weathering

They compare two new zero-shot forecasting papers and why treating time like language could change prediction. The conversation covers model architectures, synthetic data, and tradeoffs between bespoke and foundation approaches. They explore industrial workflows, probabilistic forecasts, and implications for atmospheric and fluid modeling.

<p>Bias has a pretty well-known definition in the world of AI/ML programming. But today’s paper asks us to expand that definition and consider how cultural, organizational, and human forces can intersect with development. The authors of “Identifying and Categorizing Bias in AI/ML for Earth Sciences” argue that, beyond considering bias and before mitigating bias, developers ought to be able to identify bias in all its modes &amp; forms. Their taxonomy of bias is helpful, and inspired us to create an actionable check list of our own.</p> <hr> <h2>Paper</h2> <p><strong><a href="https://journals.ametsoc.org/view/journals/bams/105/3/BAMS-D-23-0196.1.xml">Identifying and Categorizing Bias in AI/ML for Earth Sciences</a></strong>, McGovern et al</p> <hr> <h2>Chapters</h2> <ul> <li><strong>00:00:03</strong> - Intro</li> <li><strong>00:02:39</strong> - Abstract</li> <li><strong>00:03:33</strong> - Discussion of framing</li> <li><strong>00:09:03</strong> - A taxonomy of bias 09:03</li> <li><strong>00:47:52</strong> - Our proposed check list for mitigating bias</li> <li><strong>01:08:14</strong> - Reading recs</li> </ul> <hr> <p><strong>A check list for mitigating bias (in developing AI or for really any kind of technological development)</strong></p> <ul> <li>Use the language of “bias”, acknowledging both social and technical bias</li> <li>Define your specific user and use case early</li> <li>Establish a clear baseline for comparison</li> <li>Ensure diversity of perspectives on your team</li> <li>Practice reflexivity (question your assumptions, and continue that line of questioning throughout the development process.)</li> <li>Examine your incentives vs. end-user incentives</li> <li>Prioritize transparency (in your decision-making and in your incentives &amp; goals.)</li> </ul> <p><strong>Recommended reading</strong></p> <ul> <li><a href="https://backup.ai2es.org/">AI2ES Newsletter</a></li> <li><em>Groundhog Day</em> (film <em>AND</em> musical theater adaptation)</li> <li><a href="https://sociology.sas.upenn.edu/sites/default/files/Weber-Science-as-a-Vocation.pdf">“Science as a Vocation</a>” by Max Weber</li> <li><em>Book of the New Sun</em> by Gene Wolfe (four-book series)</li> <li><em>Infinite Powers</em> by Steven Strogatz</li> </ul>

<p>While staring into uncertainty might sound abyssal and frightening, ensemble models have proven (to us, at least) that this isn’t the case. In today’s episode, we’re exploring two papers with different approaches to ensemble forecasting. This &quot;choir&quot; approach to weather prediction is one that embraces chaos rather than striving to chart a single line of truth on a graph by generating dozens or even hundreds of slightly different predictions that together map a full range of possible outcomes.</p> <p>We’re unpacking FGN, the latest variant of AIFS, and their differing approaches to the same challenge: <strong>how can we create a confident forecast of an assuredly uncertain future.</strong></p> <hr> <h2>Featured Papers</h2> <ul> <li><em><a href="https://arxiv.org/abs/2506.10772">Skillful joint probabilistic weather forecasting from marginals</a></em>, Alet et al.</li> <li><em><a href="https://arxiv.org/html/2412.15832v1">AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score</a></em>, Lang et al.</li> </ul> <hr> <h2>Chapters</h2> <ul> <li><strong>00:00</strong> — Intro</li> <li><strong>01:38</strong> — Abstract (Meet the paper!)</li> <li><strong>10:17</strong> — Weather report &amp; reading list</li> <li><strong>19:37</strong> — A lil history (Newton, Blake, Goethe, &amp; voices of dissent)</li> <li><strong>28:43</strong> — Paper time!</li> </ul> <hr> <h2>Further Reading</h2> <ul> <li><em>The Solace of Open Spaces</em> by Gretel Ehrlich (especially “On Water”)</li> <li><em>North Woods</em> by Daniel Mason</li> <li><em>Frederick</em> by Leo Lionni</li> <li><em>Hyperion</em> and <em>The Fall of Hyperion</em> by Dan Simmons</li> <li><em><a href="https://www.tate.org.uk/art/artworks/blake-newton-n05058">Newton</a></em> by William Blake</li> <li><em><a href="https://www.youtube.com/watch?v=iv-5mZ_9CPY">But how do AI images and videos actually work?</a></em> by WelchLabs and 3Blue1Brown (video)</li> <li><em>Chaos: Making a New Science</em> by James Gleick</li> <li><a href="https://www.upstream.tech/posts/from-chaos-to-clarity-seasonal-forecasts-for-confident-risk-management">“From chaos to clarity: Seasonal forecasts for confident risk management”</a> with Marshall, Alden, and Phil Butcher from the HydroForecast team</li> </ul>

