Introduction
Last week marked a significant milestone in my professional journey as I joined the team led by Dominic Ligot at Data Ethics Ph. In the midst of our vibrant discussions, a critical point emerged – the necessity of establishing a research program dedicated to AI alignment. It's a concept that's not just futuristic but essential in the realm of ethical AI development. This blog post is a contemplation and a blueprint of that very idea, encapsulating our early visions and the potential paths this endeavor might take.
Understanding AI Alignment
At its core, AI alignment is about developing AI systems that can understand and adhere to human values, ensuring that their actions are in harmony with human intentions. This concept becomes increasingly significant as AI systems grow more complex and autonomous. The alignment challenge is not just a technical hurdle; it's a pivotal point in ethical AI development, ensuring that the technology we create benefits humanity as a whole.
The Vision for Our Research Program
Our research program at Data Ethics Ph is in its nascent stage, but its ambitions are vast. We aim to concentrate on two primary areas: robustness and practicality. Robustness, to ensure our solutions are resilient and effective across various scenarios; and practicality, to ensure these solutions are implementable in real-world settings. This balance is crucial, especially considering our objective to find an alignment solution for AGI (Artificial General Intelligence) within the next three years. The importance of this goal cannot be understated as AGI represents a significant leap in AI capabilities, potentially surpassing human cognitive abilities.
Upscaling and Downscaling Scenarios
In envisioning the future of our research program, we are looking at two contrasting scenarios – upscaling and downscaling. The upscaled version is ambitious and involves conducting state-of-the-art (SOTA) safety experiments on AI. This approach would allow us to be at the forefront of AI safety research, setting benchmarks and contributing significantly to the global conversation on ethical AI.
In contrast, the downscaled scenario is more introspective. Here, the research would be led primarily by me (Miguel), focusing on the core aspects of AI alignment as they pertain to our immediate environment. While this approach may seem limited, it holds immense value in understanding and addressing localized challenges in AI ethics.
A major factor influencing these scenarios is resource availability. In the Philippines, accessing advanced training resources for AI is challenging. The need for high-end hardware (like MacBooks with M-series chips) or access to compute servers which are crucial for intensive AI research, poses a significant financial challenge for many researchers in the region.
Exploring the Upscaled Research Model
In our journey toward expanding the reach and impact of our research in AI alignment, overcoming resource constraints stands as a pivotal challenge. Our strategy for upscaling involves a comprehensive enhancement of the scope and intricacy of our theoretical experiments. At the heart of this endeavor is the development of datasets meticulously designed to embody human values accurately.
We are currently exploring a promising approach known as Reinforcement Learning using Layered Morphology (RLLM). This method involves training language models on datasets that encapsulate desired behaviors. While this technique is still in its refinement stage, it represents a strategic move towards establishing a robust research infrastructure. Our goal is to harness these datasets to educate state-of-the-art AI models, thereby delving into the complexities of AI alignment with unprecedented depth.
Equally critical to the development of these models is their rigorous testing. Through exhaustive evaluation, we aim to gauge the efficacy of our alignment strategies, thereby initiating a cycle of perpetual refinement. This iterative approach is crucial for the creation of AI systems that not only resonate with human values but are also flexible enough to adapt to shifting ethical paradigms.
AI Governance and Audits
As we delve deeper into AI alignment, the realms of AI governance and audits emerge as critical areas of exploration. Over the next three years, as AI technologies become more embedded in our daily lives, the importance of AI audits will grow exponentially. These audits will serve as crucial tools in assessing and ensuring the ethical deployment of AI systems.
However, effective AI governance and auditing are contingent upon established AI safety standards, which are currently lacking. This gap highlights the urgency of our research and the need to develop comprehensive safety standards that can guide the ethical development and deployment of AI technologies.
Conclusion
Our journey in establishing a robust AI alignment research program at Data Ethics Ph is just beginning. We aim to contribute significantly to the global discourse on AI ethics, balancing the dual demands of theoretical robustness and practical applicability. Recognizing the enormity of this task, we understand that it cannot be accomplished in isolation. We envision a collaborative effort, requiring the passion and intellect of numerous researchers to join us in this crucial endeavor.
In the coming months, we are planning to host a series of seminars aimed at fostering interest in AI safety and ethics. These events will serve not only as platforms for knowledge exchange but also as a litmus test to gauge market interest and identify potential collaborators who are equally committed to ensuring an AI-aligned future. Our hope is to cultivate a community of like-minded individuals and organizations who are eager to contribute to this field, whether through research, policy-making, or public advocacy.
We invite tech enthusiasts, researchers, and policymakers to engage with us in this vital conversation, helping shape a future where AI is not only advanced but aligned with the greater good of the Filipino People. The fight for an AI-aligned future is a collective one, and together, we can pave the way for ethical AI development that truly benefits society.