Singapore's AI Testing Toolkit: An Inside Look at Project Moonshot

In a bold move to establish itself as a global leader in responsible AI development, Singapore has unveiled Project Moonshot, a groundbreaking initiative aimed at creating a comprehensive AI testing toolkit. The project, spearheaded by the Infocomm Media Development Authority (IMDA) and the AI Verify Foundation (AIVF), seeks to address critical concerns surrounding the safety, fairness, and transparency of large language models (LLMs).

Project Moonshot brings together experts from diverse fields to create a suite of tools designed to evaluate and improve AI systems. The toolkit encompasses various components, including red-teaming techniques, explainability assessments, and ethical considerations. By focusing on these key areas, Singapore AI is positioning itself at the forefront of responsible AI development. This article will explore the origins of Project Moonshot, delve into its main components, examine its global impact, and discuss potential partnerships that could shape the future of AI governance.

The Genesis of Project Moonshot
Singapore's AI Strategy

Singapore's journey towards becoming a global AI leader began with the introduction of its National AI Strategy 2.0. This comprehensive plan aims to achieve excellence and empowerment in AI for the public good, benefiting both Singapore and the world at large. A key enabler of this strategy is the creation of a trusted environment for AI, which has become the foundation for Project Moonshot's development.

The Infocomm Media Development Authority (IMDA) recognized the need to establish guardrails to manage AI risks while fostering innovation. This approach emphasizes the importance of adopting an agile, test-and-iterate methodology to address key risks in model development and use By extending the work of the AI Verify Foundation into the realm of Generative AI, IMDA has taken a proactive stance in tackling the challenges posed by Large Language Models (LLMs).

Need for Transparent and Trustworthy AI

The rapid advancement of AI technologies, particularly LLMs, has brought to light significant concerns regarding their potential negative impacts. Without proper safeguards, these models can reinforce biases and generate harmful content, leading to unintended consequences. This realization has underscored the critical need for transparent and trustworthy AI systems.

To address these challenges, Project Moonshot was conceived as one of the world's first tools to combine benchmarking, red-teaming, and testing baselines. This innovative approach allows developers to focus on what's most important to them while ensuring the safety and quality of AI models and applications.

The project provides intuitive results that unveil the quality and safety of a model or application in an easily understood manner, even for non-technical users. It employs a 5-tier scoring system, where each "exam paper" completed by the application is scored on a five-tier scale. The grade cut-offs can be determined by the author of each "exam paper," allowing for flexibility and customization.

To ensure the tool's usefulness and alignment with industry needs, IMDA and the AI Verify Foundation (AIVF) have collaborated with like-minded partners such as DataRobot, IBM, Singtel, Temasek, and Resaro. This collaborative effort has been crucial in providing design inputs and feedback for Project Moonshot, making it a truly industry-driven initiative.

The genesis of Project Moonshot represents a significant step towards establishing global testing standards for AI. By bringing together leading AI testing organizations like AI Verify Foundation and MLCommons, the project aims to build a common testing benchmark for large language models. This collaboration provides developers with clarity on what and how they need to test their applications to ensure they meet certain thresholds of safety and quality.

Key Components of the AI Testing Toolkit

Project Moonshot, developed by the AI Verify Foundation, is one of the first tools to combine benchmarking and red-teaming capabilities for evaluating Large Language Models (LLMs) and LLM applications. This innovative toolkit addresses critical aspects of AI testing, including explainability, robustness, fairness, and safety.

Explainability and Robustness Testing

The toolkit offers a range of benchmarks to measure an LLM application's performance in capability, quality, and trust & safety. These benchmarks serve as "exam questions" to test the model across various competencies, such as language and context understanding. To streamline the testing process, Moonshot helps identify and run only the most relevant tests, optimizing the evaluation procedure.

For unique use cases, users can tailor their evaluation process with custom datasets by creating their own "recipes" in Moonshot. This flexibility allows for a more comprehensive assessment of AI systems across different scenarios.

Fairness and Safety Assessments

A crucial component of Project Moonshot is its red-teaming capabilities. Red-teaming involves adversarial prompting of LLM applications to induce behavior incongruent with their design. This process is essential for identifying vulnerabilities in AI systems. Moonshot simplifies red-teaming by providing an easy-to-use interface that allows for the simultaneous probing of multiple LLM applications.

The toolkit equips users with red-teaming tools such as prompt templates, context strategies, and attack modules. These features enable thorough testing of AI models for potential biases, safety issues, and unintended behaviors.

User-Friendly Interface and Reporting

Project Moonshot aims to provide intuitive results of the quality and safety of a model or application in an easily understood manner, even for non-technical users. The toolkit employs a web-based user interface that allows users to produce HTML reports visualizing test results in easy-to-read charts. For more in-depth analysis, users can access raw test results through JSON files that log full prompt-response pairs.

To ensure the tool's usefulness and alignment with industry needs, IMDA and the AI Verify Foundation collaborated with partners such as DataRobot, IBM, Singtel, and Temasek. This collaborative effort has been crucial in providing design inputs and feedback for Project Moonshot, making it a truly industry-driven initiative.

