Plugin pricing in the AI era

[VariKusBrainZ]

corrupted audio buffers Will be all the rage next year

[stoopicus]

I tried the most recent GPT models with JUCE as an experiment recently and found that they still make a ton of mistakes. You really need to steer it very carefully, ask it to do research on API docs/examples etc, otherwise it will still make a mess of it, causing corrupted audio buffers etc. I found that surprising given how advanced they have become, but maybe it is just an issue with the training data. So I wouldn't call coding over quite yet, although I am sure it is coming eventually.

Usually more often that not it comes down to bad prompting and/or not using established software engineering practices.

For example...stage gate it. TASK -> PLAN -> IMPLEMENTATION.

Focus on making the task as complete and as well specified as possible. Spend most of your time here. When you think the task seems complete, get it to create a new agent to challenge and review the task.

Then from that get the A.I to generate a detailed implementation plan. Feed it reference to coding style, practices, existing paradigms and idioms. Once more, get it to generate a new agent to challenge the plan and fix any issues.

From there, get it to implement the plan. Get it to generate unit tests and allow it to measure for success/failure. When done, you've got it, get it to create a new code review agent to adversarially challenge the implementation and fix any issues.

Then get it to present a summary of work done.

Most of all, don't abdicate responsibility. You must know what you're doing. A.I. is an amplifier. It can amplify dross as much as it can amplify excellence. As always, garbage in, garbage out.

You'll be amazed how far this gets you. When used with a model like Fable, it's truly astonishing.

Thank you. This post nails what I was getting at. In actual, real terms for software engineering, we are seeing a shift more towards the more senior level roles of architecture, design, and managing engineering resources; meanwhile AI is quite good at the tasks of more rote coding and testing and other aspects that while fundamental, are also more entry level table stakes. I feel bad for GenZ because AI is effectively replacing the role of entry level software engineer.

It's still spectacularly bad at unsguided, one-shot development.

For now.

But coding? Coding is over. None of the best engineers I know spend much if any time in editors these days.

[stoopicus]

Code is art.

I want to also say I am not arguing with this - some of the most beautiful code I have seen was assembly written long after asm was required. I'm thinking stuff like the demo scene, or more easily tangible the BSD bootloader, which from the very first line is a work of art.

[ksandvik]

Yes, we will miss the talent of engineers that for example managed to squeeze in most of MacOS into 64k ROM... LLM uses a lot of open-source, github et rest material, and as most of us know, that codebase is not exactly elegant.