Thomson Reuters goes its own way
Thomson Reuters is launching 'Thomson,' a proprietary large language model built on Alibaba's Qwen framework, according to a report from The Decoder. The company's CTO Joel Hron emphasized the strategic rationale: 'What matters isn't intelligence itself, but knowing which intelligence you need to own.' The model, costing roughly $40 million over two years, is designed to integrate deeply with Thomson Reuters' own content, such as Westlaw legal databases. Benchmarks show top marks only when the model can tap into that exclusive data, suggesting the company's competitive edge lies more in its curated information than in raw model capability.
Inherent's research teammate
Separately, Inherent, a London-based startup founded by DeepMind alumni, has unveiled 'Faraday,' an AI agent specialized in replicating scientific research. According to TechCrunch, Faraday outperforms models from Anthropic and OpenAI in reproducing experimental results from published papers. Inherent positions Faraday as a 'teammate' for scientists, automating the verification and extension of existing work. The agent uses a combination of retrieval-augmented generation and code execution to simulate experiments.
A shift toward ownership
Both developments reflect a growing skepticism among businesses and research institutions about relying on general-purpose AI from dominant providers. By owning the model and the data pipeline, organizations can tailor capabilities, control costs, and protect proprietary knowledge. Thomson Reuters' approach also hedges against potential price hikes or policy changes from companies like OpenAI and Anthropic. However, the $40 million investment is significant and may not be feasible for smaller firms. Inherent, by contrast, aims to offer Faraday as a service, but its success depends on the accuracy and reproducibility of its outputs—a perennial challenge in AI-driven science.
Critics warn about reproducibility and lock-in
Some researchers caution that AI agents like Faraday may overstate their ability to replicate human-led experiments, as subtle contextual details are often lost. Meanwhile, the move to build proprietary models risks creating new data silos, making cross-institutional collaboration harder. Thomson Reuters' reliance on Alibaba's Qwen also raises questions about geopolitical dependencies, especially as US-China tech tensions persist.