<p>Machine learning dominates the conversation, but what will happen to the centuries-old physical equations that built our understanding of the atmosphere? How are the two approaches at odds? How might they coexist?</p> <p>Today’s paper, <em>NeuralGCM</em>, sheds light on how physics-based and AI approaches might be a powerful pairing for weather forecasting. For the first truly hybrid model we’ve discussed, it’s only fitting that we’ve also taken a hybrid approach in this conversation. So join us for hybrid models, data compression, <em>Dragon Ball Z</em>, and the strange future of primitive equations.</p> <hr> <h2>Featured paper</h2> <p><strong><a href="https://www.nature.com/articles/s41586-024-07744-y">Neural general circulation models for weather and climate</a></strong></p> <hr> <h2>Chapters</h2> <ul> <li>00:00 Intro</li> <li>01:31 Weather report &amp; books</li> <li>18:01 Paper time</li> <li>25:26 A theory of compression</li> <li>48:07 Closing thoughts: Resolution, interpretability, &amp; the future of primitive equations</li> </ul> <hr> <h2>Recommended reading</h2> <ul> <li><em>Chaos: Making a New Science</em> — <strong>James Gleick</strong></li> <li><em>Landmarks</em> — <strong>Robert McFarlane</strong></li> <li><em>Dandelion Wine</em> — <strong>Ray Bradbury</strong></li> <li><a href="https://www.google.com/url?q=https://www.websters1913.com/words/Wind"><em>Webster’s 1913 Dictionary</em></a></li> <li><a href="http://www.incompleteideas.net/IncIdeas/BitterLesson.html">“<strong>The Bitter Lesson</strong>”</a> — <strong>Rich Sutton</strong></li> <li><a href="https://arxiv.org/pdf/2101.05186">MC-LSTM: Mass-Conserving LSTM</a></li> <li><em>Charisma and Disenchantment</em> — <strong>Max Weber</strong></li> </ul>

<p>Where does the end begin? Where does the end end? And what the HELL is happening in the middle?</p> <p>These are some of the questions we tackle in this episode as we explore Aardvark Weather, an end-to-end forecasting system. End-to-end models are super cool, and likely to be the future of weather forecasting—but they also revitalize old questions about transparency, end-user trust, and the fidelity &amp; availability of raw observations.</p> <p>Join us as we unpack what an end-to-end model is, the Aardvark Weather approach, and how these models at-large might shape the landscape of forecasting (And books. Always books.)</p> <h2>Featured paper</h2> <p><a href="https://www.nature.com/articles/s41586-025-08897-0"><strong>End-to-end data-driven weather prediction</strong></a></p> <h2>Additional reading</h2> <ul> <li><em>The Forest Unseen</em> by David George Haskell</li> <li><em>“<a href="https://harpers.org/archive/2024/07/the-gods-of-logic-benjamin-labatut-ai/">The Gods of Logic</a>”</em> by Benjamín Labatut</li> <li><em>Moon of the Crusted Snow</em> by Waubgeshig Rice</li> <li><em>“<a href="http://www.incompleteideas.net/IncIdeas/BitterLesson.html">The Bitter Lesson</a>”</em> by Rich Sutton</li> </ul>

<p>How far into the future can we actually predict the weather? Typically, this question is rhetorical, a slight (or smite⚡) against meteorologists everywhere after someone’s wedding gets rained out. But in this inaugural episode of <em>Weathering</em>, we’re asking in earnest—what is the horizon for accurate weather forecasts?</p> <p>Here, we look at a paper from the University of Washington (link below) that suggests the limit might be more than double the long-held belief of fourteen days. Thus, too, challenging the foundational theory of chaos—“the butterfly effect”—that has informed how we think and forecast weather since it was coined in the 1970s. We examine the methods that the researchers use and engage in some armchair philosophy: What does chaos mean if our foundational example of chaos—weather—is actually predictable? Is there a difference between chaos and predictability? And how might knowing next month’s weather change our relationship to the environment and weather itself?</p> <p>We can’t promise you answers, we didn’t even articulate the questions that well, but we can promise to add to your never-ending TBR.</p> <h2>Featured paper</h2> <p><strong>Testing the Limit of Atmospheric Predictability With a Machine Learning Weather Model</strong><br> <a href="https://arxiv.org/pdf/2504.20238">https://arxiv.org/pdf/2504.20238</a></p> <h2>Further reading</h2> <ul> <li> <p><strong>Science and Method</strong> by Henri Poincaré<br> <em>Chaos theory before it was chaos theory.</em></p> </li> <li> <p><strong>Isaac's Storm</strong> by Erik Larson<br> <em>The 1900 Galveston hurricane through the eyes of meteorologist Isaac Cline, revealing the human and scientific failures that led to the deadliest natural disaster in U.S. history.</em></p> </li> <li> <p><strong>Immeasurable Weather: Settler Colonialism and the American Weather Enterprise</strong> by Sara J. Grossman<br> <em>Weather science in the U.S. is entangled with settler colonialism (omg no way).</em></p> </li> <li> <p><strong>Strange as This Weather Has Been</strong> by Ann Pancake<br> <em>A novel set during the coal boom in southern Appalachia, centered on mountaintop removal mining and catastrophic flooding. You will weep.</em></p> </li> <li> <p><strong>“A Sound of Thunder”</strong> by Ray Bradbury<br> <em>A short story in which stepping on a butterfly in the past alters the future. Published BEFORE the term was coined, and that’s Bradbury for ya.</em></p> </li> </ul>