5 Key Lessons for Building Effective Foundation Models in AI from Ordnance Survey's CTO

Many businesses are just starting to explore the potential of artificial intelligence (AI), but some, like Ordnance Survey (OS), the UK's national mapping service, have been leveraging machine learning (ML) and other advanced technologies for over a decade. Manish Jethwa, CTO at OS, shared insights with ZDNet on how the organization is integrating its AI and ML expertise with recent advances in generative AI to enhance and distribute its vast geospatial data. Here are five essential tips from Jethwa’s experience for building foundation models for AI: Develop a Strong Use Case Jethwa emphasized the importance of creating foundation models tailored to specific needs. OS builds models from scratch, using its internally labeled data for tasks like extracting environmental features. These models, while smaller in scale compared to those from large tech companies, outperform them in specialized domains because they are trained on a narrower, more relevant dataset. For instance, OS's models identify features accurately by focusing on a particular mix of urban and rural environments. This approach allows the organization to reuse the models for various tasks, such as analyzing roof materials or biodiversity, by simply fine-tuning them. Establish Purposeful Methods Cost efficiency is crucial when training foundation models. OS starts with a small model using a few hundred labeled examples to validate the concept and ensure they are heading in the right direction. They then gradually increase the labeled dataset, eventually reaching hundreds of thousands of examples. By breaking down the training process into manageable chunks, OS avoids excessive resource consumption and ensures that the model’s performance aligns with their specific requirements. Jethwa highlighted that executing models requires far fewer resources than training them, making this approach fiscally prudent. Use Other Large Language Models (LLMs) for Fine-Tuning While OS builds its own foundation models, it leverages commercially available LLMs for fine-tuning and complementary tasks. The organization uses Azure machine learning models and Python-based tools, often collaborating with partners like IBM to enhance its data processing capabilities. By building on existing models and tailoring them to their unique needs, OS achieves better performance without reinventing the wheel. This hybrid approach also helps in rationalizing the development process and keeping costs under control. Think About Commercialization OS’s foundation models represent valuable intellectual property due to the Crown copyright that applies to assets created by UK public sector employees. Jethwa noted that while commercialization is a possibility, it must be handled carefully to prevent the exploitation of taxpayer-funded assets without benefiting the UK. OS is exploring ways to share its models while maintaining the integrity and value of its data. The goal is to strike a balance between protecting sensitive information and maximizing the utility and distribution of their AI capabilities. Keep One Eye on the Future Jethwa envisions a future where generative AI significantly enhances user interaction with geospatial data. He imagines a scenario where users can engage in a conversational interface to query detailed maps. For example, a user interested in a particular area could zoom in and specify their interest, such as "schools." The AI would then ask follow-up questions to refine the search, translating these queries into definitive answers from authoritative sources like OS. This forward-thinking approach highlights the importance of using APIs and trusted data to create accurate and useful responses to user requests. These tips from Jethwa offer valuable guidance for businesses looking to build and deploy AI foundation models effectively. They underscore the need for targeted use cases, efficient training methods, strategic use of existing models, cautious commercialization, and a focus on future user interfaces. Industry insiders highlight the growing trend of organizations combining internal expertise with commercial AI tools to create customized solutions. Ordnance Survey, known for its precision and reliability in mapping, demonstrates how this approach can yield significant advantages in specialized applications. By prioritizing relevance and user experience, OS is setting a benchmark for how businesses can harness the power of AI while navigating complex regulatory landscapes and cost considerations.