Daily Note (10th August 2026)
Another low-science day, with most of the day spent in admin stuff or in trying to get quotes for desktops (why is ram so expensive!?1).
Had a fun meeting with Gene Magnier, and talked about the All the Catalogs™ project. Gene brought up the fact that the PanSTARRs DR2 dataset should contain a lot more than the objects it currently has, and is worth looking into. Turns out, the way the PanSTARRS data was ingested (along with GALEX/ALLWISE/DES) was basically only objects that had crossmatches in the existing set of object positions from the surveys at that time (PS1 DR2 has over 10 billion detections over 30,000 deg). Have started running a new script over the larger area now, we'll see how much end up in the final catalogs. Here's roughly how much is in the current catalogs, and what we could get to:
| Survey | Full Catalog size | Currently ingested | VizieR mirror |
|---|---|---|---|
| AllWISE | 747M sources, all-sky, W1-W4 | 3.99M (0.5%) | II/328 |
| GALEX GUVcat_AIS | 82M sources, 24,790 deg | 1.09M (1.3%) | II/335 |
| DES DR2 | 691M sources, 5000 deg grizY | 1.96M (0.3%) | II/371 |
| PanSTARRS DR2 | 1868M sources, , grizY | 322k (0.02%) | II/389 |
We also talked about courses I'm thinking of designing/teaching over the next few years, and about evaluations and goals from academic courses in the era of LLMs, and touched on how collaboratively designing an AI-usage policy before starting a class might be the way to go, similar to a session 0 before a D&D campaign2. For the Astrostats/ML course I'm thinking of, there's three different directions I can take it in - (i) a pure astrostats course, focusing on things like what a distribution is, why different distributions arise naturally in different contexts, hypothesis testing, moments and expectation values, aleatoric and epistemic uncertainties and systematics, inference and sampling, all with an emphasis on astronomy (ii) an astrostats+ML course, focusing on a stats primer, followed by how ML (starting with kernel methods + regression like Josh S. likes) is essentially applied statistics, and how understanding this allows us to build better (or at least more targeted) algorithms, architectures and loss functions, and then jumping straight into real-world problems with a focus on interpretability, (iii) an astrostats+LLM course, focusing on combining elements from my stats course and Vivi's LLMs in research course, again thinking of LLMs as sampling engines in latent space and how that helps develop methods to better harness transformer-based architectures in research workflows (and better understand their limitations and failure modes). The last one needs a bit more work though3 and might best be saved for a more advanced grad seminar (kinda like a special topics course), but is also probably the most timely. Stuff to think about.
Highlights: Ms. Coach had another great show today (opening for Omnivores) with a decent crowd. Fun lines include 'did anyone lose a grandpa', and 'am i talking to anybody really?' in response to 'are you talking to me?'