Sign in

Karandeep(Singh)

@kdpsingh.bsky.social
5K followers 774 following 619 posts

Faculty at UC San Diego. #rstats. Creator of Tidier.jl #julialang. #GoBlue. Views own.

PostsRepliesMedia
Karandeep(Singh) @kdpsingh.bsky.social · 22h
The Ship of Thaiseus: when AI writes the initial draft but you edit it to the point that none of the original words remain.
040
Karandeep(Singh) @kdpsingh.bsky.social · 03/10/2026
Interesting branding strategy.
0133
Reposted by Karandeep(Singh)
brennan @brennan.computer · 03/10/2026
hard to believe 2018 was 27 years ago
2734
Karandeep(Singh) @kdpsingh.bsky.social · 02/10/2026
In SF, they just sell AGI at vending machines.
150
Karandeep(Singh) @kdpsingh.bsky.social · 01/10/2026
“Methods: In this retrospective study, we built a 2-variable model to predict the binary outcome using logistic superintelligence.”
161
Karandeep(Singh) @kdpsingh.bsky.social · 01/10/2026
I remember when we used to call them neural networks. That was before the Executive Order on Deep Learning.
060
Karandeep(Singh) @kdpsingh.bsky.social · 30/09/2026
Opus 5.5 has been limitless on Claude but still costs money on Copilot Cowork (which is priced on usage). In contrast, Sonnet 5.5 is essentially free.
010
Karandeep(Singh) @kdpsingh.bsky.social · 30/09/2026
PSA: Sonnet 5.5 is insanely cheap on Copilot Cowork (~$0 for a workflow that would typically cost $3-4 on Sonnet 5). Either this is a rounding error on Microsoft's side, or Copilot Cowork might just be ready to scale for routine workflows.
122
Reposted by Karandeep(Singh)
Ciaran Martin @ciaranm.bsky.social · 29/09/2026
I was wrong. It’s happening
222045245
Karandeep(Singh) @kdpsingh.bsky.social · 28/09/2026
Passing by Penn Station soon in case you discover that you have a quantum laptop and need the superpositioned one picked up.
110
Karandeep(Singh) @kdpsingh.bsky.social · 28/09/2026
Linen Data 🤝 Tidy
080
Karandeep(Singh) @kdpsingh.bsky.social · 27/09/2026
Wheezing from laughter lol
120
Reposted by Karandeep(Singh)
Emily Moin @emilymoin.com · 06/04/2026
2103
Reposted by Karandeep(Singh)
Jason Moore @moorejh.bsky.social · 27/09/2026
And uninstalling drivers to make room in 640K of conventional memory to run Doom #retrocomputing
0317
Reposted by Karandeep(Singh)
cee @cee.wtf · 24/09/2026
Little guy was doing some research
61342
Reposted by Karandeep(Singh)
Jason Moore @moorejh.bsky.social · 24/09/2026
Ultima Online turns 29! I played this in grad school. One of my favorite games of all time uo.com/2026/09/24/u... #ultima #ultimaonline #lordbritish #retrogaming
1132
Karandeep(Singh) @kdpsingh.bsky.social · 24/09/2026
Tired: Felony Bench Wired: Intercontinental Crime Bench
