Andrew Kern Profile
Andrew Kern

@pastramimachine

Followers
14K
Following
93K
Media
3K
Statuses
28K

Evolution, population genetics, machine learning. rhythm section.

Eugene, Oregon
Joined December 2011
Don't wanna be here? Send us removal request.
@pastramimachine
Andrew Kern
5 years
Hiroshi Sugimoto, South Pacific Ocean Maraenui, 1990
Tweet media one
4
31
203
@pastramimachine
Andrew Kern
2 days
unit tests take all the freedom out of programming. feels un-american.
3
0
6
@pastramimachine
Andrew Kern
3 days
One thing that rules about academia is you can go with your lab up a rather insane mountain and no one blinks an eye.
0
0
6
@pastramimachine
Andrew Kern
3 days
RT @stoolpresidente: Top 1 American moment when Joey Chestnut beat up a protestor mid hot dog eating contest.
0
4K
0
@pastramimachine
Andrew Kern
4 days
asking Soham to review the next paper I'm sent.
0
0
1
@pastramimachine
Andrew Kern
4 days
only one thing could make driving in Boston worse.
@pli_cachete
Rota
4 days
The Waymos are coming to Boston
Tweet media one
0
0
3
@pastramimachine
Andrew Kern
5 days
0
3
0
@pastramimachine
Andrew Kern
9 days
Okay i'm not going to show all the results in this thread-- go read the paper! . What I will say is this is very exciting to me, and I think the first use of a LLM framework specifically for population genetics.
1
2
12
@pastramimachine
Andrew Kern
9 days
5) because of the generative nature of cxt, we can do multiple decoding passes and treat those passes as an approximate posterior sample. Incredibly such samples are quite well calibrated, meaning we can rapidly get a sense of uncertainty
Tweet media one
1
0
4
@pastramimachine
Andrew Kern
9 days
4) cxt, because it is a deep language model, has the capacity to be trained on huge corpus. we decided to train it on the entire v0.2 stdpopsim catalog-- it does really well
Tweet media one
1
0
4
@pastramimachine
Andrew Kern
9 days
@KevinKorfmann 3) our method -- that we call cxt -- is highly accurate. in benchmark comparisons against Singer and gamma-SMC, cxt looks very good
Tweet media one
1
1
3
@pastramimachine
Andrew Kern
9 days
@KevinKorfmann 2) the model is small (for a language model) -- 10-20M parameters -- so it trains in minutes or hours, not days. and prediction is blazing fast (decoding human chroms in minutes).
2
0
6
@pastramimachine
Andrew Kern
9 days
@KevinKorfmann there are a few things to highlight: . 1) we perform generative training using the task of 'next coalescence prediction' where the model is asked to predict the next time moving across a chromosome. this is akin to next token prediction in natural language
Tweet media one
1
0
3
@pastramimachine
Andrew Kern
9 days
Really excited about our lab's latest paper led by @KevinKorfmann. Here, Kevin has developed a deep, transformer-based language model for population genetics for inferring coalescent times (times of common ancestry) along recombining chromosomes.
1
38
118
@pastramimachine
Andrew Kern
10 days
"AI is crap" i whisper to myself as i scrape the wrong code block from stackoverflow.
0
0
7
@pastramimachine
Andrew Kern
11 days
AI agents at the command line make me so pumped for a future with robots at home.
0
0
2
@pastramimachine
Andrew Kern
11 days
RT @leo_speidel: Postdoc position in my group in Tokyo! .Please get in touch if you are interested. And happy to discuss projects - rang….
0
36
0
@pastramimachine
Andrew Kern
13 days
claude is the best git interface.
0
0
0
@pastramimachine
Andrew Kern
14 days
Any plant pathologists who can diagnose what happened to this dahlia?
Tweet media one
3
1
2
@pastramimachine
Andrew Kern
14 days
RT @robertsietsema: Here's my idea of the five best pastrami sandwiches in NYC in Monday's free Substack [link in first comment] https://t.….
0
2
0
@pastramimachine
Andrew Kern
14 days
Roblox stock up 25% today on news of nuclear war.
1
0
2