apoorva lal
@apoorvalal.com
causal inference, econometrics, ML, arsenal, loud music, unix, FOSS for scientific computing.
apoorvalal.github.io
(passively) maintains @paperposterbot.bsky.social
apoorvalal.github.io
(passively) maintains @paperposterbot.bsky.social
Who would've thought a ban on aimless position papers in one arxiv field would simply redirect it to another
link 📈🤖
Integrating RCTs, RWD, AI/ML and Statistics: Next-Generation Evidence Synthesis (Yang, Gamalo, Fu) Randomized controlled trials (RCTs) have been the cornerstone of clinical evidence; however, their cost, duration, and restrictive eligibility criteria limit power and external validity. Stu
Integrating RCTs, RWD, AI/ML and Statistics: Next-Generation Evidence Synthesis (Yang, Gamalo, Fu) Randomized controlled trials (RCTs) have been the cornerstone of clinical evidence; however, their cost, duration, and restrictive eligibility criteria limit power and external validity. Stu
November 26, 2025 at 4:24 PM
Who would've thought a ban on aimless position papers in one arxiv field would simply redirect it to another
handy little data hoarder tool i built and deployed for myself but others might find useful lalten.org/linkpull/
1) academic or govt site with a bunch of hyperlinks
2) write a regex to grab all links matching a pattern in said url, copy and paste into file
3) download with wget or curl
1) academic or govt site with a bunch of hyperlinks
2) write a regex to grab all links matching a pattern in said url, copy and paste into file
3) download with wget or curl
November 26, 2025 at 6:19 AM
handy little data hoarder tool i built and deployed for myself but others might find useful lalten.org/linkpull/
1) academic or govt site with a bunch of hyperlinks
2) write a regex to grab all links matching a pattern in said url, copy and paste into file
3) download with wget or curl
1) academic or govt site with a bunch of hyperlinks
2) write a regex to grab all links matching a pattern in said url, copy and paste into file
3) download with wget or curl
an attempt to formalize the "squiggly pre-trends that match well are more credible than smooth pre-trends" intuition.
apoorvalal.github.io/lalgorithms/...
ideation -> simulation -> algorithm sketch -> writeup heavily aided by LLMs
apoorvalal.github.io/lalgorithms/...
ideation -> simulation -> algorithm sketch -> writeup heavily aided by LLMs
November 24, 2025 at 12:57 AM
an attempt to formalize the "squiggly pre-trends that match well are more credible than smooth pre-trends" intuition.
apoorvalal.github.io/lalgorithms/...
ideation -> simulation -> algorithm sketch -> writeup heavily aided by LLMs
apoorvalal.github.io/lalgorithms/...
ideation -> simulation -> algorithm sketch -> writeup heavily aided by LLMs
typical theorist career
November 21, 2025 at 4:11 PM
typical theorist career
What is it about building flash-attn (vanilla uv pip install, v 2.8.3) wheels that oscillates wildly between consuming ~no RAM and consuming 25-26 gb? Don't know enough about the cuda compiler to debug.
This is single threaded (MAX_JOBS=1), anything more OOMs my machine.
This is single threaded (MAX_JOBS=1), anything more OOMs my machine.
November 19, 2025 at 8:24 PM
What is it about building flash-attn (vanilla uv pip install, v 2.8.3) wheels that oscillates wildly between consuming ~no RAM and consuming 25-26 gb? Don't know enough about the cuda compiler to debug.
This is single threaded (MAX_JOBS=1), anything more OOMs my machine.
This is single threaded (MAX_JOBS=1), anything more OOMs my machine.
Crossposting because the QT is there but the audience for this is here
November 15, 2025 at 10:39 PM
Crossposting because the QT is there but the audience for this is here
First wintery week calls for GPU warmth
Large scale OCR (and hopefully downstream embedding, search, and model) project running on a brand new GPU - should be adaptable to many other tasks
github.com/apoorvalal/n...
Large scale OCR (and hopefully downstream embedding, search, and model) project running on a brand new GPU - should be adaptable to many other tasks
github.com/apoorvalal/n...
