@chenhaotan.bsky.social
Associate professor at the University of Chicago. Working on human-centered AI, NLP, CSS. https://chenhaot.com, https://substack.com/@cichicago
Due to a speaker conflict, we will have the online seminar on **Thursday** at 12pm ET/11am CT/9am PT. Come to learn about AI & Scientific Discovery in material science!
ai-scientific-discovery.github.io
ai-scientific-discovery.github.io
November 3, 2025 at 4:50 PM
Due to a speaker conflict, we will have the online seminar on **Thursday** at 12pm ET/11am CT/9am PT. Come to learn about AI & Scientific Discovery in material science!
ai-scientific-discovery.github.io
ai-scientific-discovery.github.io
October 27, 2025 at 6:02 PM
This week we will have Rose Yu from UCSD sharing her work on grounding foundation models for physical sciences!
ai-scientific-discovery.github.io
ai-scientific-discovery.github.io
October 20, 2025 at 10:10 PM
This week we will have Rose Yu from UCSD sharing her work on grounding foundation models for physical sciences!
ai-scientific-discovery.github.io
ai-scientific-discovery.github.io
This week we will have two speakers at the AI & Scientific Discovery online seminar: Yihang Wang and Jiachen Yao!
ai-scientific-discovery.github.io
ai-scientific-discovery.github.io
October 14, 2025 at 5:15 PM
This week we will have two speakers at the AI & Scientific Discovery online seminar: Yihang Wang and Jiachen Yao!
ai-scientific-discovery.github.io
ai-scientific-discovery.github.io
Jilian Fisher gave a thought provoking talk on political bias. 1. Biased LLMs influence people’s political decision making. 2. We should seek approximations of political neutrality given the impossibilities of a perfect definition.
October 10, 2025 at 2:00 PM
Jilian Fisher gave a thought provoking talk on political bias. 1. Biased LLMs influence people’s political decision making. 2. We should seek approximations of political neutrality given the impossibilities of a perfect definition.
The amazing @jmendelsohn2.bsky.social kicking off the NLP4Democracy workshop!
October 10, 2025 at 1:32 PM
The amazing @jmendelsohn2.bsky.social kicking off the NLP4Democracy workshop!
Excited that Jean Kossaifi will speak at AI & Scientific Discovery this week!
ai-scientific-discovery.github.io
ai-scientific-discovery.github.io
October 6, 2025 at 6:12 PM
Excited that Jean Kossaifi will speak at AI & Scientific Discovery this week!
ai-scientific-discovery.github.io
ai-scientific-discovery.github.io
Thank Lei Li for for giving a great talk at the inaugural AI & Scientific Discovery online seminar!
The analogy between language and molecule generation is interesting. Also, it is nice to see markov random field after a long exposure to LLM papers!
ai-scientific-discovery.github.io
The analogy between language and molecule generation is interesting. Also, it is nice to see markov random field after a long exposure to LLM papers!
ai-scientific-discovery.github.io
October 6, 2025 at 6:11 PM
Thank Lei Li for for giving a great talk at the inaugural AI & Scientific Discovery online seminar!
The analogy between language and molecule generation is interesting. Also, it is nice to see markov random field after a long exposure to LLM papers!
ai-scientific-discovery.github.io
The analogy between language and molecule generation is interesting. Also, it is nice to see markov random field after a long exposure to LLM papers!
ai-scientific-discovery.github.io
@uchicagoci.bsky.social will be at #COLM2025. Say hi if you are around!
October 6, 2025 at 4:49 PM
@uchicagoci.bsky.social will be at #COLM2025. Say hi if you are around!
@ari-holtzman.bsky.social kicking off @uchicagoci.bsky.social seminar this year! Full house!
October 2, 2025 at 5:55 PM
@ari-holtzman.bsky.social kicking off @uchicagoci.bsky.social seminar this year! Full house!
Excited that we are having the first talk in AI & Scientific Discovery online seminar on Friday at 12pm ET/11am CT/9am PT by the awesome Lei Li from CMU!
🧪 Generative AI for Functional Protein Design🤖
#artificialintelligence #scientificdiscovery
ai-scientific-discovery.github.io
🧪 Generative AI for Functional Protein Design🤖
#artificialintelligence #scientificdiscovery
ai-scientific-discovery.github.io
September 29, 2025 at 5:57 PM
Excited that we are having the first talk in AI & Scientific Discovery online seminar on Friday at 12pm ET/11am CT/9am PT by the awesome Lei Li from CMU!
