NOTE: The following blog is Part 1 of a five-part series that will be published through end of year discussing areas of the AI field that the Atlas AI team is passionate about and working to advance in 2024.
The leadership team of Atlas AI has been discussing our roadmap for 2024 and the research questions we’re most excited to tackle in a rapidly advancing field of artificial intelligence. One area of R&D that I’m particularly interested in is how recent advances in Large Language Models (LLMs) offer us a new lens on human development.
Atlas AI was founded with a mission to help organizations navigate the complexities of our rapidly changing world. Our approach has always been dynamic, utilizing various data sources from satellite imagery to mobile phone activity. However, the significant advances with LLMs in 2023 has opened a new avenue for us to advance our mission, with unexpectedly powerful results .
Although trained on text to perform “next-token” predictions, LLMs have shown a remarkable ability to do all sorts of things other than just generate text, like solve math problems, play board games, or forecast time series. Given this, it’s not too much of a stretch to think that LLMs might be helpful in describing local conditions around the world.
Last year, a student at Stanford, Rohin Manvi, spearheaded research into how LLMs can map socioeconomic conditions. Things like how many people live in a location, how wealthy are these households, and how much do houses cost? The findings were encouraging. For instance, by fine-tuning GPT3.5 with a limited dataset, we achieved an r2 of 0.73 in predicting household asset wealth in Africa. This not only surpasses some results obtained from our core satellite imagery-based models but also hints at the untapped potential of LLMs in geospatial intelligence.
The performance also seems to scale with both the complexity of the model and the number of samples used to fine-tune, suggesting that LLM performance could be pushed even higher. How exactly these early models are performing so well is still under investigation, just like the performance of LLMs on most tasks not directly related to text completion. But it seems that there is enough geographic information embedded in online data such as Wikipedia articles or blog posts that LLMs learn a reasonable mapping of the world. Some other recent work has come to similar conclusions, even claiming to identify ‘space’ neurons that embed geospatial knowledge.
What these results mean for generating accurate maps is still unclear. For example, it’s not yet obvious whether this information is mostly redundant to what we already know from images, or if combining the two will improve overall performance. In the coming year, we will delve into several key research questions:
- How can we better harness LLMs to predict urban growth and its socio-economic impacts?
- What role can these models play in enhancing our existing maps to forecast future societal trends?
- How might they complement our satellite imagery to offer a more comprehensive view of global changes?
As an academic co-founder, my enthusiasm lies not just in Atlas AI's potential but in the broader implications of our work in advancing the field of geospatial intelligence. Deeper insights into the complex tapestry of global societies will enable us to better anticipate and respond to the needs of a rapidly changing world. The integration of LLMs into our platform will be a step towards a more nuanced and comprehensive understanding of the world, one that empowers organizations to make informed decisions in an ever-evolving global landscape.
David Lobell is a Co-Founder of Atlas AI, the Benjamin M. Page Professor at Stanford University in the Department of Earth System Science and the Gloria and Richard Kushel Director of the Center on Food Security and the Environment. He is also the William Wrigley Senior Fellow at the Stanford Woods Institute for the Environment, and a senior fellow at the Freeman Spogli Institute for International Studies (FSI) and the Stanford Institute for Economic Policy and Research (SIEPR).
Lobell's research focuses on agriculture and food security, specifically on generating and using unique datasets to study rural areas throughout the world. His work has been recognized with various awards, including the Macelwane Medal from the American Geophysical Union (2010), a Macarthur Fellowship (2013), the National Academy of Sciences Prize in Food and Agriculture Sciences (2022) and election to the National Academy of Sciences (2023).