By bringing together leading AI testing organizations like AI Verify Foundation and MLCommons, Project Moonshot aims to build a common testing benchmark for large language models. This collaboration provides developers with clarity on what and how they need to test their applications to ensure they meet certain thresholds of safety and quality.

Global Implications and Partnerships
Collaboration with MLCommons

Project Moonshot represents a significant step towards establishing global testing standards for AI. Two leading AI testing organizations, AI Verify Foundation (AIVF) and MLCommons, have joined forces to build a common safety benchmark suite for large language models. This collaboration aims to provide developers with clarity on what and how they need to test their applications to ensure they meet certain thresholds of safety and quality.

On May 29 2024, AIVF and MLCommons signed a memorandum of intent (MOI) to collaborate on building this common safety benchmark suite. Peter Mattson, MLCommons President and co-chair of the AI Safety working group, expressed pride in this partnership, stating that it would positively impact AI safety by providing a globally accepted approach to safety testing for generative AI.

MLCommons, supported by over 125 founding members and affiliates, including startups, leading companies, academics, and non-profits from around the globe, plays a crucial role in democratizing AI. The organization focuses on:

1. Developing open industry-standard benchmarks to measure quality and performance

2. Building open, large-scale, and diverse datasets to improve AI models

3. Continually measuring and improving the accuracy, safety, speed, and efficiency of AI technologies

This collaboration between AIVF and MLCommons is particularly significant as MLCommons is recognized by the US National Institute of Science and Technology (NIST) under its AI Safety Consortium.

Potential Impact on International AI Standards

The partnership between AIVF and MLCommons has the potential to shape international AI standards significantly. By combining their expertise and resources, these organizations are working towards creating a globally accepted approach to AI safety testing. This effort could lead to more consistent and reliable evaluation methods for AI systems worldwide.

In addition to the AIVF-MLCommons collaboration, other international initiatives are underway to address AI governance and safety:

4. Singapore's AI Safety Institute (SG AISI): Housed within NTU's Digital Trust Center, SG AISI will collaborate internationally with other AI Safety Institutes to advance the sciences of AI safety for national AI governance and international frameworks.

5. Digital Forum of Small States (DFOSS) AI Governance Playbook: Singapore and Rwanda are collaborating on developing an AI Governance playbook for small states through open consultations at DFOSS.

These initiatives demonstrate the growing global recognition of the need for standardized AI safety measures and governance frameworks. As Project Moonshot continues to evolve and gain international support, it has the potential to become a cornerstone in the development of global AI testing standards, fostering trust and accountability in AI systems worldwide.

Conclusion

Project Moonshot represents a significant step forward in the realm of AI governance and safety. By bringing together experts from various fields and collaborating with industry partners, Singapore has created a toolkit that has a substantial impact on how we test and evaluate AI systems. This initiative not only addresses critical concerns about AI safety and fairness but also sets the stage for global partnerships to develop common testing standards.

The implications of Project Moonshot extend far beyond Singapore's borders. As AI continues to shape our world, the need for reliable, transparent, and ethical AI systems becomes increasingly crucial. With its comprehensive approach to AI testing, Project Moonshot offers a blueprint for other nations to follow, potentially leading to more consistent and trustworthy AI development worldwide. In the end, this initiative showcases Singapore's commitment to responsible AI innovation and its potential to influence the future of AI governance on a global scale.

FAQs
What is the new toolkit launched by Singapore for testing the safety of general AI models?

Singapore has introduced an AI testing toolkit developed in collaboration with partners such as DataRobot, IBM, Singtel, and Temasek. This initiative is part of Singapore's commitment to leveraging the global open-source community to address AI risks and ensure the tool meets industry needs.

What does AI Verify refer to?

AI Verify is an AI governance testing framework and software toolkit designed to enhance transparency in AI applications across various industries, thereby fostering trust.

Reference

[1] - https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/2024/project-moonshot https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/2024/project-moonshot

[2] - https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/press-releases/2024/sg-launches-project-moonshot https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/press-releases/2024/sg-launches-project-moonshot

[3] - https://github.com/aiverify-foundation/moonshot https://github.com/aiverify-foundation/moonshot

[4] - https://cybersecurityasean.com/news-press-releases/project-moonshot-singapores-ai-toolkit-safer-large-language-models https://cybersecurityasean.com/news-press-releases/project-moonshot-singapores-ai-toolkit-safer-large-language-models

[5] - https://www.prnewswire.com/in/news-releases/singapore-launches-project-moonshot--a-generative-artificial-intelligence-testing-toolkit-to-address-llm-safety-and-security-challenges-302160481.html https://www.prnewswire.com/in/news-releases/singapore-launches-project-moonshot--a-generative-artificial-intelligence-testing-toolkit-to-address-llm-safety-and-security-challenges-302160481.html

[6] - https://mlcommons.org/ https://mlcommons.org/

Author Bio

Cecil Su

Co-opted Committee Member

AiSP

Cecil leads the Cyber Security & DFIR unit for BDO Advisory. He currently leads various engagement teams on diversified advisory, security testing and incident response projects across vertical industries. His current area of focus is in IT Security research and security testing. Prior to this, he was the Practice Lead for Trustwave SpiderLabs SEA, which is the advanced security services and research team that specialises in penetration testing, incident response/forensics, education services, and security research.