🧪 Generative AI for Functional Protein Design🤖
#artificialintelligence #scientificdiscovery
ai-scientific-discovery.github.io
🧪 Generative AI for Functional Protein Design🤖
#artificialintelligence #scientificdiscovery
ai-scientific-discovery.github.io
🚀 We’re thrilled to announce the upcoming AI & Scientific Discovery online seminar! We have an amazing lineup of speakers.
This series will dive into how AI is accelerating research, enabling breakthroughs, and shaping the future of research across disciplines.
ai-scientific-discovery.github.io
This series will dive into how AI is accelerating research, enabling breakthroughs, and shaping the future of research across disciplines.
ai-scientific-discovery.github.io
September 25, 2025 at 6:28 PM
🚀 We’re thrilled to announce the upcoming AI & Scientific Discovery online seminar! We have an amazing lineup of speakers.
This series will dive into how AI is accelerating research, enabling breakthroughs, and shaping the future of research across disciplines.
ai-scientific-discovery.github.io
This series will dive into how AI is accelerating research, enabling breakthroughs, and shaping the future of research across disciplines.
ai-scientific-discovery.github.io
As AI becomes increasingly capable of conducting analyses and following instructions, my prediction is that the role of scientists will increasingly focus on identifying and selecting important problems to work on ("selector"), and effectively evaluating analyses performed by AI ("evaluator").
September 16, 2025 at 3:07 PM
As AI becomes increasingly capable of conducting analyses and following instructions, my prediction is that the role of scientists will increasingly focus on identifying and selecting important problems to work on ("selector"), and effectively evaluating analyses performed by AI ("evaluator").
cannot tell whether this is a typo in the AI act
July 24, 2025 at 1:41 AM
cannot tell whether this is a typo in the AI act
But isn’t Mechanistic Interpretability all you need to understand LLMs? We argue that prompt science and MI are complementary. So far, MI has been better at confirming hypotheses, but prompt science has excelled at helping us develop novel hypotheses when we’re still in the dark.
July 9, 2025 at 8:08 PM
But isn’t Mechanistic Interpretability all you need to understand LLMs? We argue that prompt science and MI are complementary. So far, MI has been better at confirming hypotheses, but prompt science has excelled at helping us develop novel hypotheses when we’re still in the dark.
Prompting is our most successful tool for exploring LLMs, but the term evokes eye-rolls and grimaces from scientists. Why? Because prompting as scientific inquiry has become conflated with prompt engineering.
This is holding us back. 🧵and new paper with @ari-holtzman.bsky.social .
This is holding us back. 🧵and new paper with @ari-holtzman.bsky.social .
July 9, 2025 at 8:07 PM
Prompting is our most successful tool for exploring LLMs, but the term evokes eye-rolls and grimaces from scientists. Why? Because prompting as scientific inquiry has become conflated with prompt engineering.
This is holding us back. 🧵and new paper with @ari-holtzman.bsky.social .
This is holding us back. 🧵and new paper with @ari-holtzman.bsky.social .
Nerdy addendum: why does predicting past workload measure fatigue?
It's a bit like predicting chronological age to measure biological age.
If workload and other shocks affect fatigue, and that's what affects note-writing...
Predicting workload gets you fatigue for free!
It's a bit like predicting chronological age to measure biological age.
If workload and other shocks affect fatigue, and that's what affects note-writing...
Predicting workload gets you fatigue for free!
July 2, 2025 at 7:29 PM
Nerdy addendum: why does predicting past workload measure fatigue?
It's a bit like predicting chronological age to measure biological age.
If workload and other shocks affect fatigue, and that's what affects note-writing...
Predicting workload gets you fatigue for free!
It's a bit like predicting chronological age to measure biological age.
If workload and other shocks affect fatigue, and that's what affects note-writing...
Predicting workload gets you fatigue for free!
LLMs predict the next word—their writing is predictable by definition.
Tired physician notes turn out to be highly predictable too.
So LLM notes might LOOK fine…
…but encode the same subtle problems as fatigued human writing.
Tired physician notes turn out to be highly predictable too.
So LLM notes might LOOK fine…
…but encode the same subtle problems as fatigued human writing.
July 2, 2025 at 7:26 PM
LLMs predict the next word—their writing is predictable by definition.
Tired physician notes turn out to be highly predictable too.
So LLM notes might LOOK fine…
…but encode the same subtle problems as fatigued human writing.
Tired physician notes turn out to be highly predictable too.
So LLM notes might LOOK fine…
…but encode the same subtle problems as fatigued human writing.
