Conceptual Integrity in the Age of Prompt Engineering
In the age of prompt engineering, conceptual integrity is as important as ever. Natural language is becoming the next, higher level of abstraction in programming languages. With it, the scale at which we can write software and solve problems is unprecedented, but it's not free. Much of the decision-making process has been relinquished by the individual programmer, and the line separating design and implementation has been blurred as our prompts encode both. Also, while we can trust our programming language and compiler to formally produce assembly language (likewise assembly to produce machine code) so that we virtually never have to read or understand assembly, there is no such technology yet to formally map any natural language to any programming language, so we cannot give AI-generated code the same degree of blind trust. The more details a statistical model handles for us, the more room there is for its decisions to drift from our own reasoning, and over time even its own. This risks the loss of conceptual integrity, which leads to the creation of an incoherent black box that will inevitably become legacy tech debt.
So as we scale up programming, we must scale up our decision-making as well. Our processes and prompts must be coherent and powerful enough to match the level of AI design and implementation, in order to maintain a level of control and conceptual integrity of the systems we build. It’s up to us to remain diligent, and to train ourselves in the theory and practice of what we’re doing so that we keep our decisions - our prompts, as effective as possible. Also, much as we need to learn a programming language before we can code, and having an understanding of the hardware and infrastructure we run our software on improves it, it will pay to understand how generative AI works under the hood. At a basic level we can be more efficient with token usage and compute, but we can perhaps see improvement in other virtues such as correctness and maintainability. AI is a powerful teacher; as we use it, we can learn from it, reinforcing our own knowledge so that we can in turn prompt, build and solve problems better.