Comment by seniorThrowaway
6 hours ago
AI's / LLM's have already been trained on best practices for most domains. I've recently faced this decision and I went the LLM custom app path, because the software I needed was a simple internal business type app. There is open source and COTS software packages available for this kind of thing, but they tend to be massive suites trying to solve a bunch of things I don't need and also a minefield of licensing, freemium feature gating, and subject to future abandonment or rug pulls into much higher costs. Something that has happened many times. Long story short, I decided it was less work to build the exact tool I need to solve my "right now" problem, architected for future additions. I do think this is the future.
> AI's / LLM's have already been trained on best practices for most domains.
I've been at this long enough to see that today's best practices are tomorrow's anti-patterns. We have not, in fact, perfected the creation of software. And the your practices will evolve not just with the technology you use but the problem domains you're in.
I don't mean this as an argument against LLMs or vibe coding. Just that you're always going to need a fresh corpus to train them on to keep them current... and if the pool of expertly written code dries up, models will begin to stagnate.
Also, they've been trained on common practices more than they've been trained on best practices. And best practice is heavily context dependent anyways.
I've been doing this a long time too. The anti-patterns tend to come from the hype cycles of "xyz shiny tool/pattern will take away all the nasty human problems that end up creating bad software". Yes, LLMs will follow this cycle too, and, I agree we are in a kind of sweet spot moment for LLMs where they were able to ingest massive amounts of training material from the open web. That will not be the case going forward, as people seek to more tightly guard their IP. The (open) question is whether the training material that exists plus whatever the tools can self generate is good enough for them to improve themselves in a closed loop cycle. LLM generated code was the right tool for my job today; doesn't mean it's the right tool for everyone's job or that it always will be. One thing constant in this industry is change. Sold as revolutionary, which is the truth, in the sense of going in circles/cycles.
What if there is a new domain.
Then it is new for everyone, no?
Humans can learn from new experiences. LLMs have to be retrained (continuous learning isn't good enough yet), or you have to fit enough information into the context while still having enough for the task itself.