When we detected fatigue in notes, those doctors made *worse decisions* for the patient
For example: Yield of testing for heart attack was *far lower*
(Past studies of fatigue show little impact on patients—but they looked only at schedules: mental state is hard to measure!)
For example: Yield of testing for heart attack was *far lower*
(Past studies of fatigue show little impact on patients—but they looked only at schedules: mental state is hard to measure!)
July 2, 2025 at 7:25 PM
When we detected fatigue in notes, those doctors made *worse decisions* for the patient
For example: Yield of testing for heart attack was *far lower*
(Past studies of fatigue show little impact on patients—but they looked only at schedules: mental state is hard to measure!)
For example: Yield of testing for heart attack was *far lower*
(Past studies of fatigue show little impact on patients—but they looked only at schedules: mental state is hard to measure!)
Here's what tired doctors do differently:
- Write more *predictable* notes (i.e., an LLM can easily predict the next word)
- Use fewer *insight* words ("I think," "believe")
- Use more *certainty* words ("clear," "definite")
- Write less *complex* sentences
- Write more *predictable* notes (i.e., an LLM can easily predict the next word)
- Use fewer *insight* words ("I think," "believe")
- Use more *certainty* words ("clear," "definite")
- Write less *complex* sentences
July 2, 2025 at 7:25 PM
Here's what tired doctors do differently:
- Write more *predictable* notes (i.e., an LLM can easily predict the next word)
- Use fewer *insight* words ("I think," "believe")
- Use more *certainty* words ("clear," "definite")
- Write less *complex* sentences
- Write more *predictable* notes (i.e., an LLM can easily predict the next word)
- Use fewer *insight* words ("I think," "believe")
- Use more *certainty* words ("clear," "definite")
- Write less *complex* sentences
It predicts pretty well—not just shifts in the last week, but also:
1. Who’s working an overnight shift (in our data + external validation in MIMIC)
2. Who’s working on a disruptive circadian schedule
3. How many patients has the doc seen *on the current shift*
1. Who’s working an overnight shift (in our data + external validation in MIMIC)
2. Who’s working on a disruptive circadian schedule
3. How many patients has the doc seen *on the current shift*
July 2, 2025 at 7:24 PM
It predicts pretty well—not just shifts in the last week, but also:
1. Who’s working an overnight shift (in our data + external validation in MIMIC)
2. Who’s working on a disruptive circadian schedule
3. How many patients has the doc seen *on the current shift*
1. Who’s working an overnight shift (in our data + external validation in MIMIC)
2. Who’s working on a disruptive circadian schedule
3. How many patients has the doc seen *on the current shift*
Chao-Chun led this work, training an algorithm that reads the notes ER doctors write…
…and detect whether the doc who wrote it was tired.
(Basically: predict quasi-random variation in number of shifts worked in the past week—the result is a “tiredness detector”)
…and detect whether the doc who wrote it was tired.
(Basically: predict quasi-random variation in number of shifts worked in the past week—the result is a “tiredness detector”)
July 2, 2025 at 7:23 PM
Chao-Chun led this work, training an algorithm that reads the notes ER doctors write…
…and detect whether the doc who wrote it was tired.
(Basically: predict quasi-random variation in number of shifts worked in the past week—the result is a “tiredness detector”)
…and detect whether the doc who wrote it was tired.
(Basically: predict quasi-random variation in number of shifts worked in the past week—the result is a “tiredness detector”)
When you walk into the ER, you could get a doc:
1. Fresh from a week of not working
2. Tired from working too many shifts
@oziadias.bsky.social has been both and thinks that they're different! But can you tell from their notes? Yes we can! Paper @natcomms.nature.com www.nature.com/articles/s41...
1. Fresh from a week of not working
2. Tired from working too many shifts
@oziadias.bsky.social has been both and thinks that they're different! But can you tell from their notes? Yes we can! Paper @natcomms.nature.com www.nature.com/articles/s41...
July 2, 2025 at 7:22 PM
When you walk into the ER, you could get a doc:
1. Fresh from a week of not working
2. Tired from working too many shifts
@oziadias.bsky.social has been both and thinks that they're different! But can you tell from their notes? Yes we can! Paper @natcomms.nature.com www.nature.com/articles/s41...
1. Fresh from a week of not working
2. Tired from working too many shifts
@oziadias.bsky.social has been both and thinks that they're different! But can you tell from their notes? Yes we can! Paper @natcomms.nature.com www.nature.com/articles/s41...