ChatGPT class maps 1.5M space objects

Space discovery usually sounds like a telescope story. New mirror, new launch, new image. But this one is really a data story — and a pretty wild one. Matteo Paz, one of the students featured in OpenAI’s new ChatGPT Futures class, built an AI pipeline that reprocessed NASA’s NEOWISE archive and flagged about 1.5 million previously unknown variable objects hiding in old infrared observations. (openai.com) ### What was actually found? Not 1.5 million new planets or asteroids. The haul is mostly variable sources — objects whose brightness changes over time. That bucket can include quasars, binary stars, supernova candidates, pulsating stars, and other things that flicker, fade, or flare in ways astronomers care about. The important part is that these sources were sitting in the data without being systematically pulled out and labeled before. (neowise.ipac.caltech.edu) ### Why was NEOWISE the right place to look? Because NEOWISE was huge. The mission collected infrared observations across the whole sky for more than a decade, creating a single-exposure database with nearly 200 billion detections. That is the kind of archive humans cannot inspect source by source, and older pipelines were not built to chase every faint, irregular signal buried in it. Basically, t(neowise.ipac.caltech.edu)ding. (arxiv.org) ### What did Paz build? He built a model called VARnet. The paper describes it as a Fourier- and wavelet-based system for fast time-series analysis — meaning it looks for meaningful changes in brightness over time, not just static snapshots. In validation tests on real infrared variable sources, the model hit an F1 score of 0.91, and the paper says it could process a source in under 53 microseconds on suitable GPU hardware. That spee(arxiv.org) long, an archive this big becomes basically unworkable. (arxiv.org) ### Why does “variable” matter so much? Because brightness changes are often the clue, not the noise. A quasar can pulse because matter is falling into a black hole. A binary star can dim because one star passes in front of the other. A supernova can show up as a sudden brightening and fade. If you think of a sky survey as a giant photo album, variability turns it into a movie — and the movie is where a lot of the physics lives. (ne([arxiv.org)ltech.edu/news/neowise20250411/)) ### Where does ChatGPT fit in? The news hook here is OpenAI featuring Paz in its inaugural ChatGPT Futures class, a program announced on May 6, 2026 that gives 26 student builders a $10,000 grant and access to frontier models. OpenAI frames the class as a showcase for students using ChatGPT to build real projects, and Paz’s space-mapping work is one of the standout examples attached to that launch. (openai.com) ### Is this just a nice student project? No — it already crossed into real research. Caltech said Paz published a single-author, peer-reviewed paper tied to the work, and the project won the top $250,000 prize in the 2025 Regeneron Science Talent Search. So the point is not “teen uses AI, internet claps.” The point is that the pipeline was strong enough to matter to working astronomy. (neowise.ipa([openai.com) the bigger lesson? Astronomy is entering an era where the limiting factor is often not collecting data but extracting signal from archives that are already overflowing. That is true for NEOWISE, and it will be even more true for newer surveys. Tools like this do not replace telescopes — they turn neglected back catalogs into active discovery machines. (neowise.ipac.caltech.edu)r is 1.5 million. The deeper story is that a student-built AI system turned old sky data into a fresh map of changing objects. That is why this matters — not because AI magically invented new space, but because it made a giant, messy archive legible enough for science to use. (neowise.ipac.caltech.edu)