If I want to invest time in fixing some junior developer's probably broken code, I'd much rather do it with a real junior developer who has the potential to learn over time and become a not junior developer
I'll give you an example of something that I moved to LLM code generation.
I have used Excel for probably more than two decades to generate certain types of code structures that are simply tedious to both write by hand and also maintain. One such example are lookup table definitions for state machines. Using Excel for this, once setup, makes it fast and infinitely more maintainable.
Today, I can do this with LLM's in a heartbeat. The code is good and it is easy to maintain.
Over the years I have written custom code generation tools for the same reason. Sticking with the state machine example, if you have 100 states and each state requires entry, during and exit handler functions, well, you have to take that lookup table and write 300 functions. In a language like C this means both the declaration and implementation skeleton. LLMs are good for this as well.
Like I said, it's a tool in the toolbox. It can be used a million different ways.
As I said, this topic is likely to remain controversial for some time. I have a feeling that a lot of people are reacting to it rather than actively using and understanding where we are today.
I have been using it actively and across a wide range of application domains (not just coding) for probably somewhere around a year. This use has been under a "trust but verify" modality, being careful and skeptical. And, yes, I have come across issues here and there. And yet, in the aggregate, it would be impossible to envision not using this tool. It is most-definitely a potentially huge productivity and time-to-market booster. And it will only get better.
I believe in this to the extent that we decided to pivot a hardware-centric startup we've been pursuing to make it software-centric with AI agents at the core of the offering. These tools will change the way we conduct business, write software, design hardware, conduct research, administer healthcare and more. They might not be ideal today. These are just babies learning to walk.
My impression so far is that they're already in diminishing returns territory. They might just be puppies learning to walk, in the sense of having a fundamental limit to their potential
Well, the demand from every potential customer we have spoken to is very real. And the pain point we have identified cannot be addressed without using AI. Now it is just a matter of getting to a reasonable MVP. We already have customers who want to be in the beta program.