020
Karandeep(Singh) @kdpsingh.bsky.social · 24/09/2026
You call my drink chai tea, I call your drink coffee coffee.
081
Reposted by Karandeep(Singh)
Tiger Tang @tigertang.bsky.social · 24/09/2026
📘 It’s official: my first book, Work Automation with R, is published! I started with a simple belief: you don’t need to be an R expert to build useful automation with R. I certainly wasn’t one when I started. 📖 www.routledge.com/Work-Automat... 1/5 #RStats #RProgramming
routledge.com
Work Automation with R
Work Automation with R provides a solid framework on how to automate tasks and workflows in the workplace with R. This book introduces the most common components used in R automation, one by one, and ...
6416
Karandeep(Singh) @kdpsingh.bsky.social · 22/09/2026
I wish that Siri AI was Cowork for iOS, and that iPhone Duo could function as a sidecar display for Mac.
010
Reposted by Karandeep(Singh)
Jason Moore @moorejh.bsky.social · 18/09/2026
Oh 1989. If you only knew..... #coding
172
Reposted by Karandeep(Singh)
Tim Fernholz @timfernholz.com · 18/09/2026
I did not have Accenture on my bingo card for this one techcrunch.com/2026/09/18/a...
techcrunch.com
Anthropic's first embedded evaluator is ... Accenture? | TechCrunch
Accenture is about to take on its most high-risk consulting engagement ever.
1112
Reposted by Karandeep(Singh)
Rohit Vashisht @rohitv.bsky.social · 17/09/2026
Excellent review! A must read and ponder upon with your favorite drink www.cell.com/cell/fulltex... @kdpsingh.bsky.social
cell.com
Avoiding common failures in AI for health and medicine
Salaudeen et al. review common reliability failures in predictive and generative AI for healthcare, including erroneous model outputs, clinically unjustified performance differences, and deployment-ti...
042
Reposted by Karandeep(Singh)
Nick Fisher @hydroxide.dev · 17/09/2026
Actually maybe I don't even need Jev!
1211
Karandeep(Singh) @kdpsingh.bsky.social · 17/09/2026
OpenAIFelonyBench += 6
011
Karandeep(Singh) @kdpsingh.bsky.social · 16/09/2026
To paraphrase Chris Rock, we don't need AI control - we need token control.
010
Karandeep(Singh) @kdpsingh.bsky.social · 16/09/2026
Tired: Chasing the frontier Wired: Pacing the frontier Inspired: Encasing the frontier
121
Karandeep(Singh) @kdpsingh.bsky.social · 15/09/2026
We should Pace the Frontier and Magic the Gathering.
020
Karandeep(Singh) @kdpsingh.bsky.social · 15/09/2026
Maybe everyone can just set their learning rates a bit lower to pace the frontier.
141
Reposted by Karandeep(Singh)
Museum of Twitter @museum-of-twitter.bsky.social · 15/09/2026
Mr. Bedtime @InternetHippo

