Just some notes from a sharing session on Generative AI:

Large language models

Generative AI leverages of large language models (LLM) – essentially dense neural networks with many parameters for tuning. These natural networks operate on the concept of “embedding” – translating non-vector/numerical inputs such as human text, or images into vectors/numerical inputs that can then be used by computers for computation

Different companies have trained up their own LLM – think of this as a basic functioning “human worker” capable to generic speech and comprehension. For example, ChatGPT has Davinci LLM, Cohere has Command LLM

Fine tuning

You can then fine-tune this basic “worker” to do more specialised tasks by training the basic LLM with additional, specialised data sets (prompt + desired answer) – what happens here is parameters can be adjusted or even added and you end up with your own custom LLM (which you can potentially IP, or safeguard within your own data centers if data privacy is important to you)

On biases

If your training data set is biased, your LLM is going to end up giving you biased responses. Even as humans we find it hard to check our own biases, but with LLM this is hardcoded into their very nature.

One thing we should do at this point of GenAI maturity, is to ensure there is a reviewer of the generated content before it is used for official purposes, to vet for any unintended biases.