Raw LLM Responses

It’s amazing how the tech leaders managed to fek this up so badly… gen AI is amazing at writing code. HOWEVER, it has important limitations like ‘context dilution’ that anyone with a brain would realize poses real challenges for generating and maintaining large code bases… so significant that trying seems like a foolish errand. On the other hand my team has had great success using AI for code generation - the approach we take is pretty simple… do some significant systems engineering upfront - and gen AI is really useful for this - to identify core functionality and interfaces and implement each function as a microservice with a well-defined interface… and when I say microservice I really mean ‘micro’ as in no more than a couple of 1000 of lines of code… we then write software requirements for each service, and a PR schedule to implement… and again AI is great at supporting this activity. Then we have AI write the code for each PR - we test in between each PR implementation to make sure we’re getting what we think we’re getting… and once we have all the building blocks we can pull together those microservice wrapped core functions in a multitude of ways to build full solutions for our customer with some custom glue and a slick UI. This is not ‘vibe’ coding. And it might seem like the process I’ve described is complicated and burdensome, but with it we have built some pretty impressive solutions in weeks that only a few years ago would have taken months, if not years. Gen AI thrives on well-defined manageable (I.e., not 50000 tokens in a single request) context… so if you do your due diligence with planning and take a very modular approach, the resulting generated code quality is much higher and can be a significant productivity boost. Seems to me that too many users of gen AI are either lazy, hugely over-estimating it’s capabilities, don’t understand the limitations, or how to mitigate them, or are using it to do things they’re not qualified to do themselves - you really need to know what you’re doing.