Jason

In somebody else's words:

How to use R packages such as ggplot in Julia

Tidier a great idea, but tbh last time i tried it (2yrs ago?) the small diffs between it and og tidyverse drove me crazy. If thats still the case, it might be easier to just write R with rcall: avt.im/blog/archive.... IIRC the only downside here was the lack of support for the vscode plot pane.

part of this is that Pkg is so, so much better than pip etc, and another part of this is that there’s no AD / c++ backend conflict weirdness b/c Julia packages are mostly just pure Julia.

a very, very short Bayesian hmm in Julia

turinglang.org/docs/tutoria...

This example I added to the docs I think really shows off how well Julia packages tend to work together! If you think about the total call stack here it’s kinda crazy that I’ve never seen package conflict issues, but it’s the truth!

GitHub - gdalle/HiddenMarkovModels.jl: A Julia package for simulation, inference and learning of Hidden Markov Models. A Julia package for simulation, inference and learning of Hidden Markov Models. - gdalle/HiddenMarkovModels.jl

I also haven’t touched Julia a minute since I started working an R job, but in the past I’ve had a great time with hmms.jl + turing!

github.com/gdalle/Hidde...

I have never had a single issue installing a Julia library, and I’ve installed a lot of really random low star count stuff. What issues are you referencing?

Python is a different story tho— plenty of issues even with giant, popular packages.

wonder if the programming languages we have now are basically all we're ever going to get. it feels like the more people get used to coding via prompt, the higher the cost of switching to a language that doesn't exist in the training set.

this is also like 60% of my linkedIn lol

actual interviews have been mercifully light on this stuff but the total quantity of related slop online is not great

Partner is interviewing rn and I really feel this. The standard data science interview prep deck is full of the same trivia questions about, like, bagging v boosting or whatever.

really sad and unproductive way to interact with work!

1. I try an llm and am impressed by the improvements since I last tried an llm
2. I try harder tasks over the next few weeks until the LLMs start to produce garbage
3. “Damn, this sucks”

[wait four months]

1. I try an llm and am impressed by the improvements since I last tried an llm

feel like youd like the diaconis book "ten great ideas about chance"

I have one separating my half of the office from my partner — underrated feature is that it slowly accumulates doodles and nice messages / jokes from guests. At this point less than half those board is actually usable, but makes for good decor!

x <- val -> y

A more definitive answer on A/B testing run times if you're Baysian

dpananos.github.io/posts/2026-0...

lmao

in the reproducing kernel hilbert spaces approach, as the name implies, we... uh...

OpenAI Platform Explore developer resources, tutorials, API docs, and dynamic examples to get the most out of OpenAI's platform.

you can play around with the tokenizer here if you want to see this in action

platform.openai.com/tokenizer

that something is (mostly, IIRC) tokenization! Instead of seeing the string "a strawberry", the llm is trained on token ids [64, 101830].

"strawberry" alone tokenizes to [302, 1618, 19772]. the chatbot only learns statistical associations between tokens – it never sees the actual strings.

looks pretty cool 👀

lol

not sure what people do to get around this? just follow a bunch of people on twitter? live in sf and go to in person meetups? any thoughts @cameron.stream (basically the only AI person I follow on here lol)

so many terrible medium blogs, seo optimized articles from companies selling their services, etc.

meanwhile searching up topics in bayes stats will still surface some pretty great personal blogs the first few results. also,: no branded search!

one think i appreciate about gp / bayesian stats stuff being fairly niche is the lack seo spam. been looking at some llm stuff lately and its interesting that if i want to find the state of the art on, like, curating pretraining datasets or something this will be essentially impossible on google–

Calories in, calories out revisited “All models are wrong, but some are useful.” George E. P. Box A couple of months ago, I wrote about my experience “counting calories,” particularly about the accuracy of a v…

the dumbest possible cico model has pretty killer holdout accuracy (n = 1)

possiblywrong.wordpress.com/2015/01/01/c...

It even shows up in probability textbooks sometimes (Whittle)

There’s a refrain you see online re: DSA coding screens that, yeah, some of this stuff is useful– but not dynamic programming! that’s just there to weed ppl out, DP never shows up in prod code etc.

I’ve always found this odd b/c DP flavoured stuff is, like, the *only* DSA topic I’ve seen at work?

geolift and tidysynth both coming out meta / industry instead of academia might be relevant? not sure tho 🤔

It’s interesting because usually R has a pretty good doc / vignette culture, at least compared to Julia!

It’s pretty good but I still catch semi-frequent bugs (~1per200 lines)? also lots of weird unidiomatic defensive programming– guards that are impossible to hit, etc.

R may be uniquely bad here because of the low average quality of R code tho …