Year-round
Research Internship
Aristo
Posted on 7/19/2023
The Allen Institute for AI
201-500 employees
Non-profit AI research institute and tools
Compensation Overview
$86.5k - $123.6k/mo
Seattle, WA, USA
Bachelor's, Master's, PhD
 AI & Machine Learning (1) |
|---|
 Data & Analytics (1) |
|---|
Get referred to The Allen Institute for AI
See people who can refer or advise you
- - Pursuing a BS, MS, or PhD in computer science or a related field
- - Interest in cutting-edge artificial intelligence research
- - Reading the scientific literature to investigate and understand AI research
- - Executing AI research under guidance of the mentor, which may include building AI systems or system components, developing datasets, and/or conducting experiments
- - Meeting with the mentor to discuss research progress and ideas
AI2 advances artificial intelligence through nonprofit research and open engineering that benefits society. It builds open projects—such as AllenNLP, Aristo, Semantic Scholar and others—that researchers use to do language understanding, reasoning, science Q&A, and literature search. Unlike many AI firms that sell products, AI2 is funded by grants and donations and releases its tools and datasets openly to the public. Its goal is to improve AI reasoning and language understanding and apply these advances to real-world problems in education, science, and policy for the public good.
Company Size
201-500
Company Stage
N/A
Total Funding
N/A
Headquarters
Seattle, Washington
Founded
2014
Get referred to The Allen Institute for AI
See people who can refer or advise you
Simplify's Take
What believers are saying
- August 12, 2026 OlmoEarth embeddings with Hugging Face widen adoption in geospatial AI.
- August 2026 Shippy expands Skylight into free maritime analysis for governments and nonprofits.
- May 2026 OLMoEarth v1.1 and June 2026 MolmoMotion show sustained release velocity.
What critics are saying
- Microsoft hired Ali Farhadi, Hanna Hajishirzi, and Ranjay Krishna in March 2026.
- FFST shifted 2026 funding to proposals, pressuring frontier-model work before 2027.
- Open-source leadership loss to Microsoft weakens AI2; talent drain can hollow core research.
What makes The Allen Institute for AI unique
- AI2 open-sources models, datasets, and evals, unlike closed frontier labs.
- OLMoEarth and MolmoBot combine domain-specific training with reproducible infrastructure.
- Shippy and Semantic Scholar turn research models into operational public-interest tools.
Help us improve and share your feedback! Did you find this helpful?
Benefits
Medical
401(k)
Visa Sponsorship
Time off
Catered meals & free snacks
Training & development
Tuition reimbursement
Remote & hybrid work options
Company News
OlmoEarth embeddings launch for geospatial data analysis. Allen Institute for AI (AI2) and Hugging Face have introduced OlmoEarth embeddings, enabling researchers to export customized geospatial data representations for diverse analytical tasks. This development streamlines the application of foundation models to satellite imagery and environmental data. Published August 12, 2026 Leveraging the power of large language models for specialized domains like earth observation has been a significant challenge. The recent introduction of OlmoEarth embeddings represents a pivotal step in bridging this gap, offering a novel approach to extracting meaningful, contextualized representations from satellite imagery. This initiative by the Allen Institute for AI (AI2), in collaboration with Hugging Face, aims to empower researchers and developers with tools to delve deeper into geospatial analysis, fostering advancements across environmental science, urban planning, and climate monitoring. It marks a significant evolution in how foundation models can be tailored and applied to complex visual data streams. Custom feature extraction for Earth observation. OlmoEarth embeddings provide a robust mechanism for generating high-quality vector representations of specific regions or phenomena captured in satellite imagery. Unlike traditional methods that might rely on hand-engineered features or general-purpose computer vision models, OlmoEarth is built upon the OlmoEarth foundation model, which has been pre-trained on a massive dataset of multi-spectral satellite images. This specialized training allows the model to capture nuanced spatial and temporal patterns inherent in Earth observation data, producing embeddings that are more semantically rich and contextually relevant. The key innovation lies in the ability for users to customize these embeddings, essentially guiding the model to focus on particular aspects of the imagery relevant to their specific analytical goals. Researchers can utilize the OlmoEarth Studio interface on Hugging Face to specify regions of interest and export these tailored embeddings. This flexibility is crucial for applications ranging from monitoring deforestation and agricultural health to tracking urban development and disaster response. By providing a streamlined workflow for obtaining custom embeddings, the platform significantly lowers the barrier to entry for advanced geospatial analysis, enabling a broader community of scientists and practitioners to leverage cutting-edge AI for environmental insights. This functionality extends beyond simple feature extraction, allowing for a deeper interrogation of complex land cover changes and environmental dynamics. Technical underpinnings and accessibility. At its core, OlmoEarth leverages a transformer-based architecture adapted for multi-spectral satellite imagery. The foundation model, OlmoEarth-v1, was trained on petabytes of publicly available satellite data, including imagery from Sentinel-2, Landsat, and other sources, covering diverse geographies and time periods. This extensive pre-training imbues the model with a comprehensive understanding of Earth's surface characteristics and their evolution. The exported embeddings are high-dimensional vectors that encapsulate these learned features, making them suitable for a variety of downstream machine learning tasks such as classification, clustering, anomaly detection, and similarity search. Accessibility is a cornerstone of this release. By integrating with Hugging Face, a widely adopted platform for AI model sharing and deployment, OlmoEarth embeddings are readily available to a global community. The OlmoEarth Studio provides a user-friendly interface for generating and downloading these embeddings, abstracting away the underlying computational complexity. Furthermore, the availability of the base OlmoEarth model on Hugging Face allows advanced users to fine-tune the model for even