Over three decades since, I published my first papers on Artificial Intelligence as part of research in Universities and industries. At that time, my interest was in Neural Networks, with applications in pattern processing; the concepts do not differ much from those currently in fashion, but the tools and interfaces have evolved a lot to the point where these ideas and models are viable.
I remember that one of the first successful applications I worked on was for the oil industry; it was about neural networks for analyzing charts produced in the inspection of wells. The charts were scanned, and the images were analyzed by a first neural network that determined the state of the deposit based on pre-determined categories. A second neural network generated actions based on the analysis and historical data of the deposit. Get to have close to 90% agreement with the analysis and recommendations of human experts.
That application used one of the most widely used reinforcement learning models. The neural networks were trained using examples provided by human experts; the more samples, the better the performance. A refined version of that model has been renamed “Reinforced Learning with Human Feedback (RLHF)” and is used by applications like ChatGPT.
Something inquisitive is that ChatGPT doesn’t know the acronym of the model it is based on:
And when I clarified the acronym, he apologized and replied with the following:
This example allows me to talk about what has made AI finally come to the mainstream.
The dialogue I included above can lead us to conclude that the agent understands what we are saying and therefore generates an apologetic response before elaborating on the concept of RLHF. The reality is that, as in the primitive neural networks of 3 decades ago, the underlying program has no fundamental understanding of the language but instead uses pattern processing mechanisms…