John F Wu

#astrosci

Figure 3: Discovered Low Mass Satellites. The HI mor- phology of the isolated dwarf group dm1623 overlaid on a g-band optical image of the system. The HI contours are at the levels of 7.8, 10.4, 13.0, 18.2, 23.4, 28.7, 33.9, 39.1 × 1020 atoms cm −2, and the size of the restoring beam is indicated with the filled ellipse in the lower right hand corner. The arrows indicate the six dwarf group members with the let- tering scheme adopted in this work. The newly discovered low mass satellites (e and f) are easily identified via their gas content but are only barely visible in the optical image.

arxiv.org/abs/2512.09174

New satellites in a compact dwarf group, discovered in HI, down to 10^6 M_HI/M_sun.

Very interesting to study these satellite systems surrounding LMC-like hosts!

I love when people train transformers to do math problems, because the outcomes illustrated so succinctly how transformers identify patterns and where and why they are making mistakes

Garamond, Baskerville, and Nunito semibold.

FYI this is not fully up to date; see Alberto's other response.

I'm happy to report that "Scenario 2" won't be necessary, as NASA has just communicated to us that SciX funding will continue in 2026 (albeit at a reduced level). Therefore, we will not be forcing astronomers to leave ADS, but rather develop a plan that allows a longer transition.

Presumably the frontier labs are all using some kind of sparse/linear attention, too? At least I'd imagine GDM and OAI are doing so, given their context lengths.

Gotcha, I'm inclined to agree with all that. But given determinism, the manifold hypothesis, and the approximate guarantees of stochastic gradient descent, then I'm also somewhat not so shocked that this should work. (But maybe I've been primed after watching my NNs learn astro phenomena for years)

Do you think that animal intelligence is non-deterministic? (Genuinely asking.)

In the era of LLMs, what does learning look like? What happens when are overconfident in our understanding due to AI sycophancy? Is it even useful to use LLMs for education without outsourcing our thinking?

New blog post: jwuphysics.github.io/blog/2025/12...

In the era of LLMs, what does learning look like? What happens when are overconfident in our understanding due to AI sycophancy? Is it even useful to use LLMs for education without outsourcing our thinking?

New blog post: jwuphysics.github.io/blog/2025/12...

Fig 3 from the paper, showing how the galaxy morphology helps break degeneracies in dust and age in SED fitting-like posterior inference.

Ever wished that galaxy spectral energy distribution fitting actually made use of galaxy morphology? See the ML4PS paper/poster led by Mikaeel Yunus at NeurIPS tomorrow!

Also check out the paper on arXiv: arxiv.org/abs/2512.05078!

A note on learning physics ⚛️ for starting PhD students:
At the risk of saying something completely useless because everyone has it worked out

There is no magic trick, your classmates *may* be faster than you, but they are still studying as long as they need to and you are quite capable of the same

sorry, that's you now

btw I use ____

Prompt: How likely are you to recommend Windows 10 to a friend or colleague? Response: 1/5 (radio button - not at all likely) Prompt: Please explain why you gave this score. Answer: I need you to understand that people don't have conversations where they randomly recommend operating systems to one another

Same energy

Ken Liu's The Literomancer just about broke me

Scientific poster with dark background and two black holes illustrated in the center. The paper visualizes gravitational waves, and explains how parameter estimation is performed with DINGO. The standard DINGO model and the DINGO-T1 architectures are illustrated and results are shown. For example, it is possible to reanalyze the same event with different detector configurations with DINGO-T1, illustrated bz a corner plot.

1/ 🌀 New paper alert! We introduce Dingo-T1, a flexible transformer-based deep learning model for gravitational-wave (GW) data analysis. It adapts to different detector & frequency settings, improving inference efficiency and flexibility

🚀 #AI #MachineLearning #Physics #Astronomy #AcademicSky

How do people productively use Discord for collaborations? It's pure chaos to me. Am I just old?

Polymarket @Polymarket BREAKING: OpenAl ready to roll out ads in ChatGPT responses. X.com

actually gemini is kinda good

Glad that the sun includes Limb darkening!

This might say something about the sorry state of VCs more than anything

AI + Lean proof of a version of Erdos problem #124.

NOT the full solution, which has evaded mathematicians for 30 years, but an impressive and elegant proof nonetheless.

www.erdosproblems.com/forum/thread...

Yes!

They were already out of most sizes by the time I went (around 4:30pm).

Some Lidl ugly holiday sweaters for $12.99

Aldi crewneck sweater with the all caps, colorful letters "ALDI" plastered over it

Honestly it's hard to compare because I think Lidl is actually trying to make ugly holiday sweaters.

Meanwhile Aldi is not playing around with their designs.

(My daughter was actually wearing the Aldi crewneck shown in the 2nd pic when we got groceries at Lidl today.)

Okay so now Lidl also has holiday drip? I'm intrigued

Headed to NeurIPS just for the workshops; give me a shout if you want to chat and you're interested in scientific ML/AI and interpretability!

In June 2025 Sam Altman claimed about ChatGPT that "the average query uses about 0.34 watt-hours". In March 2020 George Kamiya of the International Energy Agency estimated that "streaming a Netflix video in 2019 typically consumed 0.12-0.24kWh of electricity per hour" - that's 240 watt-hours per hour at the higher end. Assuming that higher end, a ChatGPT prompt by Sam Altman's estimate uses: 0.34 Wh / (240 Wh / 3600 seconds) = 5.1 seconds of Netflix Or double that, 10.2 seconds, if you take the lower end of the Netflix estimate instead. I'm always interested in anything that can help contextualize a number like "0.34 watt-hours" - I think this comparison to Netflix is a neat way of doing that. This is evidently not the whole story with regards to AI energy usage - training costs, data center buildout costs and the ongoing fierce competition between the providers all add up to a very significant carbon footprint for the AI industry as a whole.

Out of curiosity I decided to try and run the numbers on how much Netflix you can watch for the energy cost of a ChatGPT prompt

As far as I can tell it's between 5.1 and 10.2 seconds, depending on which end of the 2019 IEA Netflix energy usage estimate you use

simonwillison.net/2025/Nov/29/...

Why support Baltimore Beat? Because Baltimore deserves journalism that doesn't ignore us.https://support.baltimorebeat.com/page/NEIGHBORHOOD

A thread of directed acyclic graphs (DAGs) that look like record covers... because that's EXACTLY what the world needs

1. Huey Lewis and the News: link.springer.com/chapter/10.1...