more specialized tasks or integrate it into custom pipelines. This combination of powerful technology and accessible tooling democratizes access to state-of-the-art geospatial AI. Implications for geospatial AI and research. The introduction of OlmoEarth embeddings carries significant implications for the field of geospatial artificial intelligence. Historically, the application of deep learning to satellite imagery has required substantial expertise in both remote sensing and machine learning, often involving complex data preprocessing and model development. By offering pre-trained, customizable embeddings, OlmoEarth streamlines this process, enabling researchers to focus on their analytical questions rather than the intricacies of model training. This development is expected to accelerate discoveries in areas such as climate change impact assessment, biodiversity monitoring, sustainable land management, and humanitarian aid. For businesses, it opens new avenues for leveraging satellite data in sectors like agriculture, insurance, and resource management, leading to more informed decision-making and operational efficiencies. The ability to export embeddings for specific tasks means that AI solutions can be tailored with unprecedented precision, moving beyond generic models to highly specialized tools for understanding its planet. Why it matters. OlmoEarth embeddings signify a crucial advancement in making sophisticated AI capabilities for Earth observation more accessible and applicable. By enabling customized feature extraction from satellite imagery, this tool empowers a broader range of users to conduct detailed geospatial analyses, fostering innovation in environmental monitoring, climate research, and various industrial applications. It represents a concrete step towards democratizing advanced AI for understanding and addressing global challenges related to its planet. The integration with Hugging Face further amplifies its potential impact, ensuring wide adoption and collaborative development within the AI and remote sensing communities.
MolmoMotion: advancing language-guided 3D motion forecasting. Allen Institute for AI introduces MolmoMotion, a new model capable of 3D human motion forecasting guided by natural language instructions, enhancing interaction in virtual environments and robotics. Published June 17, 2026 The Allen Institute for AI (AI2) has unveiled MolmoMotion, a novel approach to 3D human motion forecasting. This development represents a significant step forward in the ability of AI systems to predict complex human movements, offering a more nuanced and context-aware understanding of how people interact within digital spaces. By integrating natural language instructions directly into the prediction process, MolmoMotion aims to bridge the gap between abstract human commands and the precise physical actions required to fulfill them. The challenge of motion forecasting. Predicting human motion in a 3D environment is a multifaceted challenge. Traditional methods often rely heavily on observed trajectories and statistical models, which can be effective for short-term predictions or repetitive actions. However, these methods typically lack the capacity to understand the intent or high-level goals behind a person's movement. The variability of human behavior, the multitude of possible actions, and the continuous nature of movement in three-dimensional space make it difficult for AI to generate realistic and purposeful predictions. This limitation restricts the sophistication of interactions possible in applications such as virtual reality, robotics, and assistive technologies. Introducing language-guided prediction. MolmoMotion addresses these limitations by incorporating natural language as a guiding input. Users can provide textual descriptions of desired movements, such as "walk towards the door" or "pick up the cup," and the model generates a plausible 3D motion sequence consistent with both the initial observed motion and the linguistic command. This approach moves beyond simple trajectory prediction to encompass a deeper semantic understanding of actions. The model leverages large language models (LLMs) to interpret the linguistic input, translating high-level goals into actionable constraints and preferences for motion synthesis. This allows for more intuitive and flexible control over generated human movements, enhancing realism and utility. Technical underpinnings and implementation. At its core, MolmoMotion employs a diffusion-based generative model. Diffusion models have shown remarkable success in generating complex data, including images and audio, by iteratively refining a noisy input into a coherent output. In MolmoMotion's case, the model learns to denoise a random sequence of poses into a realistic human motion, with the denoising process continually steered by both the observed initial body poses and the encoded linguistic guidance. The architecture integrates modules that process 3D skeletal data and text embeddings, allowing for cross-modal interaction that effectively fuses the visual and linguistic information. This multi-modal integration is crucial for generating motions that are not only physically accurate but also semantically aligned with the given instructions. Implications for diverse industries. The capabilities of MolmoMotion hold significant implications across various industries. For robotics, it could enable robots to understand and execute more complex human-like tasks in unstructured environments, moving beyond pre-programmed routines to respond dynamically to verbal cues. In virtual and augmented reality, it could facilitate more natural and expressive avatar movements, enhancing immersion and user experience. Game developers could utilize this technology to generate more diverse and context-aware character animations with less manual effort. Furthermore, in areas like human-computer interaction, it could lead to more intuitive interfaces where users can describe desired actions rather than relying solely on graphical input. Why it matters. MolmoMotion represents a notable advancement in the field of AI-driven human motion synthesis. By effectively integrating natural language understanding with 3D motion generation, it opens up new avenues for more intelligent and interactive AI systems. This development pushes the boundaries of how machines can interpret and respond to human intent, paving the way for applications that are more intuitive, versatile, and deeply integrated into its daily lives. The ability to command complex physical actions through natural language marks a significant step towards more human-centric AI.
