Emily Hunt
@emily.space
Postdoc @univie.ac.at
Researches the Milky Way & star clusters with machine learning
Founded the Astronomy feeds (@astronomy.blue)
🏳️🌈 🏳️⚧️ (she/her), Ⓥ
Website: https://emily.space
GitHub: https://github.com/emilyhunt
Researches the Milky Way & star clusters with machine learning
Founded the Astronomy feeds (@astronomy.blue)
🏳️🌈 🏳️⚧️ (she/her), Ⓥ
Website: https://emily.space
GitHub: https://github.com/emilyhunt
the codebase is no longer clean after ✨ today ✨
January 14, 2026 at 4:18 PM
the codebase is no longer clean after ✨ today ✨
Pwahaha! Amazing!
I don't even think this work doesn't have a use (someone should do it I guess), but this paragraph is really something
I don't even think this work doesn't have a use (someone should do it I guess), but this paragraph is really something
January 13, 2026 at 11:12 AM
Pwahaha! Amazing!
I don't even think this work doesn't have a use (someone should do it I guess), but this paragraph is really something
I don't even think this work doesn't have a use (someone should do it I guess), but this paragraph is really something
I need to explore more of central & eastern Europe now I'm in Austria, but this sure looks a lot more filled in than before my PhD!
January 7, 2026 at 1:36 PM
I need to explore more of central & eastern Europe now I'm in Austria, but this sure looks a lot more filled in than before my PhD!
Yes, my wedding has a website
Yes, I have also included modelling of the expected temperature on the day
Yes, I did this by fitting a Bayesian model to 45 years of temperature data
Yes, the wedding website has an easter egg with a little link to find a writeup of how I got the numbers
Yes, I have also included modelling of the expected temperature on the day
Yes, I did this by fitting a Bayesian model to 45 years of temperature data
Yes, the wedding website has an easter egg with a little link to find a writeup of how I got the numbers
December 29, 2025 at 8:47 PM
Yes, my wedding has a website
Yes, I have also included modelling of the expected temperature on the day
Yes, I did this by fitting a Bayesian model to 45 years of temperature data
Yes, the wedding website has an easter egg with a little link to find a writeup of how I got the numbers
Yes, I have also included modelling of the expected temperature on the day
Yes, I did this by fitting a Bayesian model to 45 years of temperature data
Yes, the wedding website has an easter egg with a little link to find a writeup of how I got the numbers
Before vs. after of just doing the classic 'pick a round number I like' vs. using an unbiased estimator to make your histogram:
☄️ #astrocode
☄️ #astrocode
December 19, 2025 at 9:24 AM
Before vs. after of just doing the classic 'pick a round number I like' vs. using an unbiased estimator to make your histogram:
☄️ #astrocode
☄️ #astrocode
But like wtf is that mix of black and silver, just pick a colour my guy
December 18, 2025 at 6:06 PM
But like wtf is that mix of black and silver, just pick a colour my guy
Incredible that we live in a world where the President of the United States freely gloats about the deaths of political adversaries, yet when RandomBlueskyUser952 writes that [redacted] was a [redacted]-hole they get permabanned and lose their job
December 16, 2025 at 7:01 AM
Incredible that we live in a world where the President of the United States freely gloats about the deaths of political adversaries, yet when RandomBlueskyUser952 writes that [redacted] was a [redacted]-hole they get permabanned and lose their job
New paper led by Fabian Polnitzcky & with @sratzenboeck.bsky.social + @joaoalves.bsky.social: based on the ages of stars with infrared excess in Sco-Cen, it seems planet-forming disks last around twice as long as previous estimates suggest - giving twice as long for planets to form. 🔭☄️ #exoplanets
December 9, 2025 at 10:57 AM
New paper led by Fabian Polnitzcky & with @sratzenboeck.bsky.social + @joaoalves.bsky.social: based on the ages of stars with infrared excess in Sco-Cen, it seems planet-forming disks last around twice as long as previous estimates suggest - giving twice as long for planets to form. 🔭☄️ #exoplanets
It's an absolutely gorgeous and historical building that's annoyingly difficult to photograph as it sits in a lovely bit of woodland
This was it just after construction (with fewer trees)
This was it just after construction (with fewer trees)
November 19, 2025 at 10:48 AM
It's an absolutely gorgeous and historical building that's annoyingly difficult to photograph as it sits in a lovely bit of woodland
This was it just after construction (with fewer trees)
This was it just after construction (with fewer trees)
golden hour at the @univie.ac.at observatory
November 19, 2025 at 10:46 AM
golden hour at the @univie.ac.at observatory
These quotes from this article live rent-free in my head now: Meta saying their security is "industry-leading" when they're *eight years* behind on implementing a fix for a known issue 🫠
November 19, 2025 at 8:13 AM
These quotes from this article live rent-free in my head now: Meta saying their security is "industry-leading" when they're *eight years* behind on implementing a fix for a known issue 🫠
their website seems pretty odd. For instance, that's... a lot of money to charge for some Zoom talks...
November 6, 2025 at 9:30 AM
their website seems pretty odd. For instance, that's... a lot of money to charge for some Zoom talks...
my fiancée knows me well 🥺
November 4, 2025 at 6:44 PM
my fiancée knows me well 🥺
... the study is:
- some survey data
- conducted by economics profs (not: social scientists)
- does some batshit stats to reach any conclusion at all
- finds a tiny 0.4%-2.5% work time improvement!
... that's about as effective a productivity intervention as doing yoga once a week to sleep better
- some survey data
- conducted by economics profs (not: social scientists)
- does some batshit stats to reach any conclusion at all
- finds a tiny 0.4%-2.5% work time improvement!
