apoorva lal
@apoorvalal.com
causal inference, econometrics, ML, arsenal, loud music, unix, FOSS for scientific computing. opinions my own.
apoorvalal.github.io
(passively) maintains @paperposterbot.bsky.social
apoorvalal.github.io
(passively) maintains @paperposterbot.bsky.social
yeah there's a whole genre of games that's made for slightly theatrical people who got really really into final fantasy in their teenage years and i am not one of them. de gustibus.
January 13, 2026 at 2:56 AM
yeah there's a whole genre of games that's made for slightly theatrical people who got really really into final fantasy in their teenage years and i am not one of them. de gustibus.
I kinda hate jrpgs and tried COE33 and was so confused that the jrpg-est japanese role playing game to ever role play japanese-ly is ... french?
January 13, 2026 at 2:27 AM
I kinda hate jrpgs and tried COE33 and was so confused that the jrpg-est japanese role playing game to ever role play japanese-ly is ... french?
"here's a fork. Let me now spend 8000 words on how you could poke your eyes out with your fork"
CC runs in a sandbox and will complain+be extra conservative if you're not version-controlled. The average v marginal user point is worth making, but the actual computing parts of this were nonsense
CC runs in a sandbox and will complain+be extra conservative if you're not version-controlled. The average v marginal user point is worth making, but the actual computing parts of this were nonsense
January 13, 2026 at 1:49 AM
"here's a fork. Let me now spend 8000 words on how you could poke your eyes out with your fork"
CC runs in a sandbox and will complain+be extra conservative if you're not version-controlled. The average v marginal user point is worth making, but the actual computing parts of this were nonsense
CC runs in a sandbox and will complain+be extra conservative if you're not version-controlled. The average v marginal user point is worth making, but the actual computing parts of this were nonsense
+1
As a practical matter, learn to use symlinks
ln -s /place/to/look current/directory
Lets agentic editors see things you point them to. They're (justifiably) sandboxed by default and will do well w examples
As a practical matter, learn to use symlinks
ln -s /place/to/look current/directory
Lets agentic editors see things you point them to. They're (justifiably) sandboxed by default and will do well w examples
January 13, 2026 at 12:43 AM
+1
As a practical matter, learn to use symlinks
ln -s /place/to/look current/directory
Lets agentic editors see things you point them to. They're (justifiably) sandboxed by default and will do well w examples
As a practical matter, learn to use symlinks
ln -s /place/to/look current/directory
Lets agentic editors see things you point them to. They're (justifiably) sandboxed by default and will do well w examples
you need a horn section to go apeshit for that song though. gotta save that for the 50sth dads band where someone invariably gets really into trombones for some reason
January 12, 2026 at 1:39 AM
you need a horn section to go apeshit for that song though. gotta save that for the 50sth dads band where someone invariably gets really into trombones for some reason
Weird fishes is fun to play. Drummer has to be tight as hell though.
January 12, 2026 at 1:13 AM
Weird fishes is fun to play. Drummer has to be tight as hell though.
nice, filing this away for when i get yelled at for posting random articles from the arxiv archive
January 10, 2026 at 11:11 PM
nice, filing this away for when i get yelled at for posting random articles from the arxiv archive
you kid but what we treat as the lalonde data is dehejia-wahba's trimmed sample, where PSM works suspiciously well, while it really really doesn't in the full sample. conveniently enough this was published with a paper arguing PSM was the best thing since sliced bread
yiqingxu.org/papers/2024_...
yiqingxu.org/papers/2024_...
yiqingxu.org
January 9, 2026 at 10:20 PM
you kid but what we treat as the lalonde data is dehejia-wahba's trimmed sample, where PSM works suspiciously well, while it really really doesn't in the full sample. conveniently enough this was published with a paper arguing PSM was the best thing since sliced bread
yiqingxu.org/papers/2024_...
yiqingxu.org/papers/2024_...
ah good old Lalonde. our tools may change, but the dataset for causal inference classes will always remain the same
January 9, 2026 at 9:54 PM
ah good old Lalonde. our tools may change, but the dataset for causal inference classes will always remain the same
I once tried to trick it into admitting it was the protagonist of GTA3
January 9, 2026 at 6:22 PM
I once tried to trick it into admitting it was the protagonist of GTA3
another pro tip - define it as a snippet and map to ctrl/cmd+shift+semicolon in vscode.
January 9, 2026 at 4:16 PM
another pro tip - define it as a snippet and map to ctrl/cmd+shift+semicolon in vscode.
specifically this guy
github.com/nteract/hydr...
github.com/nteract/hydr...
GitHub - nteract/hydrogen: :atom: Run code interactively, inspect data, and plot. All the power of Jupyter kernels, inside your favorite text editor.
