LLM's Future Hinges on Post-Training and Proprietary Models

RihardJarc · x · 2026-07-15

A former LLM lab employee shared insights on the future of LLM usage and the growing importance of post-training:

Application/product companies and model providers are converging; long-term, product companies will likely train their own models.
Training cost structures have shifted: pre-training previously accounted for roughly 80% and post-training 20%, but the split is now closer to 50/50, encouraging more companies to pursue pre-training.
Relying solely on private user data as a moat is overestimated; data reflecting real-world usage patterns is far more valuable.
App companies without proprietary models will be increasingly dependent on closed-source model vendors, as it's easier for model companies to build products than vice versa.
The future will rely heavily on prompt routing to various models, with smaller, cheaper models handling many tasks instead of defaulting to frontier models.

More from Companies & People