... that's about as effective a productivity intervention as doing yoga once a week to sleep better
October 30, 2025 at 8:52 PM
... the study is:
- some survey data
- conducted by economics profs (not: social scientists)
- does some batshit stats to reach any conclusion at all
- finds a tiny 0.4%-2.5% work time improvement!
... that's about as effective a productivity intervention as doing yoga once a week to sleep better
- some survey data
- conducted by economics profs (not: social scientists)
- does some batshit stats to reach any conclusion at all
- finds a tiny 0.4%-2.5% work time improvement!
... that's about as effective a productivity intervention as doing yoga once a week to sleep better
Yet funnily enough, this claim goes almost completely uncited except for one (1!) link to a blog post (NOT a peer-reviewed paper!) by the Federal Reserve of St Louis (?)
October 30, 2025 at 8:52 PM
Yet funnily enough, this claim goes almost completely uncited except for one (1!) link to a blog post (NOT a peer-reviewed paper!) by the Federal Reserve of St Louis (?)
This is SO unhinged
These wild assertions made me think, hmmm surely you're going to put some effort into citing those
These wild assertions made me think, hmmm surely you're going to put some effort into citing those
October 30, 2025 at 8:52 PM
This is SO unhinged
These wild assertions made me think, hmmm surely you're going to put some effort into citing those
These wild assertions made me think, hmmm surely you're going to put some effort into citing those
wtf, my little blog post about uv is third on hacker news right now
October 30, 2025 at 1:57 PM
wtf, my little blog post about uv is third on hacker news right now
I recall now a good blog post about python tooling from a couple years back. Hatch is much less well developed: (excerpt from chriswarrick.com/blog/2024/01...)
October 24, 2025 at 2:52 PM
I recall now a good blog post about python tooling from a couple years back. Hatch is much less well developed: (excerpt from chriswarrick.com/blog/2024/01...)
This was a really fun hybrid approach to try! My favourite part is that the final MCMC model fit is virtually as accurate as the *best possible* (>2000 trials) XGBoost model we tried, despite XGBoost having >1000× more trainable parameters than our little logistic function 🤯 🔭☄️ #astrocode
October 24, 2025 at 8:45 AM
This was a really fun hybrid approach to try! My favourite part is that the final MCMC model fit is virtually as accurate as the *best possible* (>2000 trials) XGBoost model we tried, despite XGBoost having >1000× more trainable parameters than our little logistic function 🤯 🔭☄️ #astrocode
So basically: using machine learning to make finding & fitting a traditional model much faster. 🔭☄️ #astrocode
Because the final model is just a simple logistic on 3 params, the model parameters are interpretable, like n_50 - the median number of stars a cluster needs to be detectable
Because the final model is just a simple logistic on 3 params, the model parameters are interpretable, like n_50 - the median number of stars a cluster needs to be detectable
October 24, 2025 at 8:40 AM
So basically: using machine learning to make finding & fitting a traditional model much faster. 🔭☄️ #astrocode
Because the final model is just a simple logistic on 3 params, the model parameters are interpretable, like n_50 - the median number of stars a cluster needs to be detectable
Because the final model is just a simple logistic on 3 params, the model parameters are interpretable, like n_50 - the median number of stars a cluster needs to be detectable
My favourite part of this paper is the model. In principle, 15+ different parameters impact the detectability of a cluster - that'd be complex! 🔭☄️ #astrocode
I got that down to just three (!!!) by using SHAP to say which were important, & PySR/symbolic regression to build intuition about models
I got that down to just three (!!!) by using SHAP to say which were important, & PySR/symbolic regression to build intuition about models
October 24, 2025 at 8:40 AM
My favourite part of this paper is the model. In principle, 15+ different parameters impact the detectability of a cluster - that'd be complex! 🔭☄️ #astrocode
I got that down to just three (!!!) by using SHAP to say which were important, & PySR/symbolic regression to build intuition about models
I got that down to just three (!!!) by using SHAP to say which were important, & PySR/symbolic regression to build intuition about models
I mean, depending on how you measure it they already did 😭 about ~33% of the vote both times but a dip in popular vote. Only because of the UK's absurd voting system does he have a 'mandate' at all
October 24, 2025 at 7:13 AM
I mean, depending on how you measure it they already did 😭 about ~33% of the vote both times but a dip in popular vote. Only because of the UK's absurd voting system does he have a 'mandate' at all
yes: i was playing minecraft when i named this package
no: i am not ashamed of the acronym
no: i am not ashamed of the acronym
October 22, 2025 at 3:20 PM
yes: i was playing minecraft when i named this package
no: i am not ashamed of the acronym
no: i am not ashamed of the acronym
Selection functions are REALLY important - that's the conclusion of our new paper on the open cluster selection function! 🔭☄️ #galactic
Even simple things (like a cluster's orbit) can cause a *major* boost to cluster detectability, biasing results not corrected for selection effects in *big* ways.
Even simple things (like a cluster's orbit) can cause a *major* boost to cluster detectability, biasing results not corrected for selection effects in *big* ways.
October 22, 2025 at 12:54 PM
Selection functions are REALLY important - that's the conclusion of our new paper on the open cluster selection function! 🔭☄️ #galactic
Even simple things (like a cluster's orbit) can cause a *major* boost to cluster detectability, biasing results not corrected for selection effects in *big* ways.
Even simple things (like a cluster's orbit) can cause a *major* boost to cluster detectability, biasing results not corrected for selection effects in *big* ways.