Jonathan Cheng
@jonathancheng.bsky.social
English PhD turned Machine Learning Researcher. I medicate my imposter syndrome with cold brews.
Current:
Foundation Models @ Apple
Prev:
LLMs/World Models @ Riot Games
RecSys @ Apple
Seattle 🏳️🌈
Current:
Foundation Models @ Apple
Prev:
LLMs/World Models @ Riot Games
RecSys @ Apple
Seattle 🏳️🌈
BLANK!
This is pretty BLANKing BLANK
January 20, 2026 at 7:57 PM
BLANK!
Got to be mere inches from a tiger!
January 10, 2026 at 8:03 PM
Got to be mere inches from a tiger!
Reposted by Jonathan Cheng
1/ We found that deep sequence models memorize atomic facts "geometrically" -- not as an associative lookup table as often imagined.
This opens up practical questions on reasoning/memory/discovery, and also poses a theoretical "memorization puzzle."
This opens up practical questions on reasoning/memory/discovery, and also poses a theoretical "memorization puzzle."
January 8, 2026 at 8:31 PM
1/ We found that deep sequence models memorize atomic facts "geometrically" -- not as an associative lookup table as often imagined.
This opens up practical questions on reasoning/memory/discovery, and also poses a theoretical "memorization puzzle."
This opens up practical questions on reasoning/memory/discovery, and also poses a theoretical "memorization puzzle."
Reposted by Jonathan Cheng
Extracting Books From Production Language Models ahmeda14960.github.io/blackbox_web...
Extracting Books from Production Language Models
ahmeda14960.github.io
January 8, 2026 at 7:33 AM
Extracting Books From Production Language Models ahmeda14960.github.io/blackbox_web...
As chrome adds split tabs, I do start to wonder how much of my visual experience will become an ide 😂
January 6, 2026 at 6:21 PM
As chrome adds split tabs, I do start to wonder how much of my visual experience will become an ide 😂
This hits especially hard as someone who emerged from English departments — listening to a lot of tenured profs contrast their perceived professional practices from coding/computation, at the active cost of their students… while they accrue debt.
W-i-l-d.
W-i-l-d.
I can tell you as a former psych grad who WASN'T taught to code in grad school that choice was a monumental financial loss for so many of us and then I taught myself to code and consulted in R and basically all the consulting I did is obsolete or the time cost is radically different now
it's wild
it's wild
January 4, 2026 at 3:28 AM
This hits especially hard as someone who emerged from English departments — listening to a lot of tenured profs contrast their perceived professional practices from coding/computation, at the active cost of their students… while they accrue debt.
W-i-l-d.
W-i-l-d.
When you happen to look down from a pier and a harbor seal emerges, looking up at you— before going underwater again, teasing me into waiting with my binoculars for half an hour…
—
Dear reader, the cute harbor seal did not return. 😔😔😔 instead I was treated to a smorgasbord of sea lions
—
Dear reader, the cute harbor seal did not return. 😔😔😔 instead I was treated to a smorgasbord of sea lions
January 2, 2026 at 10:24 PM
When you happen to look down from a pier and a harbor seal emerges, looking up at you— before going underwater again, teasing me into waiting with my binoculars for half an hour…
—
Dear reader, the cute harbor seal did not return. 😔😔😔 instead I was treated to a smorgasbord of sea lions
—
Dear reader, the cute harbor seal did not return. 😔😔😔 instead I was treated to a smorgasbord of sea lions
My partner and I were looking for the last few edge pieces of a jigsaw puzzle, but there were quite a few pieces to sort through.
So as a joke he took a picture of all the pieces asked it to find them, and… it found them 😂
Fun benchmark idea: visual needle in a haystack (if not already a thing)
So as a joke he took a picture of all the pieces asked it to find them, and… it found them 😂
Fun benchmark idea: visual needle in a haystack (if not already a thing)
January 2, 2026 at 1:47 AM
My partner and I were looking for the last few edge pieces of a jigsaw puzzle, but there were quite a few pieces to sort through.
