Launch HN: Silurian (YC S24) – Simulate the Earth
A YC-backed startup is building a generative transformer model to forecast global weather, claiming performance that rivals or exceeds traditional physics-based simulations like GraphCast and Aurora at far lower computational cost. Commenters probe whether data-driven models can truly replace explicit physics, how they handle chaos, climate change and rare events such as hurricanes, floods, and volcanic eruptions, and what new applications this could unlock in sectors like energy, insurance, agriculture, and defense. The thread also raises questions about business viability in a field dominated by government and big-tech efforts, model transparency and benchmarking, and the limits of extending this approach to harder domains such as earthquakes or agent-based economic simulations.
Model performance and comparisons
- Startup claims its Generative Forecasting Transformer (GFT) outperforms Microsoft’s Aurora, which in turn outperforms GraphCast, based on internal and published metrics.
- A Google-affiliated commenter notes that another model (NeuralGCM) is actually top of the WeatherBench leaderboard and includes explicit physics, indicating the landscape is competitive and evolving.
- Users request public benchmark results (e.g., WeatherBench scores) and clearer quantitative comparisons.
Physics-based vs data-driven forecasting
- Several commenters frame this as another instance of “The Bitter Lesson”: scaling general-purpose methods beating hand-crafted physics.
- Others are skeptical: physics-based ensemble models have well-defined skill metrics and handle non-stationarity (e.g., climate change) more transparently.
- Clarification: modern ML weather models are trained on 4D “movies” of reanalysis fields and learn to predict the next frame; errors still grow chaotically over time, as with physics models.
Scope, use cases, and business viability
- Claimed plan: start with weather, then integrate into domains dependent on it—energy, agriculture, logistics, defense, insurance.
- Commenters see clear commercial value in better forecasts for power trading, grid line ratings (regulation-driven), flood and wildfire risk, surf and sports forecasting, and hurricane tracking.
- Others question defensibility given heavy government and big-tech investment and lack of proprietary data.
Other phenomena: earthquakes, flooding, grid
- Multiple requests to “do earthquakes next”; replies note limited data, strong chaos, and likely short useful lead times (minutes–hours).
- Flooding and wildfire called out as especially high-value but hard: need much finer spatial resolution and high-quality terrain / land-use data.
- Grid-level simulation seen as politically and data-access constrained; focusing on adequacy and renewables correlation is advised.
Visualization and product clarity
- Significant side-thread about the site looking like a clone of nullschool.
- Clarified: it uses an open-source version of that visualization, with attribution; the startup’s contribution is the forecast data/model.
- Some argue the UI should better highlight what’s new to non-experts.
Climate and long-term simulation
- Plan to extend from weather to climate with distributional, not pointwise, forecasting.
- A related Google effort suggests training on short-term weather can yield realistic multi-year climate behavior, supporting this direction.