Maurizio: What are the challenges you face in working with artificial intelligence and large language models when it comes to working at the intersection of global, country, and local levels?
What should policy makers be thinking about when it comes to the regulations of these technologies?
Could you also provide a short explanation of large language models?
Jaron: A large language model is a series of mathematics and computation that looks to understand how humans use language. It looks at enormous repositories of text like Wikipedia and Google News, and learn, through a series of tasks, how humans put sentences together.
A language model has billions of different parameters. If you think about how a human speaks, we make jokes. We use expressions of love and hate, and of scientific and technical sophistication. A language model must think about how to learn all of that. And that's done by sequencing different data points that come together.
I see two main challenges working with across global, country, and regional levels.
One thing I think computer and data scientists often forget to communicate is that large-language models must be trained in order to be useful for specific tasks. In my work, I spend a lot of time working with experts to come up with concepts and datasets that can help train a model, identifying things like interventions, outcomes, populations, and other data points that are important for evidence-based decision making.
One challenge we face is the availability of data. We already know through 50 to 100 years of work that we have more published data coming from North America and Europe than from Africa and Southeast Asia. This implicitly biases a model in decision-making ability, especially if it is not properly trained. We know that context matters when looking for opportunities to work together to solve problems.
Languages are another challenge. Most models are trained on English-language materials. It's challenging to find models trained to work with non-English languages, especially local or regional languages, and this inherently biases what can be done with data models. The good news is that this issue is being taken seriously and I see concerted efforts to ensure that, where possible, models have looked at different languages and can tell you about the differences in what you can expect to achieve for any language.
Maurizio: You were a co-director of Ceres2030, a unique research project that presented a real evidence-based roadmap calling to double food-related aid to end hunger by 2030. Hesat2030 is the next phase. Could you tell us more about what Hesat2030 is and where it aims to take us next?
Jaron: Hesat2030: a global roadmap to end hunger sustainably and nutritiously is a new partnership driving change in global agrifood systems through better evidence, advocacy and innovation. We are led by the Food and Agriculture Organization of the United Nations (FAO), Shamba Centre for Food and Climate, and the University of Notre Dame and in partnership with CABI, Havos.Ai, Global Alliance for Improved Nutrition (GAIN); Global Donor Platform for Rural Development (GDPRD); International Food Policy Research Institute (IFPRI); and the University of Chicago. Hesat2030 was launched during the UN Food Systems Summit +2 Stocktaking Moment.
Hesat2030 will build on the evidence and costing of Ceres2030 by integrating additional outcomes, such as women’s empowerment, climate adaptation, and nutrition, to better understand how to improve the quality and quantity of official development assistance (ODA). We will be updating some of our global modelling figures, our evidence, and increasing outreach with the global community. We will continue to leverage advances in artificial intelligence and economic modelling. A global community of stakeholders is very important to the work of Hesat2030. The Donor Platform is a key stakeholder and community partner.
Ceres2030 was organized around a global modelling effort and a publication with Nature Research. With Hesat2030, we are publishing new findings more frequently while at the same time we work to update global modelling and a series of comprehensive recommendations by 2025.
We know that governments, funders, and stakeholder groups (including the Donor Platform’s SDG 2 Roadmap Working Group) need information faster. We want to increase efficiency in the global knowledge value chain, especially around science-policy information. Through our collaborations with groups like the Zero Hunger Coalition, Hesat2030 is producing a series of country-specific cost roadmaps for Madagascar, Zambia, and other countries. And the Juno Evidence Alliance will be releasing the State of the Field: Research in Agrifood Systems, which will use AI to look at more than 6 million papers in agrifood systems to better understand where we have high-quality evidence as well as how to identify complementary innovations from science faster.
The Platform's SDG2 Roadmap Working Group is a unique platform because we're able to receive feedback from donors on an ongoing basis. This better aligns our agendas with the messages donors are working on and ensures the evidence and cost modelling is time responsive to their concerns.
Maurizio: What gives you hope in 2023 and beyond?
Jaron: Communities are coming together to learn from each other. There is a willingness to work differently and to rely more on data and evidence. We're seeing incredible opportunities and commitments from groups working across the private sector, and groups convening farmer organizations and academics. Groups like the Juno Evidence Alliance and the Zero Hunger Coalition give me hope because they are bringing together a broad community of partners to ensure SDG2 is achieved by 2030.