So as a joke he took a picture of all the pieces asked it to find them, and… it found them 😂
Fun benchmark idea: visual needle in a haystack (if not already a thing)
So as a joke he took a picture of all the pieces asked it to find them, and… it found them 😂
Fun benchmark idea: visual needle in a haystack (if not already a thing)
Let’s goooooooo!!! 🤪😂
January 1, 2026 at 7:35 PM
Let’s goooooooo!!! 🤪😂
TIL, San Luis Obispo features the very first Jamba Juice.
Which in my historical mental model was the popularizer of the health smoothie? Idk if there was one before that?
Which in my historical mental model was the popularizer of the health smoothie? Idk if there was one before that?
December 31, 2025 at 10:16 PM
TIL, San Luis Obispo features the very first Jamba Juice.
Which in my historical mental model was the popularizer of the health smoothie? Idk if there was one before that?
Which in my historical mental model was the popularizer of the health smoothie? Idk if there was one before that?
Reposted by Jonathan Cheng
California’s state slogan should really be “nothing you do will make your state as beautiful as ours, surrender.”
December 30, 2025 at 3:36 AM
California’s state slogan should really be “nothing you do will make your state as beautiful as ours, surrender.”
Little does Katseye know that 21st century kpop stars 🤝 19th century horticulturists
December 30, 2025 at 12:31 AM
Little does Katseye know that 21st century kpop stars 🤝 19th century horticulturists
Will post prettier pics in a bit, but today I got to witness 3 park rangers coral a bobcat away from humans. Barely visible in this picture.
December 29, 2025 at 11:52 PM
Will post prettier pics in a bit, but today I got to witness 3 park rangers coral a bobcat away from humans. Barely visible in this picture.
The tour guide’s sermon on the perils of yellow journalism doesn’t quite have the intended effect when you follow it up with the afforded opulence.
December 28, 2025 at 10:42 PM
The tour guide’s sermon on the perils of yellow journalism doesn’t quite have the intended effect when you follow it up with the afforded opulence.
Monarch butterfly grove near Mismo beach! The sleeping clusters are very hard to spot.
Also bonus image of sleepy California beach towns being unreal.
Also bonus image of sleepy California beach towns being unreal.
December 28, 2025 at 1:24 AM
Monarch butterfly grove near Mismo beach! The sleeping clusters are very hard to spot.
Also bonus image of sleepy California beach towns being unreal.
Also bonus image of sleepy California beach towns being unreal.
Downside: flight delays, missed connections, stuck in sf for most of Christmas Day.
Upside: there’s an open dim sum restaurant tomorrow that looks out at the bay?
Har gow-ing exactly as planned. 😏
Upside: there’s an open dim sum restaurant tomorrow that looks out at the bay?
Har gow-ing exactly as planned. 😏
December 25, 2025 at 8:40 AM
Downside: flight delays, missed connections, stuck in sf for most of Christmas Day.
Upside: there’s an open dim sum restaurant tomorrow that looks out at the bay?
Har gow-ing exactly as planned. 😏
Upside: there’s an open dim sum restaurant tomorrow that looks out at the bay?
Har gow-ing exactly as planned. 😏
As I listen on the phone to my mom’s holiday rant about pronouns, all I can think of is the dumb joke “wow, weird that you hate pronouns so much when your favorite movie is Ben-HUR.”
Biting my lip, holding in a laugh, is probably the best case scenario in these situations.
Biting my lip, holding in a laugh, is probably the best case scenario in these situations.
December 25, 2025 at 12:38 AM
As I listen on the phone to my mom’s holiday rant about pronouns, all I can think of is the dumb joke “wow, weird that you hate pronouns so much when your favorite movie is Ben-HUR.”
Biting my lip, holding in a laugh, is probably the best case scenario in these situations.
Biting my lip, holding in a laugh, is probably the best case scenario in these situations.
It feels like Clair Obscur is doing a lot of heavy lifting — wrt building the next gen of US francophiles.
I’m trying to think of the last time I’ve seen a piece of media excite an American audience about France.
Amélie?
I’m trying to think of the last time I’ve seen a piece of media excite an American audience about France.
Amélie?
December 20, 2025 at 11:31 PM
It feels like Clair Obscur is doing a lot of heavy lifting — wrt building the next gen of US francophiles.
I’m trying to think of the last time I’ve seen a piece of media excite an American audience about France.
Amélie?