Ai2's Shippy AI: ocean data for maritime analysis. 2h ago · 0:00 listen · Source: GeekWire Summary. The Allen Institute for AI, or Ai2, has launched an AI agent named Shippy. This agent helps maritime analysts answer questions about ocean activities, like illegal fishing or vessels that have disappeared. Shippy operates on Skylight's live vessel-tracking and satellite data. Every answer it provides links back to original records for verification. Skylight, a free ocean-monitoring platform, is used by over 300 organizations in about 70 countries. It flags suspicious behavior by combining satellite data with commercial imagery and tracking feeds. Ai2 has made the computer-vision models behind this project open-source. Shippy will also be free for governments, fisheries bodies, and eligible nonprofits. It currently has limited access but will expand to more users. Shippy focuses only on maritime questions and does not make legal judgments. It also declines defense-related requests and avoids guessing when data is insufficient. This tool aims to assist human decision-makers in addressing critical ocean issues. This is an AI-generated audio summary. Always check the original source for complete reporting.
Allen Institute introduces enhanced OLMoEarth V1.1 model suite. The Allen Institute for AI has unveiled OLMoEarth v1.1, an updated collection of open-source models designed to advance Earth system science research with improved efficiency. Published May 19, 2026 The Allen Institute for AI (AI2) recently announced the release of OLMoEarth v1.1, a significant update to its open-source family of models tailored for Earth system science. This new iteration builds upon the foundational OLMoEarth project, aiming to provide researchers with more efficient and capable tools for understanding and predicting complex environmental phenomena. Advancing Earth system modeling. OLMoEarth is a specialized collection of language models (LLMs) developed with a particular focus on Earth system data. Unlike general-purpose LLMs, OLMoEarth models are pre-trained on vast datasets relevant to climate, weather, and environmental processes. This targeted pre-training allows them to better comprehend and process the unique terminology and intricate relationships found within Earth science literature and data. The initial release of OLMoEarth laid the groundwork for leveraging large language models in scientific discovery. By creating models specifically attuned to Earth system science, AI2 sought to democratize access to advanced AI capabilities for environmental research, enabling scientists to explore complex datasets and generate hypotheses more effectively. Key enhancements in V1.1. Version 1.1 introduces several crucial improvements over its predecessor. A primary focus of this update is enhanced efficiency in both training and inference. The new models are designed to require fewer computational resources, making them more accessible to a broader range of researchers, including those with limited access to high-performance computing infrastructure. This efficiency gain does not come at the expense of performance; rather, it aims to deliver comparable or improved accuracy with a smaller footprint. Another significant enhancement involves refined architectural choices and optimization strategies. These improvements contribute to faster processing times and potentially more robust model outputs when dealing with diverse Earth system data. The models in the v1.1 suite continue to support a wide array of tasks relevant to environmental science, from understanding scientific texts to potentially aiding in the interpretation of sensor data and climate simulations. Implications for researchers and applications. The improved efficiency of OLMoEarth v1.1 has direct benefits for researchers in various Earth science disciplines. Scientists can now conduct experiments and analyze data more quickly, accelerating the pace of discovery. The reduced computational overhead also lowers the barrier to entry for smaller research groups and individual scientists who may not have access to extensive computing clusters. Potential applications span a wide spectrum, including climate change research, meteorological forecasting, environmental monitoring, and hydrological modeling. For instance, researchers could utilize these models to extract insights from vast archives of scientific papers, identify trends in environmental data, or even develop predictive frameworks for natural disasters. The open-source nature of OLMoEarth further encourages collaboration and innovation within the scientific community, allowing for continuous refinement and expansion of its capabilities. Context against previous releases and the future outlook. Prior to OLMoEarth, general-purpose LLMs often struggled to fully grasp the nuances of highly specialized scientific domains like Earth system science due to their broad training data. OLMoEarth addressed this by curating and pre-training on domain-specific datasets. Version 1.1 refines this approach by optimizing the underlying model architecture for better performance within these specialized contexts. The release of OLMoEarth v1.1 signals AI2's ongoing commitment to fostering open science and providing specialized AI tools for complex scientific challenges. The emphasis on efficiency reflects a growing awareness within the AI community of the need for sustainable and accessible AI solutions. Future iterations may see further specialization, integration with other scientific tools, and continued efforts to enhance both performance and accessibility for the global Earth science community. Why it matters. OLMoEarth v1.1 represents a significant step forward in making advanced AI more practical and accessible for Earth system science. Its enhanced efficiency and specialized training empower researchers to tackle complex environmental challenges with greater speed and fewer resources. This contributes to a more collaborative and efficient global effort in understanding and mitigating critical environmental issues.
