John Duprey At work, one of the prevailing stories of 2024 was Thomson Reuters’ journey to train our own large language models (LLMs). I primarily helped the team evaluate third-party solutions to accelerate these efforts. A significant challenge was finding the necessary computational resources that also met our information and cyber security standards. This journey ultimately led us to adopt AWS HyperPod, and we were early adopters of the Kubernetes version of the capability.
AWS London Summit in April
I presented our story at the AWS London Summit in April. I co-presented with AWS colleage Simone Zucchet.
See it on YouTube - you scan scrub to 22:42 for my segment.
Follow Up Blog Post in July
A slightly updated two-part blog post followed in July:
- Part 1/2 — Scaling Thomson Reuters’ Language Model Research
- Part 2/2 — Scaling Thomson Reuters’ Language Model Research
The Journey Continues…
The journey continued with Thomson Reuters’ acquisition of Safe Sign - infusing additional expertise into the team. I assisted with onboarding the Safe Sign team’s technology and worked with the combined team to arrive at a common technology stack. My role here was largely high-level architectural oversight and technology management - coordinating with our standards and governance and security teams.
To quote, part two of my blog:
While we do all of this, very large commercial LLMs continue to advance. With ever-increasing context windows and increasing availability/capacity, whether smaller, tuned, LLMs benefits will outweigh the cost is still to-be-determined for us.
2025 will be an important year for this effort where we, hopefully, start to realize some return on our investment. OR learn and pivot - as is the nature of R&D.