Welcome to my personal notes!
good progress on pytorch game today
cool use case of llms is indexing/searching through files
September 8, 2023
need to automate all schoolwork
todos for homework bot:
for pytorch game, could write script that takes in pytorch doc link, and creates question based on it
can also add chat so user could ask questions specifically about that torch method
September 6, 2023
todos for homework project:
might just work on pytorch game this morning though
September 5, 2023
has anyone made a general use Photomath
could be a fun way to get into computer vision
could have a couple different use cases:
September 4, 2023
maybe pytorch game is better as a companion to the docs
not sure how to turn it from tutorial to game
September 1, 2023
gonna get first training run done tonight
todos:
August 29, 2023
need a way to tell model whose turn it is
otherwise it might generalize to a move suitable for both
can just make first dimension current player's stones, and other one is opponent
instead of 0 = white, 1 = black
August 28, 2023
made cool visualization for go game
just demonstration that I am getting correct board state at any given move
should be able to start on policy network tomorrow, since training data is in correct format now
August 27, 2023
https://buttondown.email/surya/archive/suffering-from-success-2028/August 26, 2023
todos:
finished zero to one
absolute banger
reading log is deployed
need to figure out whether new home page is good idea or not
published to website, may change in the future

August 25, 2023
too tired to do anything useful
gonna build reading list rn
August 24, 2023
https://github.com/featurecat/go-datasetfirst step of go bot is policy network (just CNN trained on pro games)
should be relatively straightforward
good paper to base policy network on
https://arxiv.org/pdf/1412.3409.pdfcould be cool to look at what different layers of cnn will capture
this paper addresses symmetry of the board in a cool way
“All Rhodes Scholars had a great future in their past”
zero to one is really good so far
August 23, 2023
https://arbital.com/p/bayes_rule/?l=1zqdespite being a little verbose, there is probably some alpha still on LessWrong
lots of value in understanding bayesian thinking and rationality
even if just to understand doomer args
really need to read SICP
have had the book for months now
no better time than the present
gonna get reading list up as a way to apply social pressure
ok im making new home page
done with new home page
not gonna push it till reading log is done
https://www.youtube.com/watch?v=iGLzWdT7vGc&t=1835sAugust 21, 2023
there are so many things to learn
need to read more about
coming up with smart/interesting ideas seems to work the same as coming up with good startups according to pg
first you need to get to the edge of innovation, then look for holes
to have interesting commentary about politics/culture/economics, you first need to grind your way to the edge
read more books, use twitter/instagram/distractions less
become a well-read autodidact builder
things to be working on:

first gonna make a go bot instead of poker
might be harder to make it really good, but is much more straightforward due to complete information
would also be cool to add to my website
https://www.nature.com/articles/nature16961seems like only new topic is tree search
saying that like it is trivial
August 19, 2023
for poker model, will probably have to start with heads up(only 1 other player)
need to encode all information about a hand, including bets up to this point
betting pre-flop vs post-flop are very different
basic probabilities are easy to compute
information required per hand:
things to decide:
data might be hard to find
http://poker.cs.ualberta.ca/irc_poker_database.htmlslightly sketchy dataset:
http://web.archive.org/web/20110205042259/http://www.outflopped.com/questions/286/obfuscated-datamined-hand-historieshttp://kevinwang.us/lets-analyze-pluribuss-hands/https://archive.ics.uci.edu/dataset/158/poker+handAugust 18, 2023
could keep randomness by sending two functions to run, that both compute answer
has anyone made poker bots
that they could then use for online games
this doesn’t seem that hard to build
problem with poker model is that you don't want to recreate average player
(what you'd get by training on all games)
instead, need some kind of decision tree like alpha go
might work if you get a bunch of pro games
but they would be playing against other pros, which you don't playing against other players
maybe v1 is just on pro games
later version could be alphaGo-style decision tree
August 16, 2023
I have gotten nothing done in the past ~2 weeks
Need to build/read/learn
August 9, 2023
who is making decentralized gpu compute
also would be fun to check out smallville agents
https://arxiv.org/abs/2304.03442Need to update site
Move current one to ‘old’ subdomain
Maybe just an html page? Looks more aesthetic + faster load time
react is too bloated
reading log page
new homepage
write levels for pytorch game
create three animation for game
Want it to look something like this
https://sigilwen.ca/August 8, 2023
solution is serverless functions(aws lambda)
making solid progress
can run pytorch code and get output
August 7, 2023
once set up at school ill resume the music generation development
for now just gonna work on pytorch game

gonna take physics or chem next semester
need to become more versed in hard sciences
CS classes are likely going to be fairly straightforward until sophomore year
problem is that python-shell cant install torch on every run
probably gonna have to set up instance that will run the code😔
August 3, 2023
music gen from relative attention is really inconsistent
half the time it sounds amazing, half the time it plays 400 notes in 3s
last day at internship
didn't realize that lambda labs is constantly out of GPUs
i need a way to figure out if the model is bad because it is a bad architecture or just if the model is too small
hoping its the former
fun tool would be coding puzzles you solve with pytorch
has anyone made that?
i.e. beginner puzzle would be make tensor of this shape that has these properties
might be hard to make more complicated puzzles though
complicated levels would be to get a model to a certain loss
something like kaggle but more game-like
codeMirror as editor
python-shell to run code on nodejs
https://codemirror.net/5/index.htmlhttps://www.npmjs.com/package/python-shellshould be a game though, not a test
something like this:
https://www.apple.com/swift/playgrounds/August 2, 2023
maybe I should do electrical or computer engineering
i need more hard science knowledge
maybe ill just take a bunch of physics/chem classes for fun
i have basically no understanding of electricity/energy
https://ai.meta.com/blog/audiocraft-musicgen-audiogen-encodec-generative-ai-audio/finished virtual assistant project yesterday
want to add visual aspect though(~wav2lip)
want to do something with tweet data
something like this:
https://github.com/effectiveaccelerationism/text-to-bangervisualizations are always fun to make
August 1, 2023
still think midi is the way to go for music
actual audio files don't have enough structure
was going to train model on cloud GPUs, but lambda had zero left
they are supposed to be the cheapest by far
still ~$2/hr though
todos:
July 31, 2023
https://github.com/facebookresearch/audiocrafthttps://karpathy.medium.com/software-2-0-a64152b37c35July 30, 2023
fine tunining did not work
base model probably massively overfit
todos:
relative attention is implemented
going to try using cloud gpu tomorrow
also found a good tts model
July 29, 2023
really wish there was linear algebra in high school
sometimes it feels like all of my opinions are just conglomerations of things i've heard
i know that is how learning works, but sometimes it still feels fraudulent
