Foundation Models

A type of AI model trained on vast quantities of data and adaptable for use on a wide range of tasks. Foundation models can be used as a base for building more specific AI models and represent a paradigm shift in AI development.

Key Characteristics:

• Trained on massive datasets
• General purpose and adaptable to multiple tasks
• Can be deployed in many complex ecosystems
• Relatively small number of organizations developing them
• Varying approaches: closed control vs. open-source distribution

Regulatory Challenges:

• Lifecycle Accountability: Complex to allocate responsibility across supply chains
• Cross-Sector Impact: Don't fall neatly within single regulator remits
• Opacity: Can be challenging to identify accountability for outcomes
• Scale Effects: Improvements or defects can quickly affect all adapted products
• Open Source Complexity: Different governance challenges for open vs. closed models

UK Regulatory Approach:

• Context-specific regulation focusing on use rather than technology
• Central risk function monitoring foundation model developments
• Horizon scanning for emerging opportunities and risks
• Monitoring and evaluation with technical expertise
• Potential compute-based governance metrics for larger models
• Focus on supply chain risk management and assurance techniques

Examples: Large language models (LLMs) like ChatGPT that can write software, generate stories, provide medical advice, and perform numerous other tasks depending on deployment context.

Policy Considerations: The UK recognizes foundation models' transformative potential while taking a measured approach to avoid stifling innovation. The uk-foundation-model-taskforce supports building UK capability while the regulatory framework ensures appropriate governance as these technologies develop.