Working Notes: a commonplace notebook for recording & exploring ideas.
Home. Site Map. Subscribe. More at expLog.
A somewhat busy week.
I'm finally getting started cleaning up my dotfiles: there are a couple of decisions I've made and will be implementing as I go:
$XDG_CONFIG_DIR
, with all the configs arranged in .config
instead of being spread out throughout my home directory. I need to check
how well that works in Emacs.Ctrl-'
into elisp much more annoying than having a cleaner org file document.
If I'm particularly honest with myself, my emacs configuration is nowhere near
dense enough to need literate programming.stow
or more
scripts to pull this off.fzf
and make it generally available. In the some token, make it very
easy to bootstrap the configuration on a new machine. I find myself hopping
between devservers very frequently these days, and spending an hour or two
configuring and installing is not something I really enjoy anymore.
.zsh_history
more intentionally: so far I have a 13 million line
zsh history (with duplicates) that I don't use these days, and a more reasonable
63k line history that I do at work. My personal history is only 4k lines. JM
strongly recommends using version control: I generally like that approach,
want to spend a lot of time maintaining a repo.
Refreshing my Rust knowledge by skimming Rust in Action: I actually seem to remember a lot, but I'll find out when I actually build the thing I'm reading this book for.
I also started playing with Zig: what I really want is a small language with minimal syntax that is still very expressive. Zig seems nice and compact and enjoyable, so I'll spend some time hacking with it.
I'm not sure why all (most?) lisps I've seen follow the Scheme / Common Lisp standards: the homoiconic syntax is decoupled from the other semantics. A small part of me wants to implement my own variant of C (at the very least, compatible with C) that has the syntax of a lisp. Probably emit LLVM IR and have hygienic macros. Then attach batteries by making it trivial to interop with C and Python and other languages.
I frequently find myself hitting the wiki and reading through different discussions. Both the style, speed of loading, minimalism, and depth of the discussion make this particularly enjoyable.
(Very rough notes from stanford lectures)
Open API
Manually inspect data
Predict probabilities of the next word
Basically massively multi task learning
small vs large model -- what they learn (?)
while overall loss improves
individual tasks can improve suddenly while others are saturated
some tasks may move correctly, some may show emergent behavior
can't predict what a large model will do from smaller models
Inverse scaling/U-shaped scaling
Repeat a string
Fix a quote repeating the same task
Plot scaling curves
— Kunal