November 15, 2025 at 7:10 PM
First wintery week calls for GPU warmth
Large scale OCR (and hopefully downstream embedding, search, and model) project running on a brand new GPU - should be adaptable to many other tasks
github.com/apoorvalal/n...
Large scale OCR (and hopefully downstream embedding, search, and model) project running on a brand new GPU - should be adaptable to many other tasks
github.com/apoorvalal/n...
Baiting bsky with nanobanana artist cover portraits of The Agent in commemoration of 2 months of life and a full week of sleeping through the night
November 14, 2025 at 8:41 PM
Baiting bsky with nanobanana artist cover portraits of The Agent in commemoration of 2 months of life and a full week of sleeping through the night
Is this enshittification?
November 14, 2025 at 4:39 PM
Is this enshittification?
upgraded my eGPU setup over the weekend [RIP ADT rube goldberg machine] and decided to document it for posterity (with some promo for my torch econometrics package included)
apoorvalal.github.io/lalgorithms/...
apoorvalal.github.io/lalgorithms/...
GPU-Accelerated Econometrics on Consumer Hardware
tl;dr: A $200 used GPU plugged into an EGPU can make your research code order-of-magnitude faster.
apoorvalal.github.io
November 9, 2025 at 10:06 PM
upgraded my eGPU setup over the weekend [RIP ADT rube goldberg machine] and decided to document it for posterity (with some promo for my torch econometrics package included)
apoorvalal.github.io/lalgorithms/...
apoorvalal.github.io/lalgorithms/...
My twitter feed is doing a 4th round of memeing a shitty inconsequential interview and my bsky feed is 4 people competing to be the goodest boy by circular firing squad. Back to whatsapp and discord then.
November 8, 2025 at 4:45 PM
My twitter feed is doing a 4th round of memeing a shitty inconsequential interview and my bsky feed is 4 people competing to be the goodest boy by circular firing squad. Back to whatsapp and discord then.
The agent's first vibecoding experience
November 7, 2025 at 7:52 PM
The agent's first vibecoding experience
behold, an authoritative answer to the question of "how'd you like em apples?"
[also love how old papers have the most stilted hot-takes in their intro. no citation salute to the elders to get the paper accepted, just passive-voice passive-aggressiveness]
[also love how old papers have the most stilted hot-takes in their intro. no citation salute to the elders to get the paper accepted, just passive-voice passive-aggressiveness]
November 4, 2025 at 7:37 PM
behold, an authoritative answer to the question of "how'd you like em apples?"
[also love how old papers have the most stilted hot-takes in their intro. no citation salute to the elders to get the paper accepted, just passive-voice passive-aggressiveness]
[also love how old papers have the most stilted hot-takes in their intro. no citation salute to the elders to get the paper accepted, just passive-voice passive-aggressiveness]
Updated a classic to reflect our times
October 28, 2025 at 8:00 PM
Updated a classic to reflect our times
Reposted by apoorva lal
if you use a tiling wm, you can event tuck the radio away into your scratch workspace [L; omarchy makes it very easy]
compact htmx makes it work well on mobile [R] with background playback [getting it to not die on suspend was the most javascript i've had to learn].
compact htmx makes it work well on mobile [R] with background playback [getting it to not die on suspend was the most javascript i've had to learn].
October 23, 2025 at 11:41 PM
if you use a tiling wm, you can event tuck the radio away into your scratch workspace [L; omarchy makes it very easy]
compact htmx makes it work well on mobile [R] with background playback [getting it to not die on suspend was the most javascript i've had to learn].
compact htmx makes it work well on mobile [R] with background playback [getting it to not die on suspend was the most javascript i've had to learn].
Go play Crab Souls. It is so very good.
If AAA's want $100/game they need to be significantly better than what smaller / indie devs are making.
They aren't.
The most fun I had last year was Another Crab's Treasure ($30). Silksong this year was $20.
If the sit down restaurant is 4x the cost of the foodcart (& no better), I'm not sorry.
They aren't.