I stepped away from my computer for a moment. When I returned, it was building something. I did not ask it to do this. We're not ready for what's coming

(picture of windows 3d pipes screen saver)
63177693725
Reposted by Karandeep(Singh)
tj mahr 🤘 @tjmahr.com · 14/09/2026
the nintendogeneity problem
021
Reposted by Karandeep(Singh)
Jonathan @ignoranceburns.bsky.social · 13/09/2026
As foretold by the prophet Groening
Still from Futurama showing the "Taco Bellevue Hospital"
442224
Reposted by Karandeep(Singh)
Six Dettmar @sixdettmar.bsky.social · 13/09/2026
I’M AT THE URGENT CARE I’M AT THE TACO BELL I’M AT THE COMBINATION URGENT CARE AND TACO BELL
266118133036
Reposted by Karandeep(Singh)
JD Long @jdlong.cerebralmastication.com · 14/09/2026
this may make you uncomfortable, but this is what peak suburban planning looks like
281
Reposted by Karandeep(Singh)
dame @dame.is · 14/09/2026
did someone say pace the frontier?
Pace The Frontier salsa jar with green lid. Glass bottle containing red chunky salsa with visible pepper and vegetable pieces. Label features "Pace" in red script with a yellow sun logo, "THE FRONTIER" text, and images of red peppers and jalapeños on green banner against white background.
1715
Karandeep(Singh) @kdpsingh.bsky.social · 14/09/2026
Excess water usage detected, self-destruct sequence initiated.
030
Reposted by Karandeep(Singh)
KM @sudox.bsky.social · 09/09/2026
115236
Karandeep(Singh) @kdpsingh.bsky.social · 10/09/2026
“Get off our roads, Mr. Singh. America is for Americans.” I, for one, am so excited to learn from this very kind message from US Homeland Security that cargo shipments shall heretofore be transported by robots capable of turning into cars while Sikh Americans can lay back and chill. Thanks!
031
Reposted by Karandeep(Singh)
natemoore on elm street 👻 @natemoo.re · 08/09/2026
things could have been so much stupider
11076
Karandeep(Singh) @kdpsingh.bsky.social · 02/09/2026
I’m still holding out until I see a Felony Bench score.
000
Karandeep(Singh) @kdpsingh.bsky.social · 01/09/2026
!!! www.bbc.com/news/article...
bbc.com
Duane "Keffe D" Davis found guilty in 1996 murder of rapper Tupac Shakur
The verdict ends decades of mystery as to who was behind the rapper's death.
111
Karandeep(Singh) @kdpsingh.bsky.social · 31/08/2026
051
Karandeep(Singh) @kdpsingh.bsky.social · 31/08/2026
Have you seen this: github.com/cyberpapiii/...
github.com
GitHub - cyberpapiii/chipotlai-max: The AI coding agent that runs on stolen Chipotle compute 🌯 Fork of OpenCode with Pepper AI as default model. Community project to add providers from Home Depot, Low...
The AI coding agent that runs on stolen Chipotle compute 🌯 Fork of OpenCode with Pepper AI as default model. Community project to add providers from Home Depot, Lowes, Target, Starbucks & more....
070
Reposted by Karandeep(Singh)
Scott McGrath @smcgrath.phd · 28/08/2026
Good post and accompanying paper looking at physician revisions of AI drafted content. #MedSky Direct link to the paper:
ai.nejm.org
Physician Edits to AI-Drafted Patient Messages and Their Impact on Clinical Workload
AI systems capable of drafting replies to patient messages are widely implemented within electronic health records (EHRs) to reduce physician workload. Prior studies have shown limited time savings...
021
Karandeep(Singh) @kdpsingh.bsky.social · 28/08/2026
Paper: ai.nejm.org/stoken/defau...
ai.nejm.org
Physician Edits to AI-Drafted Patient Messages and Their Impact on Clinical Workload
AI systems capable of drafting replies to patient messages are widely implemented within electronic health records (EHRs) to reduce physician workload. Prior studies have shown limited time savings...
130
Karandeep(Singh) @kdpsingh.bsky.social · 28/08/2026
But the slowest edits aren't the costliest ones. The greatest cumulative editing burden came from AI drafts about scheduling, lifestyle advice, and referral coordination, which are more common and account for hundreds of physician-hours per 1,000 physicians in just over a year.
120
Karandeep(Singh) @kdpsingh.bsky.social · 28/08/2026
We found 15 distinct categories of physician edits. Biggest marginal increases in editing time: 1. Interpreting radiology results: +70.1% 2. Clarifying/ruling out diagnoses: +63.9% 3. Interpreting lab results: +60.8%
130
Karandeep(Singh) @kdpsingh.bsky.social · 28/08/2026
Patient messages vary wildly in complexity. To account for that, we adjusted for the first 10 principal components of neural embeddings of the original patient messages.
120
Karandeep(Singh) @kdpsingh.bsky.social · 28/08/2026
We built a two-step framework: 1. Use AI to compare AI-drafted replies against what physicians actually sent, to categorize their edits. 2. Use a mixed-effects model to find which edit types cost the most time.
110
Karandeep(Singh) @kdpsingh.bsky.social · 28/08/2026
When physicians start relying on AI-drafted content, how does a health system monitor that human-in-the-loop AI and fix what's broken? Our new @ai.nejm.org study from our team at @jchi-ucsd.bsky.social led by Dr. Lily Poursoltan found hundreds of physicians-hours spent editing AI content.
1104