The introduction of generative AI, powered by large language models (LLMs), has changed the landscape of artificial intelligence (AI). However, the current state of LLMs presents a significant challenge: they embody the definition of a monolithic system.
These models are frequently shrouded in secrecy, resistant to outside scrutiny, extremely difficult to modify, and only useful under specific conditions. Another significant concern is the environmental impact of training such large models. The computational resources required to develop and maintain cutting-edge LLMs are enormous, prompting some businesses to take drastic measures such as using nuclear power to maintain a competitive advantage and reconsidering their carbon-neutral/negative strategies.
The energy consumption raises serious questions about the long-term viability of current AI development practices. These models' outputs frequently require extensive safeguards and complex prompting techniques to address issues such as toxicity, bias, fakes, and hallucinations. These challenges highlight the importance of a more agile and iterative approach to AI development.
To address these issues, developers should use Agile methodologies to develop AI products. By doing so, developers can deconstruct the monolithic structure of current AI systems, allowing for greater innovation and continuous improvement in the design and development processes.
The Agile Edge: Iterative Development
Agile software development methodologies have several promising applications in creating more sustainable AI products. Iterative development is an important Agile principle that emphasizes releasing small increments of functionality on a consistent basis. This approach aligns with the need for AI systems to be constantly refined and updated as data and requirements change, rather than being deployed as static, monolithic systems.
Another important aspect of Agile methodologies is the emphasis on user feedback. This principle is especially important for AI systems because it ensures they remain aligned with user needs while avoiding unintended consequences during deployment. E-commerce and streaming platforms' recommendation systems are a great example of how user feedback is important for sustainability.
An Agile approach to developing recommendation systems would include conducting regular user surveys to gauge satisfaction with recommendations, identify pain points, and solicit suggestions for improvement. Setting up clear feedback loops to incorporate user feedback into the ongoing refinement of the recommendation system is critical for creating AI products that can be sustained and improved over time.
The Agile Edge: Adaptability
Adaptability, another pillar of Agile methodologies, is essential for developing AI systems that must operate reliably in dynamic, real-world environments. The ability to respond quickly to new information or requirements is critical to ensuring that AI systems remain relevant and useful in the long run.
Transparency and collaboration, both emphasized in Agile practices, can aid in improving accountability in how AI systems operate. This increased transparency may help alleviate some of the concerns about the secrecy surrounding current LLMs. A clear understanding and documentation of training data, as well as making the models open source, could significantly increase trust in the technology.
Integrating Agile testing and quality assurance practices, such as test-driven development and continuous integration/deployment, can help ensure that AI systems develop with quality, dependability, and robustness. These techniques are essential for developing AI products that can be trusted and relied on in real-world settings.
Agile techniques can also help to identify and address ethical concerns about AI development. Regular reviews of the societal impact of AI systems, utilizing Agile testing and delivery practices, can be integrated into the development process, ensuring that ethical concerns remain at the forefront of AI innovation.
By incorporating Agile methodologies into the development of LLMs and AI products, we can start to address many of the problems that the current monolithic approach causes. Agile methodologies' adaptability, user-centricity, and emphasis on continuous improvement provide a solid foundation for creating more sustainable and responsible artificial intelligence.
As part of Georgetown University’s Master of Professional Studies in Project Management, we teach how Agile methodology applies to real-world problems such as this one. Using and developing AI responsibly and sustainably is an important part of being a leader in the new economy, where AI will play a significant role and affect each organization.
Learn more
