizing texts and answering questions about topics in the public domain.  But what about questions that require analyzing a vast amount of private data?  

Large organizations generate an enormous amount of information they want to keep private (e.g. sales figures, research findings) or are legally required to keep private (e.g. customer data).  Datacakes specializes in creating custom, wholly-owned systems where an AI (like GPT) can analyze your structured data, without the need to prepare or export it.

For a demo of how we've set up a custom platform to work with private data, check out CUBIE, our AI Assistant connected to QuickBooks.

Hallucinations are one of the major risk points of current LLM technology. Sometimes, an AI will make up people, events, or data that do not actually exist in order to answer a question it was asked. This is particularly difficult to deal with, because the LLM is good at making its answer sound like fact.

Hallucinations are made worse by an over-reliance on prompt engineering. It is tempting for the human user to load their question prompt full of requirements, to ensure a certain kind of response. But there is no guarantee that the AI will meet each of those requirements, and its unpredictability may even increase with an overloaded input.

For our custom solutions, we've developed a proprietary approach that combines LLM thoughts with more classical, deterministic functions. We've found that this reduces the variability of the overall response and allows the handling of more complex questions. (Our research in this area is ongoing.)
‍