Thomson Reuters CTO: Building a Frontier Model Cost $450K, Not $40M
bigdata · x · 2026-10-08
Thomson Reuters CTO Joel Hron explains on The Data Exchange why the company built its own model, Thomson, on top of open-weight models despite frontier APIs being available.
Key points:
- Economics: training a specialized model cost $450K versus $40M for a from-scratch frontier model, leveraging decades of proprietary content and 3,000 domain experts.
- Approach: started from Qwen fine-tuning, evolved into a full training pipeline that swaps in new open-weight models; selective pre-training, RL, and agentic RL.
- Production: CoCounsel orchestrates and routes between self-hosted and frontier models, reducing third-party reliance; legal/tax workflows stress verification and human oversight.
- Takeaway: enterprises increasingly prefer open-weight models for control.
More from Companies & People
- Tavily, LangChain and Nebius host SF Tech Week event with 100K credits for startups — LangChain · 2026-10-08
- AI agent swarms have 3x the viable attack paths of human hackers, Cogent warns — cneuralnetwork · 2026-10-08
- Ecosia drops Mistral for open-source models incl. Chinese ones; Mensch jokes nuclear power is a blocker — uwukko · 2026-10-08
- 47 top Indian indie founders gathered for 3 days at a Bangalore resort — ayushtweetshere · 2026-10-08
- Sunil Pai on Reputation: 'Collecting Favours', Lessons From His Doctor Father — threepointone · 2026-10-08
- OpenAI's agentic software factory detailed — but zero verified cases of fully autonomous long-horizon coding — alex_verem · 2026-10-08