Karandeep Singh
@kdpsingh.bsky.social
Faculty at UC San Diego. #rstats. Creator of Tidier.jl #julialang. #GoBlue. Views own.
Pinned
Karandeep Singh
@kdpsingh.bsky.social
· Jun 27
No amount of AI can overcome bad health policy.
Reposted by Karandeep Singh
App: Do you like this app?
User: Yes
…15 seconds pass…
iOS: On a completely unrelated note, would you like to give this app a star rating?
User: Yes
…15 seconds pass…
iOS: On a completely unrelated note, would you like to give this app a star rating?
December 26, 2025 at 4:34 PM
App: Do you like this app?
User: Yes
…15 seconds pass…
iOS: On a completely unrelated note, would you like to give this app a star rating?
User: Yes
…15 seconds pass…
iOS: On a completely unrelated note, would you like to give this app a star rating?
They’re saying the quiet part out loud.
December 21, 2025 at 6:12 AM
They’re saying the quiet part out loud.
People worry about the AI bubble.
AI worries about the people bubble.
AI worries about the people bubble.
December 21, 2025 at 5:59 AM
People worry about the AI bubble.
AI worries about the people bubble.
AI worries about the people bubble.
Reposted by Karandeep Singh
Editorial by Eileen Kim, MD, Vincent X. Liu, MD, MSc, and Karandeep Singh, MD, MMSc: AI Scribes Are Not Productivity Tools (Yet) nejm.ai/48aLhvx
@kdpsingh.bsky.social #AI #MedSky #MLSky
@kdpsingh.bsky.social #AI #MedSky #MLSky
December 18, 2025 at 5:20 PM
Editorial by Eileen Kim, MD, Vincent X. Liu, MD, MSc, and Karandeep Singh, MD, MMSc: AI Scribes Are Not Productivity Tools (Yet) nejm.ai/48aLhvx
@kdpsingh.bsky.social #AI #MedSky #MLSky
@kdpsingh.bsky.social #AI #MedSky #MLSky
Reposted by Karandeep Singh
The problem with virtual care was never the tools.
It was the lack of shared rules for presence once proximity changed.
Im calling that missing foundation the Presence Layer. #Informatics
It was the lack of shared rules for presence once proximity changed.
Im calling that missing foundation the Presence Layer. #Informatics
December 18, 2025 at 4:53 AM
The problem with virtual care was never the tools.
It was the lack of shared rules for presence once proximity changed.
Im calling that missing foundation the Presence Layer. #Informatics
It was the lack of shared rules for presence once proximity changed.
Im calling that missing foundation the Presence Layer. #Informatics
"Crypto-Standard Science" explains a lot of what is happening right now.
FDA requiring 1 clinical trial instead of 2 for drug approval? Crypto-Standard Science
FDA proposing "impossible" standards for vaccine testing? CSS
FDA holding closed-door meetings with Russell Brand? CSS
FDA requiring 1 clinical trial instead of 2 for drug approval? Crypto-Standard Science
FDA proposing "impossible" standards for vaccine testing? CSS
FDA holding closed-door meetings with Russell Brand? CSS
December 15, 2025 at 5:15 AM
"Crypto-Standard Science" explains a lot of what is happening right now.
FDA requiring 1 clinical trial instead of 2 for drug approval? Crypto-Standard Science
FDA proposing "impossible" standards for vaccine testing? CSS
FDA holding closed-door meetings with Russell Brand? CSS
FDA requiring 1 clinical trial instead of 2 for drug approval? Crypto-Standard Science
FDA proposing "impossible" standards for vaccine testing? CSS
FDA holding closed-door meetings with Russell Brand? CSS
Personal view: “Scientists no longer have to mouth DEI shibboleths to garner funding” is an interesting way of saying that the NIH does not plan to address health disparities and the health of minorities.
thespectator.com/topic/cured-...
thespectator.com/topic/cured-...
How we cured DEI at the National Institutes of Health
The National Institutes of Health will no longer fund research based on DEI that has diverted resources from advancing health
thespectator.com
December 14, 2025 at 5:09 AM
Personal view: “Scientists no longer have to mouth DEI shibboleths to garner funding” is an interesting way of saying that the NIH does not plan to address health disparities and the health of minorities.
thespectator.com/topic/cured-...
thespectator.com/topic/cured-...
Reposted by Karandeep Singh
Nice paper Ben but it's hard to justify use of AUROC except when there is only one candidate model. AUROC should not be used to judge improvements in predictive discrimination. The width of a well-calibrated risk distribution is far better for that. #StatsSky #Statistics fharrell.com/post/addvalue
Statistically Efficient Ways to Quantify Added Predictive Value of New Measurements – Statistical Thinking
Researchers have used contorted, inefficient, and arbitrary analyses to demonstrated added value in biomarkers, genes, and new lab measurements. Traditional statistical measures have always been up to...
fharrell.com
December 13, 2025 at 3:18 PM
Nice paper Ben but it's hard to justify use of AUROC except when there is only one candidate model. AUROC should not be used to judge improvements in predictive discrimination. The width of a well-calibrated risk distribution is far better for that. #StatsSky #Statistics fharrell.com/post/addvalue
Reposted by Karandeep Singh
Our guidance regarding performance measures for medical AI models is finally out!
