The Rising Tide

LLMs set a baseline for what's possible. Builders wrap products around them, squeezing out maybe 20% more through clever prompting, RAG pipelines, or fine-tuning. No amount of engineering tricks makes it fundamentally better -- you can't turn a 3.5 into a 4. But you can apply it in new domains, find novel use cases, build workflows that weren't possible before.

The problem is attachment. Builders invest months into their scaffolding and start believing it's the secret sauce. Marketing kicks in. "Our proprietary approach" becomes "the next best thing since sliced bread."

Then a new model drops. The baseline rises. All that careful engineering is now underwater—the new model does it out of the box. Your clever retrieval system? Native context windows are longer. Your chain-of-thought wrapper? Built into the model. Your fine-tuned variant? The base model caught up.

The scaffolding becomes legacy code nobody needs.

Here's what stays valuable: the application layer. Each generation opens a wider window of what's feasible. Tasks that were too error-prone at the old baseline become reliable at the new one. The builders who win aren't optimizing the last 20% -- they're finding the next use case that just became possible.

Apply it now. It might be rough. It will make mistakes. But the tide keeps rising, and waiting for perfection means someone else ships first.