Gabe argues that vertical application companies no longer need to build everything in-house; by leveraging open-source base models, NeMo lab partnerships, synthetic data pipelines, and mode…

Vertical AI companies blocked from training on sensitive customer data can use domain experts (lawyers, doctors) to guide synthetic data generation via coding models, then scale with platfo…

In domains like legal where customer data is privileged, using domain experts to guide synthetic data creation (via tools like Mercor, Snorkel) solves the cold-start problem and enables rig…

With strong open-source base models (GLM, Kimi, NeMo-Megatron) and accessible post-training APIs (Fireworks, Baseten, Tinker), application companies can now build specialized models that co…

Models like Kimi 3, GLM 5.2, NeMo-Megatron, and Inkling have reached sufficient base capability that post-training on domain-specific data can yield frontier-level performance on narrow tas…

Before investing in post-training, companies must build robust model-serving infrastructure including multi-model routing, automated/human evaluation gates, A/B testing, and production moni…

The ultimate product is not a single best model but enabling each enterprise to continuously customize models on their private work streams while preserving data privacy, requiring breakthr…