- Stop bashing AUROC, although it does not settle things
- Calibration and clinical utility are key
- Show risk distributions
- Classification statistics (e.g. F1) are improper
www.thelancet.com/journals/lan...
- Stop bashing AUROC, although it does not settle things
- Calibration and clinical utility are key
- Show risk distributions
- Classification statistics (e.g. F1) are improper
www.thelancet.com/journals/lan...
Evaluation of performance measures in predictive artificial intelligence models to support medical decisions: overview and guidance
Numerous measures have been proposed to illustrate the performance of predictive artificial
intelligence (AI) models. Selecting appropriate performance measures is essential
for predictive AI models i...
www.thelancet.com
December 13, 2025 at 2:04 PM
Our guidance regarding performance measures for medical AI models is finally out!
- Stop bashing AUROC, although it does not settle things
- Calibration and clinical utility are key
- Show risk distributions
- Classification statistics (e.g. F1) are improper
www.thelancet.com/journals/lan...
- Stop bashing AUROC, although it does not settle things
- Calibration and clinical utility are key
- Show risk distributions
- Classification statistics (e.g. F1) are improper
www.thelancet.com/journals/lan...
Reposted by Karandeep Singh
Luxical: A new text embedding model that excels in CPU.
Blog: datologyai.com/blog/introdu...
Paper: arxiv.org/abs/2512.09015
Code: github.com/datologyai/l...
Model: huggingface.co/DatologyAI/l...
Blog: datologyai.com/blog/introdu...
Paper: arxiv.org/abs/2512.09015
Code: github.com/datologyai/l...
Model: huggingface.co/DatologyAI/l...
December 13, 2025 at 5:21 AM
Luxical: A new text embedding model that excels in CPU.
Blog: datologyai.com/blog/introdu...
Paper: arxiv.org/abs/2512.09015
Code: github.com/datologyai/l...
Model: huggingface.co/DatologyAI/l...
Blog: datologyai.com/blog/introdu...
Paper: arxiv.org/abs/2512.09015
Code: github.com/datologyai/l...
Model: huggingface.co/DatologyAI/l...
Reposted by Karandeep Singh
Officially stashing away “Their trajectory is parallel to mine on a log scale at least” for any time comparison in robbing me of job in life
(This is a post on insecurity not a knock on rstats if that wasn’t clear)
(This is a post on insecurity not a knock on rstats if that wasn’t clear)
December 13, 2025 at 1:52 AM
Officially stashing away “Their trajectory is parallel to mine on a log scale at least” for any time comparison in robbing me of job in life
(This is a post on insecurity not a knock on rstats if that wasn’t clear)
(This is a post on insecurity not a knock on rstats if that wasn’t clear)
Reposted by Karandeep Singh
The authors of a new editorial write that #AI scribes for clinical practice show early promise for improved productivity and lay out the conditions under which these tools may substantially improve productivity in the future. Read the full editorial: nejm.ai/48aLhvx
@kdpsingh.bsky.social
@kdpsingh.bsky.social
December 12, 2025 at 8:17 PM
The authors of a new editorial write that #AI scribes for clinical practice show early promise for improved productivity and lay out the conditions under which these tools may substantially improve productivity in the future. Read the full editorial: nejm.ai/48aLhvx
@kdpsingh.bsky.social
@kdpsingh.bsky.social
AI Scribe Called Quest is the crossover album we all deserve but will never see.
December 12, 2025 at 5:47 PM
AI Scribe Called Quest is the crossover album we all deserve but will never see.
Reposted by Karandeep Singh
Catch JCHI Exec Director @kdpsingh.bsky.social @ucsdhealth.bsky.social Dec 10, 1 p.m. PST for a virtual keynote at Stanford AI+HEALTH Conference. A great lineup on the latest in AI and healthcare.
Register ⬇️
medhub.stanford.edu/event/aiheal...
Register ⬇️
medhub.stanford.edu/event/aiheal...
December 9, 2025 at 7:39 PM
Catch JCHI Exec Director @kdpsingh.bsky.social @ucsdhealth.bsky.social Dec 10, 1 p.m. PST for a virtual keynote at Stanford AI+HEALTH Conference. A great lineup on the latest in AI and healthcare.
Register ⬇️
medhub.stanford.edu/event/aiheal...
Register ⬇️
medhub.stanford.edu/event/aiheal...
The Lesser-Known StreisAInd Effect
Statistician: AI gets too much hype. I’m going to fix that.
ML researcher: How?
Stats: I’m going to give a talk about AI hype…
ML: …
Stats: and then do a press release about it…
ML: To hype it up?
