Frontier LLMs are easiest to work with for now. Open models _will_ catch up. We can be excited for that future.
You are able to learn things from LLMs, and you can ask them for recommendations for an approach to implement something. Or just tell your LLM the approach to take. Sometimes it overcomplicates things. You’ll develop an instinct for when that’s likely. You can head off the overcomplication ahead of time or ask for refactorings after the initial cut is built for you. After a while you get an instinct for which way will get the work done soonest. Most fascinating of all, it’ll all change again with the next round of frontier models.
You don’t need frontier models for every task. For instance I’ve been positively surprised by Github Copilot for straightforward features and fixes. When it’s obvious how to implement, and you won’t need to go back and forth to nail down finer design details, getting an initial PR from Copilot is a great starting place.
To everyone starting out, enjoy the ride, know that none of us know what we’re doing, and share what you learn along the way!