Don’t be misled by LLMs

LLMs, the most famous one being ChatGPT, have revolutionized access to information, delivering pointed answer to your question, packaged in natural language that reads authoritatively. The natural inclination is to take the information as fact, and act on it, but beware, because the information can be misleading, wasting considerable time as you proceed based on it and find the outcome to be frustratingly poor. Hallucinations – is the term they use for this phenomenon.

Worse still, hallucinations occur not only for generic information that LLMs are pre-trained for, but also for specific information about your business that you have fed the LLM (by transfer learning). Instead of gaining productivity, hallucinations can reduce it by misleading you.

It behooves us to validate the LLM response, especially to the critical questions. How do we do it? In a couple of different ways. One is to run the same query across diverse LLMs, compare the responses you get from them, and if they are semantically equivalent, it is likely the response is correct. This can get expensive if you pay per transaction for the LLM access; also, it isn’t easy to determine if the responses are equivalent if the differences are fine-grained, even if that subtle difference can have considerable adverse impact.

Another way is to use an AI/ML platform, adept at sophisticated NLP, that queries the specific information, typically from your internal documents, finds relevant sections, compares them to what the LLM has returned, and only if it matches the fine-grained semantics, presents the LLM response as accurate, along with the relevant sections it retrieved so that experts can verify it for themselves. SLMs provides a Responsible AI/ML platform that does precisely that!

Source: eazyml

Contact us

Whether you have a request, a query, or want to work with us, use the form below to get in touch with our team.