I don't think so, at least not if you're using agents smartly for software engineering and not just firing prompts off. There's the default way agents build stuff, but they're very steerable and it's easy, with hard rules (like linting) and soft rules (like skills and AGENTS.md), to guide them to actually make half decent software architectures however you like.
Agreed. AGENTS.md and skills are the documentation level from the linked post that hands AI the context its training lacks.
However, this only works if you know a better architecture exists. Most won't and they'll use the "defaults". And these defaults is what the next model trains on. Individually you can steer, but I'm concerned it won't be enough to steer the critical mass of the training data.
And it's this output of humans that we are further training AI on. Will this progress things, or will it just keep us regurgitating 2020 code patterns?
I don't think so, at least not if you're using agents smartly for software engineering and not just firing prompts off. There's the default way agents build stuff, but they're very steerable and it's easy, with hard rules (like linting) and soft rules (like skills and AGENTS.md), to guide them to actually make half decent software architectures however you like.
Agreed. AGENTS.md and skills are the documentation level from the linked post that hands AI the context its training lacks.
However, this only works if you know a better architecture exists. Most won't and they'll use the "defaults". And these defaults is what the next model trains on. Individually you can steer, but I'm concerned it won't be enough to steer the critical mass of the training data.
No.
Just create your own guardrails. Lint rules. Hooks. Whatever. AI also keeps training on newer data. We move on.
And how many projects are actually doing this?
And if they are, how many are keeping this up to date?
> And how many projects are actually doing this?
How many are following any best practices? Whose problem is that? Not AI.
> And if they are, how many are keeping this up to date?
And how many projects use the latest Spring (mentioned in the article)?
Don't blame AI. Blame the humans.
And it's this output of humans that we are further training AI on. Will this progress things, or will it just keep us regurgitating 2020 code patterns?
[dead]