Listen to this Post
Introduction: Why Foundation Models Matter in the AI Revolution
As artificial intelligence reshapes nearly every industry, one question has become increasingly urgent for organizations of all sizes: Should you build your own AI model or buy one off the shelf? The emergence of foundation models — large, general-purpose machine learning systems — offers a promising solution for businesses looking to harness AI efficiently and at scale.
At the forefront of this transformation is Manish Jethwa, CTO of Ordnance Survey (OS), the UK’s national mapping agency. OS has been leveraging AI and machine learning long before generative AI went mainstream, and now it’s refining these systems to unlock even greater value from its decades of geospatial data. In a recent conversation with undercode, Jethwa shared five critical lessons for building AI models that are not only functional but also cost-effective, adaptable, and future-proof.
Summary: 5 Critical Tips for Building AI Foundation Models
- Build Models with a Real Use Case in Mind
Jethwa emphasizes that foundation models must be grounded in purpose. OS uses its rich archive of high-precision mapping data to train models from scratch. These models extract environmental features and serve as a reusable base for analyzing various aspects — from biodiversity to roof materials — through fine-tuning rather than retraining from zero.
2. Train Purposefully to Reduce Cost and Energy Waste
The training process can be incredibly resource-intensive. To avoid unnecessary costs, OS takes a focused approach, starting with small models trained on hundreds of examples, and gradually scaling to hundreds of thousands. This phased strategy ensures resource efficiency while maintaining high accuracy, outperforming even major tech providers’ models in specific domains.
3. Leverage Existing LLMs for Fine-Tuning
OS
4. Explore Responsible Commercialization
Though OS sees potential in sharing its models externally, Crown copyright laws limit what can be distributed. Jethwa stresses the importance of safeguarding public assets while delivering value to UK taxpayers. Any commercialization effort must strike a careful balance between innovation and protection.
5. Keep Innovating for Long-Term Success
The vision for AI at OS is interactive and intuitive. Jethwa envisions future interfaces where users can ask the AI detailed questions about maps — like school locations — and receive authoritative, non-probabilistic answers. By combining internal and external trusted data sources, OS aims to deliver AI systems that answer real-world questions with pinpoint accuracy.
What Undercode Say:
The Real Blueprint Behind OS’s AI Success
The Ordnance Survey story is a prime example of smart, domain-specific AI adoption. While Silicon Valley giants chase general-purpose AI dominance, OS is quietly leading a precision-first revolution using geospatial data — a highly specialized yet incredibly valuable asset.
- Use Case First, Hype Later: Too many businesses jump into AI without a clear plan. OS does the opposite. It builds only what’s needed, rooted in a practical application — geospatial feature extraction. This focus keeps resources aligned with outcomes.
-
Smaller, Better, Smarter: The beauty of OS’s approach lies in disproving the “bigger is always better” myth. By tailoring models to specific domains, even smaller datasets outperform massive commercial systems. This is a critical takeaway for enterprises with limited compute power but high-value niche data.
-
The Hybrid AI Stack: OS is not afraid to mix custom models with tools from tech giants. It smartly integrates Microsoft’s Azure ML and open-source Python tools. This hybrid approach maximizes flexibility and keeps the organization from being locked into one ecosystem.
-
Commercial Caution: Jethwa’s comments on Crown copyright show a refreshingly responsible stance. In an era where everyone’s rushing to open-source or sell AI models, OS stands out for prioritizing public benefit over unchecked monetization.
-
Future-Ready Design: The vision of interactive, dialog-driven geospatial AI tools is not science fiction — it’s the logical evolution of human-AI collaboration. OS’s emphasis on authoritative sourcing reminds us that hallucinations and “close-enough” answers just won’t cut it in mission-critical applications.
In essence, OS is demonstrating how specialized AI ecosystems can be more powerful than generic LLMs when properly designed, fine-tuned, and responsibly governed. Their roadmap could serve as a playbook for other government bodies, NGOs, and data-rich enterprises navigating AI deployment in sensitive sectors.
🔍 Fact Checker Results
✅ OS is indeed the UK’s national mapping agency and maintains Crown copyright on its data.
✅ Jethwa’s quotes and OS’s practices align with recent undercode reporting and official OS documentation.
✅ OS uses a mix of proprietary models and external AI tools, including Microsoft Azure-based ML solutions.
📊 Prediction
Within the next 3 to 5 years, foundation models built by public agencies like OS will become standard tools in civic infrastructure. Expect governments to adopt hybrid AI stacks — blending internal data with commercial LLMs — to improve everything from urban planning to disaster response. These models will increasingly power conversational GIS platforms, where citizens and officials alike can engage with geospatial data in natural language. As legal and ethical frameworks around data use mature, more Crown-protected models will be licensed or partially open-sourced, fueling innovation across the private sector while still preserving public interest.
References:
Reported By: www.zdnet.com
Extra Source Hub:
https://www.quora.com/topic/Technology
Wikipedia
OpenAi & Undercode AI
Image Source:
Unsplash
Undercode AI DI v2
