After 7 years as a geospatial data scientist, I quit my job. I'm going all in to tell you stories hidden in satellite data I've worked on cocoa and coffee supply chain monitoring, near real time flood maps with radar satellites that see through clouds, mapped trees in 3D to prevent wildfires and built pipelines that analyse crop yields at the scale of the US But I've always kept going back to telling stories I've done 70+ interviews of data scientists, CEOs, engineers, policy makers and artists all trying to better understand the impact of data, maps & satellites images in our lives I've made videos seen by hundreds of thousands of people on how satellite & maps work. So now I'm going all in I believe the worlds need more people telling stories, of people; that statistics and data changes the world, but people and stories are what resonate. A few weeks ago I already put a video out with @OurRadiantEarth about millions of people being miscounted in Argentina. One of the largest press outlets picked the story up, and now there's a conversation about this problem, that's what I want to go do This is probably the scariest thing I've ever done in my life, but I'm also really excited about it If you're a company doing cool, impactful stuff with satellite images, geospatial & data and can't get the world to pay attention, let's work together. I have some cool stories coming up and some collaborations in the pipes. So, if that sounds cool, I'll see you around on the internets

My lesson from August/September was harness engineering; coding agents @CommandCodeAI in particular

Introducing Atlas: The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D. Model the world, move the camera, and simulate space & time.

TimesFM-3 is here, marking a significant step forward in our time-series foundation model research by natively enabling zero-shot multivariate forecasting in a single forward pass. ✨ TimesFM-3 ranks first among all pretrained foundation models across major public benchmarks: Gift-Eval, FEV-Bench, and Time. --> Native Multivariate Support: Unlike previous versions limited to univariate forecasting, TimesFM-3 jointly predicts multiple coevolving time series, capturing cross-series dependencies without requiring task-specific fine-tuning. --> Flexible Covariates: The model natively incorporates past covariates (like historical foot traffic) and past-future dynamic covariates (like planned promotions or weather forecasts) using a clever "lookahead" token construction strategy. --> Alternating Attention Architecture: Built on top of a decoder-only transformer, it uses causal temporal attention and full variate attention in a 2D grid to seamlessly blend temporal patterns with cross-series relationships. --> Non-Autoregressive Decode: Using Contiguous Patch Masking, TimesFM-3 generates the entire forecasting horizon and its 9 probabilistic quantiles (10th to 90th percentile) in a single forward pass, eliminating iterative latency and error compounding.