The Model Factory Is the Real Asset: Lessons from Thomson Reuters and Other Companies Training Their Own Models

bigdata · x · 2026-10-22

Ben Lorica (Gradient Flow) distills lessons from interviewing Thomson Reuters CTO Joel Hron and surveying companies doing "deep model customization": starting from strong open-weight models, continued pre-training on proprietary data, followed by multi-stage post-training. Cursor, C3 AI, ByteDance, Cisco, and Reddit are exploring variants of this approach.

Key findings:

Fine-tuning isn't always enough: Thomson Reuters began with LoRA on commercial models; targeted tasks improved but other capabilities degraded.
Retrieval's limits: C3 AI's proprietary Type System doesn't appear in public training data — retrieval alone yielded an 11% pass rate on its internal coding eval; adding agents and tools pushed frontier models to 60%, but responses took about two minutes. The information was available; the models simply hadn't learned how the system worked.
Takeaway: RAG solves for fast-changing or bulky information, post-training reshapes behavior, but continued pre-training changes what the model knows about a domain — the move that matters when a company's moat is proprietary domain knowledge.

Original post →

The Model Factory Is the Real Asset: Lessons from Thomson Reuters and Other Companies Training Their Own Models

More from coding & agent