I’m trying to think of the last time I’ve seen a piece of media excite an American audience about France.
Amélie?
There’s a breakfast place in Ballard that has this *extremely* cozy bar in the back. And I now suddenly feel compelled to drink warm alcoholic beverages.
December 20, 2025 at 7:43 PM
There’s a breakfast place in Ballard that has this *extremely* cozy bar in the back. And I now suddenly feel compelled to drink warm alcoholic beverages.
Maybe one of the happiest little accidents is that I bought mussels instead of anchovies for my tomato/fennel soup.
And now I have something that tastes like shin ramyun broth.
And now I have something that tastes like shin ramyun broth.
December 18, 2025 at 3:21 AM
Maybe one of the happiest little accidents is that I bought mussels instead of anchovies for my tomato/fennel soup.
And now I have something that tastes like shin ramyun broth.
And now I have something that tastes like shin ramyun broth.
Spent the weekend learning about ml compilers. Namely aot autograd, inductor, etc
The fact that math efficiently routes to heterogenous hardware is really nothing short of a miracle
Need a word for the teleological experience of finally understanding a design and seeing the face of a cuda engineer
The fact that math efficiently routes to heterogenous hardware is really nothing short of a miracle
Need a word for the teleological experience of finally understanding a design and seeing the face of a cuda engineer
December 14, 2025 at 5:42 PM
Spent the weekend learning about ml compilers. Namely aot autograd, inductor, etc
The fact that math efficiently routes to heterogenous hardware is really nothing short of a miracle
Need a word for the teleological experience of finally understanding a design and seeing the face of a cuda engineer
The fact that math efficiently routes to heterogenous hardware is really nothing short of a miracle
Need a word for the teleological experience of finally understanding a design and seeing the face of a cuda engineer
What I’m trying to say is Monterey Seal Lofi encapsulates the focus-to-aww cycle once split between my computer and socal media on my phone.
Monterey Bay Aquarium Seal lofi:
1) amazing subgenre
2) surprisingly effective for focusing, spontaneously going aww, and then returning to work
youtu.be/r4fsBII2ZlU?...
1) amazing subgenre
2) surprisingly effective for focusing, spontaneously going aww, and then returning to work
youtu.be/r4fsBII2ZlU?...
2 Hours of Harbor Seals to Relax/Study To | Lofi Hip Hop | Monterey Bay Aquarium Krill Waves Radio
YouTube video by Monterey Bay Aquarium
youtu.be
December 10, 2025 at 11:07 PM
What I’m trying to say is Monterey Seal Lofi encapsulates the focus-to-aww cycle once split between my computer and socal media on my phone.
Monterey Bay Aquarium Seal lofi:
1) amazing subgenre
2) surprisingly effective for focusing, spontaneously going aww, and then returning to work
youtu.be/r4fsBII2ZlU?...
1) amazing subgenre
2) surprisingly effective for focusing, spontaneously going aww, and then returning to work
youtu.be/r4fsBII2ZlU?...
2 Hours of Harbor Seals to Relax/Study To | Lofi Hip Hop | Monterey Bay Aquarium Krill Waves Radio
YouTube video by Monterey Bay Aquarium
youtu.be
December 10, 2025 at 11:05 PM
Monterey Bay Aquarium Seal lofi:
1) amazing subgenre
2) surprisingly effective for focusing, spontaneously going aww, and then returning to work
youtu.be/r4fsBII2ZlU?...
1) amazing subgenre
2) surprisingly effective for focusing, spontaneously going aww, and then returning to work
youtu.be/r4fsBII2ZlU?...
Idk if this already exists, but if there was a model that was trained to estimate the amount of time taken to produce an image, producing [(image, time)].
And for different image gen models, you can see the estimated time of the outputs.
(Was musing different models of slop)
And for different image gen models, you can see the estimated time of the outputs.
(Was musing different models of slop)
December 8, 2025 at 9:19 PM
Idk if this already exists, but if there was a model that was trained to estimate the amount of time taken to produce an image, producing [(image, time)].
And for different image gen models, you can see the estimated time of the outputs.
(Was musing different models of slop)
And for different image gen models, you can see the estimated time of the outputs.
(Was musing different models of slop)