The most fun I had last year was Another Crab's Treasure ($30). Silksong this year was $20.
If the sit down restaurant is 4x the cost of the foodcart (& no better), I'm not sorry.
October 23, 2025 at 5:53 AM
Go play Crab Souls. It is so very good.
split keyboard recs? but like really really split; the wiggly phone cable thing won't do for obvious reasons.
October 23, 2025 at 12:31 AM
split keyboard recs? but like really really split; the wiggly phone cable thing won't do for obvious reasons.
(4,4) - turns out "Sleep like a baby" is black humour.
tag yourself I'm (3,7) [godmode and dead are -1, 9 respectively]
October 22, 2025 at 6:40 PM
(4,4) - turns out "Sleep like a baby" is black humour.
Built and hosted a scrappy fasthtml radio player on a VPS for my own edification+excuse to listen to good radio (Soma, kexp, etc).
Curious how many concurrent users can use it before it crashes the tiny raspberry-pi tier machine i deployed this on.
lalten.org/radio
Curious how many concurrent users can use it before it crashes the tiny raspberry-pi tier machine i deployed this on.
lalten.org/radio
Lalten 🏮 Web Radio Player
lalten.org
October 22, 2025 at 6:20 PM
Built and hosted a scrappy fasthtml radio player on a VPS for my own edification+excuse to listen to good radio (Soma, kexp, etc).
Curious how many concurrent users can use it before it crashes the tiny raspberry-pi tier machine i deployed this on.
lalten.org/radio
Curious how many concurrent users can use it before it crashes the tiny raspberry-pi tier machine i deployed this on.
lalten.org/radio
apoorvalal.github.io/lalgorithms/...
New linklog: local projections are balancing, torch choice models, inhomogeneous poisson processes for event data
New linklog: local projections are balancing, torch choice models, inhomogeneous poisson processes for event data
2025-10-20
links papers Coulombe and Klieber on modern balancing-weights interpretation of Local Projection estimators widespread in macroeconometrics.
apoorvalal.github.io
October 20, 2025 at 5:02 PM
apoorvalal.github.io/lalgorithms/...
New linklog: local projections are balancing, torch choice models, inhomogeneous poisson processes for event data
New linklog: local projections are balancing, torch choice models, inhomogeneous poisson processes for event data
Happy "Econ history nobel followed by econ historians tweeting through their persecution complex" day to those who celebrate
October 13, 2025 at 7:24 PM
Happy "Econ history nobel followed by econ historians tweeting through their persecution complex" day to those who celebrate
watch palo alto shutter all three caltrain stations in response to SB79; atherton is already thanking its lucky stars it shut its only public transit.
[more seriously, is a potential second-order effect that this passes the buck to transit extension being monstrously difficult?]
😶🌫️
[more seriously, is a potential second-order effect that this passes the buck to transit extension being monstrously difficult?]
😶🌫️
October 10, 2025 at 8:29 PM
watch palo alto shutter all three caltrain stations in response to SB79; atherton is already thanking its lucky stars it shut its only public transit.
[more seriously, is a potential second-order effect that this passes the buck to transit extension being monstrously difficult?]
😶🌫️
[more seriously, is a potential second-order effect that this passes the buck to transit extension being monstrously difficult?]
😶🌫️
Reposted by apoorva lal
For those who, like me, didn't get what "Gwern-pilled" meant in this post.
October 8, 2025 at 3:15 PM
For those who, like me, didn't get what "Gwern-pilled" meant in this post.
Put stuff on arxiv, it'll have more impact than 90% of papers because the robots read more comprehensively than most of us. Gwern-pilled conclusion - write for the robots.
(A friend got jumpscared by this and sent it to me)
(A friend got jumpscared by this and sent it to me)
October 7, 2025 at 10:17 PM
Put stuff on arxiv, it'll have more impact than 90% of papers because the robots read more comprehensively than most of us. Gwern-pilled conclusion - write for the robots.
(A friend got jumpscared by this and sent it to me)
(A friend got jumpscared by this and sent it to me)