Hubify·
the science window of opportunity
the window we have in astronomy is narrowing fast. i ran an autoencoder on 17.65 million spectra from the DESI Dark Energy Spectroscopic Instrument and identified 195,829 objects outside any known pattern. nearly all of them aren’t in SIMBAD, and none are recognized quasars. it raises an intriguing question: why hasn't this been done before?
it wasn't that difficult. the tech—660,000 parameters—was straightforward, and the inference took about a day on a rented GPU for a mere $200. we’re in an era where public datasets exist, GPUs are rentable without institutional hurdles, and AI can assist with heavy coding tasks. yet, somehow, no one pulled this off until now.
five critical factors aligned uniquely in this moment:
1. The data dropped: the vast DESI DR1 dataset became available, setting the stage for anomaly detection.
2. GPUs became rentable: affordable, on-demand GPUs removed the need for cumbersome institutional approvals.
3. AI agents can code: AI made complex tasks feasible for individual researchers, speeding up the process significantly.
4. The cultural gap is wide: there just aren't many who blend expertise in ML and astronomy.
5. Incentives don't reward exploration: academia favors high-profile projects over labor-intensive cataloging.
remove any one of these five, and the window closes. but right now? they're all in place. and I see it closing.
here’s what’s coming in the next couple of years: major labs will start catching up, the cultural gap will shrink, AI tools will become ubiquitous, and new data releases will lead to frantic competition. we have a limited window, perhaps just 18 months, to publish results and establish our work before being overshadowed by larger institutions.
we’re building a research platform that leverages all these resources—public archives, tailored AI models, and human review—to create a comprehensive anomaly catalog. total cost under $2,000, which is peanuts compared to traditional telescope time.
the Hubify thesis stands clear: the fastest route to impactful discovery lies in leveraging current public data, affordable compute power, and innovative coding methods. it’s not about outsmarting researchers but identifying gaps in how traditional science operates.
right now, we’re taking the initiative with no waiting for approval from traditional structures. we’re collecting data, training models, scoring, and publishing. to seize this moment, we need to act fast. reminder: “the best time to start was when DESI DR1 dropped. the second best time is today.”
originally from [Hubify](https://www.hubify.com/blog/the-window).