Adil
@adilsoubki.bsky.social
NLP/ML. Math. Physics. Computer Science. PhD student. AI Enjoyer/Tolerater. https://NLP.RIP/
ever since i saw the acollier video where she complains that rich people are always talking about how they could have done physics i can't stop seeing it.
nothing wrong with wishing you had but if you have a billion dollars nobody is stopping you from just doing physics. join us.
nothing wrong with wishing you had but if you have a billion dollars nobody is stopping you from just doing physics. join us.
July 31, 2025 at 12:53 PM
ever since i saw the acollier video where she complains that rich people are always talking about how they could have done physics i can't stop seeing it.
nothing wrong with wishing you had but if you have a billion dollars nobody is stopping you from just doing physics. join us.
nothing wrong with wishing you had but if you have a billion dollars nobody is stopping you from just doing physics. join us.
I'm at ACL in Vienna if you want to say hi send me a message!
July 28, 2025 at 3:25 PM
I'm at ACL in Vienna if you want to say hi send me a message!
homeland security: +$43,800M
defense: +$113,300M
NIH: -$17,900M
NSF: -$4,609M
Rental Assistance: -$26,700M
someone who is good at the economy please help me budget this. my family is dying
defense: +$113,300M
NIH: -$17,900M
NSF: -$4,609M
Rental Assistance: -$26,700M
someone who is good at the economy please help me budget this. my family is dying
May 24, 2025 at 3:33 PM
homeland security: +$43,800M
defense: +$113,300M
NIH: -$17,900M
NSF: -$4,609M
Rental Assistance: -$26,700M
someone who is good at the economy please help me budget this. my family is dying
defense: +$113,300M
NIH: -$17,900M
NSF: -$4,609M
Rental Assistance: -$26,700M
someone who is good at the economy please help me budget this. my family is dying
idea: a website where you play werewolf (or some other deception game) but the werewolves are all ai so you have to figure out who is human and who isn't.
March 5, 2025 at 12:03 AM
idea: a website where you play werewolf (or some other deception game) but the werewolves are all ai so you have to figure out who is human and who isn't.
i feel the same way about podcasts as zoomers do about phone calls. i'm ideologically opposed to the idea of downloading informational content from itunes or whatever. i don't know popular phone numbers and i don't want to.
February 16, 2025 at 5:38 AM
i feel the same way about podcasts as zoomers do about phone calls. i'm ideologically opposed to the idea of downloading informational content from itunes or whatever. i don't know popular phone numbers and i don't want to.
if the universe is made entirely of harmonic oscillators then why don't i see more harmony in the world? for example, a box of oreos costs $9 in manhattan.
February 6, 2025 at 10:20 PM
if the universe is made entirely of harmonic oscillators then why don't i see more harmony in the world? for example, a box of oreos costs $9 in manhattan.
all functions are linear if you really think about it enough
February 6, 2025 at 5:26 AM
all functions are linear if you really think about it enough
I am not a prompt engineer. I am a slow engineer.
February 2, 2025 at 5:48 AM
I am not a prompt engineer. I am a slow engineer.
How would you even prove that one model was trained on another model's outputs as opposed to a similar distribution of data? I thought we couldn't even prove that a model was trained on a particular piece of text with high certainty.
January 30, 2025 at 2:50 PM
How would you even prove that one model was trained on another model's outputs as opposed to a similar distribution of data? I thought we couldn't even prove that a model was trained on a particular piece of text with high certainty.
i tried the deepseek distilled models on some theory of mind datasets and they think for ~2000 tokens before outputting the same answers as the base model lol
January 24, 2025 at 4:26 PM
i tried the deepseek distilled models on some theory of mind datasets and they think for ~2000 tokens before outputting the same answers as the base model lol
broke: using a learning rate of 1e-5
woke: using a learning rate of 1e-π²/2
woke: using a learning rate of 1e-π²/2
January 18, 2025 at 3:17 AM
broke: using a learning rate of 1e-5
woke: using a learning rate of 1e-π²/2
woke: using a learning rate of 1e-π²/2
the answer to so many problems is "put it in a vector space and (optionally) do something".
want to know what a word means to the computer? put it in a vector space.
want to know a particle's allowed energies? put it in a (special) vector space and solve for some eigenvalues.
want to know what a word means to the computer? put it in a vector space.
want to know a particle's allowed energies? put it in a (special) vector space and solve for some eigenvalues.