Ai2 releases open-source visual AI agent that can take control of web browsers. Allen Institute for AI, a prominent Seattle-based nonprofit research organization working on advancing artificial intelligence models and systems, today launched a new open-source AI agent that can take control of web browsers on a user's behalf and automate tasks. Web agents represent the next step of what is called vision-language models, which move large language models from understanding images and text through captions and answering questions to taking actions. Today, the company announced MolmoWeb, built on the Molmo 2 multimodal model family, available in two sizes: 4 billion and 8 billion parameters. It will be available for free, along with the weights, training data and code (coming soon), as well as the evaluation tools used to build it. It's designed to be self-hosted locally or in the cloud. To take actions, AI agents must interpret instructions from humans and what can be seen. That includes a set of tasks written in conversational language and a live web page. The AI model observes the web page through a series of screenshots and then interacts directly with it via the interface by predicting what will happen when it takes actions such as clicking, typing characters into text fields, or scrolling up and down. The company said that, unlike other open-weight web agents, MolmoWeb was trained without compressing a proprietary vision-based agent. The data comes from synthetically generated text-only accessibility agents and human usage of actual web browsing activities. The agent interface supports navigating URLs, clicking on screen coordinates, typing text into fields, scrolling through pages, opening and switching browser tabs and sending a message back to the user. All of these actions work directly within the browser, with click locations represented as coordinates in pixels when executed. Ai2 said the agent was designed this way so that it won't break if the underlying webpage code or HTML changes on the fly. For example, some web pages obfuscate, or hide, how they operate under the hood in order to protect themselves. Some of them use specialized JavaScript engines in order to detect bots, stop ad blockers, display animations, track users and more. Using the underlying code can also consume tens of thousands of tokens, the essential currency of AI operations. Visual interfaces also behave much more closely to how humans interact with web interfaces: What a person sees is how they will approach the page. It means it's easier to debug why the model did what it did. In spite of the compact size, Ai2 said MolmoWeb achieves state-of-the-art results among open-weight web agents. When tested on popular evaluation suites, the 8B model scored 78.2% on WebVoyager, 42.3% on DeepShop, and 49.5% on TailBench. It outperformed leading open-weight models such as Fara-7B across all four benchmarks. The company said that MolmoWeb can also outperform agents built on GPT-4 that rely on annotated and structured page data. Ai2 said that's a particularly important result given that those models can "see" deeply into the very code of the webpage and also have substantially larger parameter sizes - by colossal orders of magnitude. like comparing a mouse to an elephant. More access to open-weight browser AI agents will also help researchers and hobbyists develop their own web-using automations. Closed-source large language model providers have already dipped their toes into the market with agentic web browsers capable of automating web tasks, including OpenAI Group PBC and Perplexity AI Inc., with ChatGPT Atlas and Perplexity Comet, respectively. Image: allen Institute for AI. A message from John Furrier, co-founder of SiliconANGLE: Support its mission to keep content open and free by engaging with theCUBE community. Join theCUBE's Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities. * 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more * 11.4k+ theCUBE alumni - Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network. About SiliconANGLE Media SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-time audience engagement. As the parent company of SiliconANGLE, theCUBE Network, theCUBE Research, CUBE365, theCUBE AI and theCUBE SuperStudios - with flagship locations in Silicon Valley and the New York Stock Exchange - SiliconANGLE Media operates at the intersection of media, technology and AI. Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Its new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.