:atom: Run code interactively, inspect data, and plot. All the power of Jupyter kernels, inside your favorite text editor. - nteract/hydrogen
github.com
January 9, 2026 at 4:09 PM
specifically this guy
github.com/nteract/hydr...
github.com/nteract/hydr...
hahaha yep sorry i didn't spell that out. atom (rip)'s great gift to scientific computing
January 9, 2026 at 4:09 PM
hahaha yep sorry i didn't spell that out. atom (rip)'s great gift to scientific computing
its good! you should disable autosave on jupyter (in case you use jupyterlab) btw - that's a footgun with these background read + sync tools.
gateway drug to start using scripts with #%% cells and eventually stop using notebooks entirely.
gateway drug to start using scripts with #%% cells and eventually stop using notebooks entirely.
January 9, 2026 at 3:55 PM
its good! you should disable autosave on jupyter (in case you use jupyterlab) btw - that's a footgun with these background read + sync tools.
gateway drug to start using scripts with #%% cells and eventually stop using notebooks entirely.
gateway drug to start using scripts with #%% cells and eventually stop using notebooks entirely.
posted this on xitter attached to a post by grus (the king of notebook haters):
switch to quarto, or pair with jupytext. notebooks + agentic edits are a disaster (which i bet you've already experienced plenty)
switch to quarto, or pair with jupytext. notebooks + agentic edits are a disaster (which i bet you've already experienced plenty)
January 9, 2026 at 3:37 PM
posted this on xitter attached to a post by grus (the king of notebook haters):
switch to quarto, or pair with jupytext. notebooks + agentic edits are a disaster (which i bet you've already experienced plenty)
switch to quarto, or pair with jupytext. notebooks + agentic edits are a disaster (which i bet you've already experienced plenty)
don't think the norm is necessarily that commit messages are super detailed, but they should be informative enough that you could identify them (and the accompanying changes) to be able to pick/revert. then again maybe this is from git log days not web ui with detailed diffs. when in doubt, commit.
January 9, 2026 at 2:47 AM
don't think the norm is necessarily that commit messages are super detailed, but they should be informative enough that you could identify them (and the accompanying changes) to be able to pick/revert. then again maybe this is from git log days not web ui with detailed diffs. when in doubt, commit.
aliases ftw; anything i'd like to remember to do on the terminal goes into my ~/.zshrc as a 3-4letter alias
eg here i'd reduce it to running the 3 lines at the bottom
eg here i'd reduce it to running the 3 lines at the bottom
January 9, 2026 at 2:12 AM
aliases ftw; anything i'd like to remember to do on the terminal goes into my ~/.zshrc as a 3-4letter alias
eg here i'd reduce it to running the 3 lines at the bottom
eg here i'd reduce it to running the 3 lines at the bottom
squashing commits is pretty easy and worth doing (i used to also just have a string of uninformative commits but that does not go over well in code review
bsky.app/profile/apoo...
bsky.app/profile/apoo...
tbf i'd do this for like 10 commits and squash commits like so:
git reset --soft HEAD~10
git commit -am "actually informative commit message"
git push -f
the gc's are purely staging small units of work.
git reset --soft HEAD~10
git commit -am "actually informative commit message"
git push -f
the gc's are purely staging small units of work.
January 9, 2026 at 1:54 AM
squashing commits is pretty easy and worth doing (i used to also just have a string of uninformative commits but that does not go over well in code review
bsky.app/profile/apoo...
bsky.app/profile/apoo...
This looks cool. Based on the blog post alone, this looks a bit like overidentified gmm with additional auxiliary data moment conditions?
direct.mit.edu/rest/article...
direct.mit.edu/rest/article...
Imposing Moment Restrictions from Auxiliary Data by Weighting
Abstract. In this paper we analyze the estimation of coefficients in regression models under moment restrictions in which the moment restrictions are derived from auxiliary data. The moment restrictio...
direct.mit.edu
January 9, 2026 at 1:32 AM
This looks cool. Based on the blog post alone, this looks a bit like overidentified gmm with additional auxiliary data moment conditions?
direct.mit.edu/rest/article...
direct.mit.edu/rest/article...
also if you let agents write your tests they will move the goalposts until they pass; i've seen
assert np.allclose(estimate, truth, atol=k)
where k gets big as the agent reward-hack.
assert np.allclose(estimate, truth, atol=k)
where k gets big as the agent reward-hack.
January 7, 2026 at 11:25 PM
also if you let agents write your tests they will move the goalposts until they pass; i've seen
assert np.allclose(estimate, truth, atol=k)
where k gets big as the agent reward-hack.
assert np.allclose(estimate, truth, atol=k)
where k gets big as the agent reward-hack.
never ceases to amuse me that the most vocal luddite denizens of this social network fail to notice it is the easiest place to scrape on the entire internet, and it is this way by design.
January 7, 2026 at 4:59 AM
never ceases to amuse me that the most vocal luddite denizens of this social network fail to notice it is the easiest place to scrape on the entire internet, and it is this way by design.
closer to 50% active [adhoc-active in last 3mo] for @paperposterbot.bsky.social what with this being the phd-haver social network and all
January 7, 2026 at 4:57 AM
closer to 50% active [adhoc-active in last 3mo] for @paperposterbot.bsky.social what with this being the phd-haver social network and all