Stats: Yes.
Statistician: AI gets too much hype. I’m going to fix that.
ML researcher: How?
Stats: I’m going to give a talk about AI hype…
ML: …
Stats: and then do a press release about it…
ML: To hype it up?
Stats: Yes.
December 9, 2025 at 5:33 AM
The Lesser-Known StreisAInd Effect
Statistician: AI gets too much hype. I’m going to fix that.
ML researcher: How?
Stats: I’m going to give a talk about AI hype…
ML: …
Stats: and then do a press release about it…
ML: To hype it up?
Stats: Yes.
Statistician: AI gets too much hype. I’m going to fix that.
ML researcher: How?
Stats: I’m going to give a talk about AI hype…
ML: …
Stats: and then do a press release about it…
ML: To hype it up?
Stats: Yes.
Reposted by Karandeep Singh
Machine learning has developed remarkable new ideas about how to develop prediction algorithms. But always baffled me why the field had to reinvent how to evaluate models. Here we show F score should not be used to evaluate medical prediction models link.springer.com/epdf/10.1186...
The F score ranks diagnostic tests and prediction models inconsistently with their clinical utility
link.springer.com
December 8, 2025 at 6:53 PM
Machine learning has developed remarkable new ideas about how to develop prediction algorithms. But always baffled me why the field had to reinvent how to evaluate models. Here we show F score should not be used to evaluate medical prediction models link.springer.com/epdf/10.1186...
iOS 26: “You are now entering Game Mode. I’m going to just leave this banner up here for 20 seconds to make sure you are aware. Apologies if this completely blocks a critical part of the screen as you play.”
iOS 26: “Wait, why did you just quit the game?”
iOS 26: “Wait, why did you just quit the game?”
December 8, 2025 at 6:01 AM
iOS 26: “You are now entering Game Mode. I’m going to just leave this banner up here for 20 seconds to make sure you are aware. Apologies if this completely blocks a critical part of the screen as you play.”
iOS 26: “Wait, why did you just quit the game?”
iOS 26: “Wait, why did you just quit the game?”
Reposted by Karandeep Singh
Really enjoyed the fantastic session we had yesterday on applications of AI in hematology at #ASH25 with @wondermixtape.bsky.social,
Sanam Loghavi, and @kdpsingh.bsky.social!
Sanam Loghavi, and @kdpsingh.bsky.social!
December 7, 2025 at 12:00 PM
Really enjoyed the fantastic session we had yesterday on applications of AI in hematology at #ASH25 with @wondermixtape.bsky.social,
Sanam Loghavi, and @kdpsingh.bsky.social!
Sanam Loghavi, and @kdpsingh.bsky.social!
Reposted by Karandeep Singh
Looking forward to our joint session Advancing Hematology Through Artificial Intelligence today at 4pm Hyatt Regency Ballroom S #ASH25
w/ Vijay Sankaran Sanam Loghavi & @kdpsingh.bsky.social on how #AI is will impact hematology across discovery science, diagnostics, & clinical care
w/ Vijay Sankaran Sanam Loghavi & @kdpsingh.bsky.social on how #AI is will impact hematology across discovery science, diagnostics, & clinical care
December 6, 2025 at 8:26 PM
Looking forward to our joint session Advancing Hematology Through Artificial Intelligence today at 4pm Hyatt Regency Ballroom S #ASH25
w/ Vijay Sankaran Sanam Loghavi & @kdpsingh.bsky.social on how #AI is will impact hematology across discovery science, diagnostics, & clinical care
w/ Vijay Sankaran Sanam Loghavi & @kdpsingh.bsky.social on how #AI is will impact hematology across discovery science, diagnostics, & clinical care
Reposted by Karandeep Singh
TITANS & MIRAS: real continual learning
MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs)
TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module
let’s dive in!
research.google/blog/titans-...
MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs)
TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module
let’s dive in!
research.google/blog/titans-...
Titans + MIRAS: Helping AI have long-term memory
research.google
December 6, 2025 at 4:04 PM
TITANS & MIRAS: real continual learning
MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs)
TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module
let’s dive in!
research.google/blog/titans-...
MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs)
TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module
let’s dive in!
research.google/blog/titans-...
Reposted by Karandeep Singh
Reposted by Karandeep Singh
Round of applause for this researcher’s work.
Brilliant backdoor attack. 🤣
Kohler Health also confirmed that it can harness the collected poop 💩 pics to train AI.
Brilliant backdoor attack. 🤣
Kohler Health also confirmed that it can harness the collected poop 💩 pics to train AI.
December 5, 2025 at 6:38 AM
Round of applause for this researcher’s work.
Brilliant backdoor attack. 🤣
Kohler Health also confirmed that it can harness the collected poop 💩 pics to train AI.
Brilliant backdoor attack. 🤣
Kohler Health also confirmed that it can harness the collected poop 💩 pics to train AI.