Less Is More: Towards Frugal, Ethical AI

Luxembourg’s Institute of Science and Technology (LIST) is committed to advancing AI research and innovation backed by a portfolio of AI projects worth €55 million for the 2023-2028 period. Francesco Ferrero, head of the Human Centered AI, Data, and Software (HANDS) Unit at LIST and leader of LIST’s Flagship Initiative on Artificial Intelligence, discusses the need for frugal AI, addressing bias and hallucinations and positioning Luxembourg as an AI leader.

Source : siliconluxembourg.lu
Date de publication : 06/12/2024

A frugal AI approach

Recently, the “bigger-is-better” approach in AI, especially with Large Language Models (LLMs) has gained a lot of traction. However, Francesco remains cautious about this trend, pointing out its significant costs: “It is rumoured that training GPT-4 has cost $100m in electricity alone,” which is “creating a barrier to entry.” This approach is “clearly not sustainable, from an energy and more generally resources point of view” and is further compounded by regulatory requirements. “The cost of training and using new models is growing exponentially, but the benefit is not,” he says.

Focusing on scale can overshadow smaller, problem-specific AI applications that could be impactful in fields such as healthcare and climate. LIST promotes “frugal AI,” which embraces smaller, more efficient models and aligns with the “less is more” philosophy. Francesco emphasises the ethical risks of a few companies controlling AI development: “The fact that a very small number of companies is controlling such powerful AI technologies raises questions.” As a public research institution, LIST advocates for “open-source AI” to ensure transparency and mitigate risks.

Reducing AI project failure rate

AI projects have a high failure rate, with more than 80% not reaching deployment—”twice the failure rate of standard IT projects.” Francesco explains that “a lot of uncertainty exists because companies are afraid to fail,” often due to poor data, lack of expertise, or unclear problem definitions. LIST’s mission is to “de-risk technological innovation” by providing expertise to ensure projects succeed. “We help design projects that make sense,” ensuring essential components like high-quality data and technical skills are in place.

Proof-of-concept (POC) projects are key: “POCs are very important because then companies can decide to deploy and invest with confidence.” LIST also offers co-funding opportunities through national and European grants, making AI projects more accessible. Francesco highlights LIST’s AI sandbox, which provides a “safe, controlled environment” where companies can experiment with AI models and benchmark results, significantly reducing the risk of failure.

Reducing bias and hallucinations

Bias and hallucinations are critical challenges in AI. LIST has developed an open-source AI leaderboard which evaluates 22 large language models (LLMs) against seven key biases, including racism, sexism, and political bias, among others. Francesco explains that “we’ve worked with the Universitat Oberta de Catalunya to integrate LangBiTe,” a tool that helps developers systematically address bias before deploying AI.

Hallucinations – AI-generated answers that contain false information – are also a problem. To mitigate these, LIST combines generative AI with classic AI (i.e., rule-based systems) to improve reliability, especially in chatbot applications. Then, for a given question, we let the user decide whether it should be answered by a deterministic and hallucination-free rule-based system or by a GenAI component. This approach can “reduce hallucinations drastically, even to zero” in controlled environments.

LIST’s AI Sandbox

LIST’s AI sandbox is a critical tool for both developers and users, providing a “safe, controlled environment” to test models and minimise risks. While still in development, it will allow developers to benchmark models against custom metrics and design tests tailored to specific business or regulatory requirements. The sandbox will also enable companies to assess the impact of their solutions before full-scale implementation, reducing the project failure rates.

Low-code and AI

With a growing shortage of developers, low-code and no-code platforms are the future of software development. Platforms like BESSER are designed to enable non-technical users to develop AI-based applications, accelerating the time-to-market for complex solutions. At the end of 2023, low-code and digital process automation market was estimated at $13.2bn (21% growth over past years). BESSER aims to enhance competitiveness by reducing the time-to-market for AI-driven products and services, while also simplifying the integration of advanced features into those solutions. BESSER empowers businesses to adopt AI without needing large teams of developers, helping Luxembourg stay at the forefront of digital innovation.

The AI Flagship Initiative and National Strategy

LIST’s AI Flagship Initiative consolidates its AI efforts into a strategic plan to position Luxembourg as a leader in AI research, development, and innovation. “Almost every team uses AI for their job,” Francesco notes, and the initiative aims to build a robust ecosystem aligned with Luxembourg’s national goals and European values, emphasising trustworthy, ethical AI applications.

“This is a coordination initiative,” Francesco explains, designed to make Luxembourg competitive on the global stage. “I’m very inspired by the example of Singapore,” which has successfully fostered AI innovation by running numerous small projects. Francesco stresses the need to “stimulate the creativity of people” through “a large number of small projects,” empowering individuals and driving rapid innovation.

Francesco sees “an opportunity in the new national strategy to accelerate” Luxembourg’s AI efforts, particularly through apprenticeship programs and “fast-track funding,” which is crucial for keeping up with the fast-evolving AI landscape.

Louis Juste

www.siliconluxembourg.lu/less-is-more-towards-frugal-ethical-ai/