I agree these things are going to move likely faster than we can imagine and become better than we can imagine. That said, last week I got a ticket that had no body but a title akin to "Add a property for this" with no specifications about the system. When I asked for clarity on the ticket the ask was very different than the title implied. When I started the work it started to look different than was described so I went back to the requestor and explained the discrepancy and we changed the direction of the change.
I say this to say, one skillset I have as a developer is taking the vague requests product owners have and figuring out how to turn them into actionable code steps in a massive existing codebase with several repos. I don't say this to say I'm impossible to replace but to say that half the time people don't even know what they want or how to describe it. Then from there you have giant codebases that wouldn't fit in anything but the biggest (current) context windows.
I agree the accuracy limitations will likely evaporate but these things aren't necessarily something an LLM can solve. I'm probably going to be proven wrong over time but I use GPT for code pretty regularly and right now I'm not too worried about my job.
I don't know for sure if an LLM could do it. But theoretically one could build a system that sends chat messages asking for clarification and also eventually with more context or something is able to translate vague or stupid "requirements" into ones that make sense.
In five years or so the capabilities may be pretty amazing.
Definitely agreed! I'm both nervous and excited. I fear for my livelihood but also if we continue to make even a percentage of the progress we've made in the last year, the next 5 years are going to be wild!
I say this to say, one skillset I have as a developer is taking the vague requests product owners have and figuring out how to turn them into actionable code steps in a massive existing codebase with several repos. I don't say this to say I'm impossible to replace but to say that half the time people don't even know what they want or how to describe it. Then from there you have giant codebases that wouldn't fit in anything but the biggest (current) context windows.
I agree the accuracy limitations will likely evaporate but these things aren't necessarily something an LLM can solve. I'm probably going to be proven wrong over time but I use GPT for code pretty regularly and right now I'm not too worried about my job.