What's Happening?
Thomson Reuters has launched its first in-house AI model, Thomson-1, aiming to decrease its dependence on costly external AI solutions like Anthropic's Claude. The new model is based on Snowdon, which was developed by 'realigning' an open-source Qwen
model from Chinese tech giant Alibaba. According to Joel Hron, Thomson Reuters' Chief Technology Officer, Thomson-1 will initially focus on tasks within the company's expertise, such as document review, which were previously handled by Claude. Hron clarified that this move is not intended to entirely replace their collaboration with Anthropic and other AI labs but rather to build internal intellectual property and manage AI costs. The company expanded its partnership with Anthropic in May for its AI legal assistant, CoCounsel, which still largely relies on Claude, but the long-term goal is for Thomson-1 to power more of CoCounsel's capabilities.
Why It's Important?
This strategic shift by Thomson Reuters highlights a growing trend among large U.S. companies to develop proprietary AI models to mitigate the high costs associated with third-party AI services and to build internal expertise. By leveraging open-source technology, even from Chinese providers, companies can customize AI solutions to their specific needs and data, potentially gaining a competitive edge and reducing long-term expenditures. The move also underscores the increasing pressure on businesses to manage AI-related expenses, which can be substantial for advanced models. For the U.S. legal and data industries, this could lead to more specialized and cost-effective AI tools tailored to their unique requirements. However, the use of Chinese open-source models also raises concerns about potential security risks and geopolitical implications, as noted by some government leaders and AI companies like Anthropic, which has called for restrictions on such practices.
What's Next?
Thomson Reuters plans to gradually integrate Thomson-1 into more of its AI-powered services, particularly within its CoCounsel legal assistant, with the objective of having it power an increasing number of capabilities over time. The company will likely continue to evaluate and adapt its AI strategy, balancing the use of in-house models with strategic partnerships with external AI providers. The broader industry may see more companies exploring similar strategies of building proprietary AI solutions based on open-source models to gain greater control over their AI infrastructure and costs. The debate surrounding the use of Chinese open-source AI models will likely intensify, potentially leading to increased scrutiny or regulatory measures from U.S. policymakers regarding data security and national interests. Companies will need to carefully assess the risks and benefits of such collaborations, ensuring robust security protocols and ethical considerations are in place.
Beyond the Headlines
Thomson Reuters' decision to build its own AI model using Chinese open-source technology delves into complex issues of technological sovereignty, intellectual property, and global supply chains in the age of AI. By comparing it to 'renting versus buying a house,' Hron articulates the long-term value of owning and developing internal AI capabilities, fostering equity and control over critical technological assets. This move could inspire other U.S. companies to invest more heavily in internal AI research and development, potentially leading to a more diversified and resilient AI ecosystem. However, it also brings to the forefront the geopolitical tensions surrounding AI development, particularly the concerns about potential backdoors or security vulnerabilities in foreign-sourced open-source models. The ethical 'de-biasing' and safety adaptation of the Qwen model by Thomson Reuters and Imperial College highlights the critical importance of responsible AI development and the need for rigorous vetting of underlying technologies, regardless of their origin.
