The Allen Institute for Artificial Intelligence

The Allen institute for Artificial Intelligence ("ai2") is an independent, scientific research institute dedicated to building breakthrough applications in the field of Artificial Intelligence ("ai"), empowering the ai research community, and benefiting society as a whole through high-impact ai research, engineering, and educational resources in the public interest.scientific research in the public interest - in 2024 alone, ai2 published 230 research papers in key scientific venues, including 10 that received awards at top conferences. Ai2 utilizes various channels to make its research and engineering accessible to the public, including multiple open platforms and datasets for ai research, as well as interactive demos of new systems. Ai2 publishes a monthly newsletter with over 384,000 subscribers; its youtube channel with over 300 videos that have received over 63,000 views on topics including demos, research presentations, and lectures from visiting research scientists; its twitter/x and linkedin channels that have over 120,000 followers combined; and its github account that offers over 500 public code repositories.educational programming in the public interest - ai2 operates multiple term employment programs for students at various stages of their careers, designed to provide valuable research experience and an opportunity to advance their specialized area of the ai field through the development of open resources and tools. These programs strengthen the pipeline of experienced, capable ai researchers both entering and starting to teach at universities. Ai2 runs a year-round internship program for undergraduate students pursuing their bachelor's degrees, a one-to-three-year predoctoral young investigator program for those preparing to apply to graduate school, and a one-to-three-year young investigator program for postgraduate researchers seeking experience before transitioning to faculty roles. In 2024, ai2 hosted 40 interns through its undergraduate program and supported 13 predoctoral young investigators and 15 young investigators through its longer-term research tracks.program overview - in 2024, ai2's research and educational activities were organized across the following project teams: allennlp, aristo, mosaic, prior, semantic scholar, climate modeling, earthranger, skylight, and wildlands. Each of these ai2 projects maintains independent short-term and long-term research goals and performance objectives that contribute to ai2's mission.allennlp - allennlp focuses on enhancing natural language processing ("nlp") systems and large language model ("llm") performance and accountability, providing open software and resources for researchers to create and streamline nlp experiments and llm development and tuning. The allennlp team envisions language-centered ai that equitably serves humanity, working to improve the performance and accountability of nlp systems and advance scientific methodologies for evaluating and understanding those systems. In 2024, several members of this team contributed to the development of the second version of ai2's open language model ("olmo"), an open-source language model that offers access to the full pipeline of model development from data to training code to model weights to evaluation tools, the first truly open model of its kind. In 2024, the team published 108 top-tier papers, two of which won best paper awards at the 2024 conference of the association for computational linguistics.aristo - aristo focuses on building systems that can successfully interact with users, demonstrate a deep understanding of the world, and solve complex problems. In 2024, the team published 15 papers at top conferences during the year, one of which won an award for outstanding paper in the qa track at the conference on empirical methods in natural language processing (emnlp).mosaic - mosaic focused on defining and developing methods of imbuing common sense, ethics, and social norms into interactive ai systems. The team's work laid important foundations for the development of ai that is safe and ethical to interact with across a variety of applications. In 2024, mosaic published 23 papers, one of which won a spotlight award at the 2024 international conference on learning representations (iclr). Mosaic sunsetted in the fall of 2024.prior - prior seeks to advance computer vision to create ai systems that see, explore, learn, and reason about the world. In 2024, the prior team published 33 papers in top venues. The prior team also released molmo, a family of fully open state-of-the-art multimodal ai models for the community to build upon. The most powerful molmo model closes the gap between open and proprietary systems across a wide range of academic benchmarks, as well as human evaluations, and the smaller models in the family outperform models ten times their size.semantic scholar - semantic scholar seeks to accelerate scientific research and enhance access to scientific knowledge via ai tools. The semantic scholar platform is a free, ai-powered research tool for scientific literature that provides better ways for researchers to discover new knowledge and key insights. In 2024, semantic scholar published 31 papers about their research into human-computer interaction and scientific paper understanding.earthranger - earthranger is a free platform that aids protected area managers, ecologists, and biologists in making more informed operational decisions for wildlife conservation. Earthranger supports ecological management, promotes sustainable coexistence through proactive planning, monitoring, and reaction coordination, and advances domain awareness for more effective and efficient security operations. Earthranger is used by 563 conservation initiatives in over 74 countries across six continents.skylight - skylight is a maritime monitoring and analysis software that helps improve maritime transparency and combat illegal fishing by surfacing suspicious events and providing actionable Intelligence to protect oceans. By applying advanced analytics, including state-of-the-art ai, to a combination of vessel-identifying information and vessel behavior history with real-time vessel tracking data and remote imaging, skylight enables analysts to identify potential illegal, unreported, and unregulated fishing and alert authorities who can take enforcement and compliance action and contribute to creating fisheries transparency. Used in over 60 countries worldwide, skylight accelerates the way organizations make operational decisions critical to protecting our ocean.climate modeling - climate modeling focuses on creating faster, more efficient, modernized open-source models to advance important research in global climate model projections. The climate modeling team develops open-source software so the broader climate modeling community can easily adopt their advances, as well as partners with a leading climate modeling center, noaa's geophysical fluid dynamics laboratory ("gfdl"), to ensure their work builds on gfdl's valuable experience. In 2024, the team released an updated version of the ai2 climate emulator (ace), which is the world's first climate emulator accurate for climate variability and climate change. The climate modeling team published 20 scientific papers over the course of the year. Wildlands - wildlands focuses on applying machine learning to support wildland management. The wildlands team works on identifying ground-level fuels (flammable forest floor detritus) that are not measurable with current satellite technology. Surface fuels are a neglected research area in wildfire research and management, and the wildlands team is applying computer vision to greatly amplify the impact of current low-tech methods. In 2024, the team extended the feature set and usage of a free app for ios and android that eases the work of practitioners collecting data in the field, provides tools to help summarize the data collected, and constructs a dataset suitable for training ai models tasked with assisting practitioners in the future.