You basically need both. Parameters and good post training. If you keep on growing both, you’ll have good models.
LLMs are still surprisingly “easy”. You need maybe a couple dozens of right people, a lot of good quality data and a lot of GPU that you know how to operate. There’s relatively little “secret sauce” needed.
How to structure experiments/scaling and hyperparameter tuning regimes are most of the secret sauce (besides massive compute). If you don't create an experimental ladder to verify scaling and optimize your hyperparameters well, you'll waste a ton of money.
The data is mostly coming from places like Scale/Mercor/etc and net dumps with some filtering and batch prioritization, and RL on verifiable domains like code/math/games.
I think there's still a ton of secret sauce needed for serving them economically
Sure, same for building a model in an economically sustainable way. But barier to entry is surprisingly low (expect for the huge amount of cash, of course). That’s fairly surprising, given how extremely powerful that tech is.
10 years ago it was super hard to have usable “frontier” ML. You needed very complex data warehouse, feature engineers, feature stores, multi level ranking, calibrations, tons of different model architectures, etc, etc. Each by itself was extremely hard engineering problem and really only handful of companies could deal with that complexity.
With LLMs, 95% of that is gone, infra to support them is greatly simplified. Of course, to make really reliable, performant, user friendly, etc - you still need to a lot of engineering. But it’s very different challenge.