Patrick Beukema's Environmental AI team uses deep learning with satellite and other geospatial data for conservation, sustainability, wildfire management, and climate work.
2
AI research should be treated like an engineering workflow from the start, with testing, continuous integration, and early use by real users.
3
Small teaching models, user feedback, and open releases help improve expensive computer vision systems without retraining the largest model after every correction.
Summary
Patrick Beukema leads the Environmental AI team at AI2, where the work focuses on conservation and sustainability rather than commercial model development. He describes projects across climate modeling, wildlife protection, wildfire management, and ocean monitoring. Much of the work uses custom deep learning models on satellite data, with different models for different sensors and tasks. In ocean monitoring, false positives can trigger expensive satellite tasking or Coast Guard trips, so the team gives precision more weight than recall. Beukema treats research as an engineering discipline. Models enter real-world workflows early, and teams use testing and user feedback to improve them. A lightweight model can learn from simple thumbs-up and thumbs-down feedback on top of a large, expensive feature extractor. He also argues for open source releases that include data, training details, weights, and models. The conversation stays grounded in practical environmental work and is skeptical of focusing on AGI at the expense of present-day public benefit.
Environmental AI covers several distinct public-interest problems
Beukema describes Environmental AI as a young field with several areas of work at AI2. Climate modeling teams simulate how the planet and atmosphere change. Earth Ranger focuses on monitoring wildlife and helping address poaching. Another team works on wildfire resources and fire management, an issue Beukema connects to the smoke affecting Seattle. Skylight is a maritime intelligence platform focused on ocean health. Each group uses artificial intelligence to advance its mission, with conservation and sustainability as the main focus rather than building models for commercial return.
The team builds custom deep learning systems around the data and task
AI2's Environmental AI work is entirely based on deep learning, with close collaboration from computer vision and natural language research teams. Satellite work involves imagery from NASA, the European Space Agency, and commercial sources. Beukema says some environmental problems have received less research attention than large language models, so the team sometimes creates new neural network architectures for tasks that are not well studied. His design approach starts by studying how expert humans perform a task, then using that process to shape a system that may combine several neural networks.
Research should enter an engineering workflow from its first experiments
Beukema argues that research and development works best when it is treated as an engineering discipline. His team adapts software engineering practices such as continuous integration and continuous delivery to machine learning research workflows. There is no separate production stage later in the process. Experiments begin with integrated testing, and models are placed in real-world settings as soon as possible. That gives the team feedback from users and lets it spend iteration time on changes that affect the actual outcome rather than polishing a system in isolation.
Real users expose failures that papers and lab scores can miss
Beukema points to poor replication and reproducibility in neuroscience and machine learning research. He cites an attempt to reproduce around 60 reinforcement learning papers submitted to NeurIPS, where the results were poor. Environmental models also face out-of-sample and out-of-distribution conditions that may not appear in published evaluations. He rejects a model-development process where researchers arrive from an institution, deploy a system, and leave. Stakeholders need to be part of development, especially when conservation technology affects communities with unequal resources.
Open releases give communities ways to inspect and improve the work
AI2 releases its environmental work openly, including data, training information, weights, and models. Beukema says this lets people find bugs, suggest improvements, and judge what the team actually did. Openness also gives the team signals about whether its work is useful. Those signals include outside groups using the technology, showing interest, funding related work, or reporting that the models perform well against systems built with better data. These measures do not reduce impact to company revenue, which is not how the nonprofit operates.
Satellite validation makes precision an operational requirement
Skylight uses low-resolution satellite data to identify possible vessels in places where they should not be. When the model finds a possible vessel, the team may task a commercial satellite for a higher-resolution image. A false positive can also lead a Coast Guard team to spend tens of thousands of dollars traveling to investigate something that is actually construction activity, such as a wind farm. Because the cost falls on resource-constrained coastal governments, Beukema says the team weights precision more heavily than recall and needs performance close to perfect for these uses.
A small teaching model can learn from user corrections
The largest computer vision models cannot be retrained after every user response because that would be too expensive. Beukema describes placing a lightweight convolutional neural network on top of a large feature extractor, such as a Swin Transformer. Users can give a simple thumbs-up or thumbs-down when the system displays what it saw. The small model learns from those supervised corrections and can be retrained as part of continuous integration and delivery. It has roughly 100,000 parameters, can run on a CPU-based GitHub runner, and helps prevent repeated mistakes such as confusing new wind turbines with vessels.
Beukema prefers bespoke models for each data source or use case when they produce better results. Satellite sensors can capture very different information, so a model designed for RGB imagery may not work well for nighttime light data. AI2 uses public and commercial sources, including sensors that detect lights across the planet several times a day. One open-source vessel-detection model uses nighttime light imagery to find vessels on the high seas. Foundational geospatial computer vision models can provide useful backbones, but Beukema's standard is task performance, not whether one model handles every satellite.
"The artificial intelligence that we're building needs to benefit Humanity right it needs to like uplift Humanity that is our Central Focus."Patrick Beukema07:58
Who should watch
You are building machine learning systems for conservation, public agencies, or other settings where a false positive has a direct operational cost.
Your research team struggles to move experiments into users' hands or to reproduce results outside the original evaluation setup.
You want a practical example of combining large vision models, small correction models, satellite data, and user feedback.
