I built a GitHub issue classifier for Apache Arrow issue language using {ellmer} - super simple and almost 100% accuracy. Blog post: niccrane.com/posts/llm-issue-triage/
#rstats #ai #llms
In the shower thinking "wouldn't it be cool to combine LLM tool calls and have them run code but in a constrained way" & then "it needs some kind of intermediate representation; how would we validate whatever it produces?" & then realised my idea wasn't novel & just the motivation for text-to-sql 😅
I remember at posit::conf last year there was mention of posit::conf Europe 2026 - anyone know if this is still a thing? #rstats #positconf #posit
Huge thanks to the organisational team for putting on such an excellent event! 💜🌈
Excited for all of the talks tomorrow, check out the schedule here if you havent' seen it! conference.rainbowr.org/schedule.html
Whew, and it's done! Thanks to everyone who came to my RainbowR workshop on LLMs for Data Analysis in #rstats! First time with that content in front of an audience, so I appreciate the excellent questions folks asked (and double thanks to everyone who filled in the feedback forms!)
"Working with agents is a lot more productive, but a lot less fun." Charlie Marsh on the weird world of building software right now. Full conversation on The Test Set.
Sounds interesting, how well does it work for R code?
It's still experimental, so potentially some rough edges, but I think it's a great example of making sure the LLM benefits are tempered with what actually makes sense for *people*.
Instead of generating a load of comment, you get suggestions one at a time, which you can then choose to accept or reject, before it moves on to the next suggestion. It generates suggestions as it goes, so if you accept some changes but reject others, its suggestions change on the basis of the code.
There's promise in using LLMs for code review, but it's tricky things to make sure it's not overwhelming.
I was looking at this new experimental package by Simon Couch and I really love how it allows you to review code iteratively. #rstats #ai #llms
github.com/simonpcouch/...
Short musings on "cognitive debt" - I'm seeing this in my own work, where excessive unreviewed AI-generated code leads me to lose a firm mental model of what I've built, which then makes it harder to confidently make future decisions simonwillison.net/2026/Feb/15/...
Should be there shortly!
Let's talk contributors! This release saw 44 contributors to the codebase! 38 worked on the C++ library, 3 on the R 📦, & 3 on both. 23 people made their first contribution! 🎉
Thanks to everyone who was involved!
Writing partitioned datasets on S3 no longer requires ListBucket permissions; useful if you have write-only access to a bucket.
The following reproducible example: library(arrow) library(dplyr) library(stringr) df <- arrow_table(x = c("Apache", "Arrow", "23.0.0")) df |> filter(str_ilike(x, "ARROW")) |> collect() #> # A tibble: 1 × 1 #> x #> <chr> #> 1 Arrow
We've added support for stringr::str_ilike() for case-insensitive pattern matching.
We're excited to announce the release of {arrow} 23.0.0 🏹📦
Here's a roundup of the new features and changes in a 🧵
Full details can be found at arrow.apache.org/docs/r/news/
#rstats #apachearrow
I mean, you could say the same thing about any R function; just a toy example - feel free to replace it with something more useful! 😉
Yeah, there's some irony in the fact that I randomly chose that specific example, and then the results even showed the new features including the web fetch thing making my example redundant! 😆 I shall have to think up a new example for when I'm teaching, but YAY, awesome new feature! 🎉
First part of code - full code can be found at https://gist.github.com/thisisnic/eae09dbd4594e2cff75d156a8bab3f59
First part of code - full code can be found at https://gist.github.com/thisisnic/eae09dbd4594e2cff75d156a8bab3f59
Tool calling lets LLMs run R functions; in this example I let an LLM ask my R session to check the latest {ellmer} updates by scraping the news page and when I ask the LLM "what's new in ellmer?", it works with what comes back.
{ellmer} website: ellmer.tidyverse.org
#rstats #llms #ai #datascience
Line chart showing percent correct on the y-axis and three conditions on the x-axis: Baseline, Intuitive, and Mocked. Three lines represent GPT-5.2, Claude Opus 4.5, and Gemini 2.5 Pro. All three models score between 93-98% on baseline, then drop on intuitive and mocked conditions. All three perform the worst on the mocked condition.
More on LLMs and plot interpretation: they do fine in normal conditions, but struggle when the plot conflicts strongly with their priors.
@simonpcouch.com and I investigated why and what might help: posit.co/blog/llm-plo...
Code in which text from wikipedia article being passed into chat_structured method to extract dates and events
I love "structured output" as a way of extracting data from text as data frame. 🎯
Image shows using the {ellmer} package and how using type_array(type_object(...)) automatically returns a data frame in R 🔧
{ellmer} website: ellmer.tidyverse.org
#rstats #llms #ai #datascience
posit::conf(2026) call for talks is now open! If you're an #RStats or #Python user, have a great DS workflow to share, or have some lessons learned, we'd love to hear from you.
🔗 posit.co/blog/posit-c...
Oh, fascinating. I'm imagining that it stops it from being over-reliant on the retrieved information or interpreting it too literally maybe? Would love to hear more about how that ends up working out!
vitals logo (a teddy bear with a stethoscope), above output from running evaluation on a dataset, with 2 correct responses from the LLM and 1 incorrect response
RAG doesn't guarantee reliability 🤔
Built a RAG chatbot for a course & tested the same question+model three times with {vitals} - one run got it wrong. ❌
This is why evals matter: catch inconsistencies you'd miss manually. 🔍
vitals.tidyverse.org
#rstats #llms #ai #datascience
Excellent post about Claude Code's actual energy usage 💡
Most AI energy posts only look at single queries. Simon breaks down full coding sessions - much higher, but still about the same as running a dishwasher once a day.
simonpcouch.com/blog/2026-01-20-cc-impact/
#rstats #ai #llms
main image for Scaling Up Data Analysis in R with Arrow R Consortium webinar
R Consortium webinar: Scaling up data analysis in R with Arrow. Learn larger-than-memory workflows, why Parquet matters, and where DuckDB fits—w/ Dr Nic Crane (Arrow R maintainer; Apache Arrow PMC). Register: r-consortium.org/webinars/sca... #rstats #arrow @niccrane.bsky.social
Speak at posit::conf(2026) and share your R & Python stories!
Accepted speakers get:
✨ Travel & lodging help
✨ Free conference pass
✨ Professional coaching
Apply by Feb 6 to join us Sept 14-16 in Houston, TX!
Submit here: pos.it/conf-talk-2026
#positconf2026 #rstats #pydata
Building LLM training materials for R users (inc a RainbowR conference workshop) and want to know what topics people care about. ~10 min survey, results shared openly!
forms.gle/uAtVwpRhKFVU...
Boosts appreciated 🤖 #rstats #llms #ai
