The rush to adopt generative AI has sparked a critical dilemma for tech leaders: Should they leverage off-the-shelf large language models or invest in building custom foundation models? For Manish Jethwa, CTO at Ordnance Survey (OS)—the UK's national mapping authority—the answer lies in a hybrid strategy that marries proprietary data with commercial tools to solve domain-specific problems. In an exclusive interview with ZDNET's Mark Samuels, Jethwa outlined five pivotal lessons from OS's journey in creating foundation models that turn geospatial data into actionable intelligence.

The Domain-Specific Advantage

Jethwa's team develops foundation models trained on OS's crown copyright data to extract environmental features like building materials or green spaces—a task where generic models falter. "We're already outperforming models from large providers because ours are purpose-built," he explained. > "Where we're trying to extract features, we build foundation models from the ground up with our labeled internal data. This allows us to connect directly to the problem we're solving." The key insight? Start with a razor-sharp use case. By focusing on precise objectives (e.g., identifying roof materials), OS reuses a single foundation model for multiple applications through fine-tuning, avoiding redundant training cycles.

Cost-Constrained Innovation

Foundation model training is notoriously resource-intensive, but OS mitigates this through incremental development. They begin with small-scale models using hundreds of examples before scaling to hundreds of thousands—far fewer than the millions used in generic LLMs. "We have to be mindful about wasting cycles," Jethwa emphasized. > "Building up labeled data takes time, but executing the models consumes less energy. We validate direction early to avoid fruitless effort." This methodical approach not only controls costs but also yields higher accuracy for niche tasks, proving that smaller, focused datasets can trump vast but diffuse training corpuses.

Strategic Fine-Tuning and Partnerships

While OS builds its foundation layers, it doesn't shun external tools. The team fine-tunes commercial LLMs like those in Microsoft Azure with OS documentation and collaborates with partners like IBM. "It's about rationalization," said Jethwa. > "We use the full breadth of available models but ensure every integration serves a clear destination." This pragmatic blend allows rapid iteration while safeguarding proprietary advantages—a blueprint for developers weighing build-vs-buy decisions in AI stacks.

Navigating Commercialization and Copyright

As OS's models mature, Jethwa confronts unique challenges around Crown copyright, which governs UK public sector data. Sharing models risks enabling third parties to monetize taxpayer-funded assets without reciprocity. "We must protect our data while delivering value to the UK," he noted, highlighting tensions many organizations face when open-sourcing AI trained on sensitive data. This underscores a broader industry imperative: Foundation models demand legal and ethical frameworks as robust as their technical foundations.

The Future: Authoritative AI Interfaces

Looking ahead, Jethwa envisions AI interfaces that combine conversational ease with authoritative data sourcing. Imagine querying a map: "Show me schools in this area," with the model pulling verified OS data rather than probabilistic guesses. "You want actual schools, not probable ones," he stressed. This future hinges on APIs that tether generative AI to trusted sources—a vision where foundation models act as bridges between human queries and ground-truth data.

For developers, OS's experience signals a shift toward specialized AI. In an era of commoditized LLMs, custom foundation models offer competitive edges in accuracy and efficiency, but only with disciplined resource allocation and ethical foresight. As Jethwa's team demonstrates, the real power emerges not from model size, but from strategic alignment with the problems worth solving.

Source: ZDNET, Mark Samuels