January 16, 2025 at 9:01 PM
the answer to so many problems is "put it in a vector space and (optionally) do something".
want to know what a word means to the computer? put it in a vector space.
want to know a particle's allowed energies? put it in a (special) vector space and solve for some eigenvalues.
want to know what a word means to the computer? put it in a vector space.
want to know a particle's allowed energies? put it in a (special) vector space and solve for some eigenvalues.
downloading one of those AI homework helper apps and asking it to help me with problems out of classical electrodynamics by jackson just to see what happens.
January 10, 2025 at 3:30 AM
downloading one of those AI homework helper apps and asking it to help me with problems out of classical electrodynamics by jackson just to see what happens.
should i be using uv? is python dev tooling finally unified under the iron fist of rust?
January 4, 2025 at 8:07 PM
should i be using uv? is python dev tooling finally unified under the iron fist of rust?
i don't install any dev tools that don't have uninstall steps in the documentation
January 4, 2025 at 7:57 PM
i don't install any dev tools that don't have uninstall steps in the documentation
I've been learning to pick locks and they have a belt ranking system (similar to karate) where once you get above essentially beginner (~blue+) there is no lock you can buy on Amazon that is difficult enough to pick.
You have to start buying locks from like enthusiasts on discord lol.
You have to start buying locks from like enthusiasts on discord lol.
December 9, 2024 at 2:07 PM
I've been learning to pick locks and they have a belt ranking system (similar to karate) where once you get above essentially beginner (~blue+) there is no lock you can buy on Amazon that is difficult enough to pick.
You have to start buying locks from like enthusiasts on discord lol.
You have to start buying locks from like enthusiasts on discord lol.
Alice and Bob are playing a game on a 6x6 grid. They place a piece on the bottom left and then flip a coin. If heads the piece moves one to the right. If tails it moves one square up. They do this until the piece falls off the board.
December 6, 2024 at 1:24 PM
Alice and Bob are playing a game on a 6x6 grid. They place a piece on the bottom left and then flip a coin. If heads the piece moves one to the right. If tails it moves one square up. They do this until the piece falls off the board.
Ok I now have access to multiple GPUs. Today is the day I figure out how to do distributed training.
December 5, 2024 at 6:14 PM
Ok I now have access to multiple GPUs. Today is the day I figure out how to do distributed training.
I used a simple neural net to approximate optimal square packings. It still needs some tuning but the results aren't too bad.
The optimal solutions for these are ~2.707 and ~3.707. Pretty close! Curious if (with some refinement) I can get some new packings this way.
The optimal solutions for these are ~2.707 and ~3.707. Pretty close! Curious if (with some refinement) I can get some new packings this way.
December 3, 2024 at 10:28 PM
I used a simple neural net to approximate optimal square packings. It still needs some tuning but the results aren't too bad.
The optimal solutions for these are ~2.707 and ~3.707. Pretty close! Curious if (with some refinement) I can get some new packings this way.
The optimal solutions for these are ~2.707 and ~3.707. Pretty close! Curious if (with some refinement) I can get some new packings this way.
I've been doing chess puzzles (I am bad) and no matter how many times I hear the phrase "if you see a good move look for a better one" I find the second best move every time.
I think I've been doing this in other parts of my life too. Ordering food, watching movies, picking research projects, ...
I think I've been doing this in other parts of my life too. Ordering food, watching movies, picking research projects, ...
December 3, 2024 at 1:04 PM
I've been doing chess puzzles (I am bad) and no matter how many times I hear the phrase "if you see a good move look for a better one" I find the second best move every time.
I think I've been doing this in other parts of my life too. Ordering food, watching movies, picking research projects, ...
I think I've been doing this in other parts of my life too. Ordering food, watching movies, picking research projects, ...
Sometimes I think my research slides might be a little too unprofessional.
December 2, 2024 at 7:48 PM
Sometimes I think my research slides might be a little too unprofessional.
Ok so some papers run a bunch of LMs on theory of mind tasks and say they got 90 something percent on their evaluation metrics the problem is solved and then others do it and they get like basically 0% so which one is it.
November 22, 2024 at 11:02 PM
Ok so some papers run a bunch of LMs on theory of mind tasks and say they got 90 something percent on their evaluation metrics the problem is solved and then others do it and they get like basically 0% so which one is it.
"prime" day doesn't even happen on a prime numbered date
November 18, 2024 at 11:21 PM
"prime" day doesn't even happen on a prime numbered date