We had implemented an AI solution in 2024. These days we are working on updating the underlying Generative AI models for better functionality and performance. The prompts used in 2024 had a lot of instructions to make AI work the way we wanted to - some transformations, a few guards, and a couple of examples. Soon we realized that the old prompts were not quite effective with the new models - in fact we did not need too many specific instructions with the newer models. It was a moment of unlearning for us, and it was quite recognizable as the experience is within one year, in the same technology and domain.

In my career spanning over 25 years, I have worked from COBOL to Generative AI, from mainframes to clouds and from terminals to mobile apps. In one perspective it was about constant learning - learning new domains, technologies and paradigms. Looking back it was more of unlearning than learning that I had to go through.

While every new technology and technique involves something to be learned, the risk is carrying the concepts and practices from the old technology. Oftentimes, this is what causes the transformations to fail, as we hold on to old concepts and practices in the new world.

This is true with digital transformations, agile practices, or DevOps platforms. I have seen quite a lot of instances where digital transformations become microservice implementations, agile implementations turn to be mini waterfalls, API platforms become wrappers of application integrations, blockchain ledgers become data repositories, and GenAI implementations become over-engineered function calls.

Learning will happen automatically while adopting new patterns and practices. However it takes effort to unlearn. Continuously unlearning through contrasting the new with the old is probably the best way of adopting new concepts and technologies.

It is also important to realize that unlearning is not undoing. Unlearning the old does not invalidate the old technologies or solutions. In the context of the new, the old best practices may need to be re-evaluated, re-defined or even avoided at times.