LLMs are getting much better at interpreting counterintuitive plots
This post is a cross-post of a post on the Posit Open Source Blog, co-written with Sara Altman.
Simon P. Couch
A data science blog
Latest posts
LLMs running on my laptop can drive coding agents now
In December, I wrote a post called “ Local models are not there (yet) .” It concluded like so:
How Posit AI’s Next Edit Suggestions work
I just loaded some data from the forested package into my R environment. It has a bunch of measurements of forest attributes across Washington State:
Introducing Posit AI
Today we released Posit AI , an AI service for data scientists. This was a huge effort that spanned many teams over several months, and I’m really excited to have it out in the world. It’s really good.
Electricity use of AI coding agents
Throughout 2025, we got better estimates of electricity and water use of AI chatbots. There are all sorts of posts I could cite on this topic, but a favorite is this blog post from Our World in Data’s Hannah Ritchie. On the electricity front:
chores 0.3.0 and local LLMs
The tl;dr:
Local models are not there (yet)
I understand the appeal of local models. Using coding agents like Claude Code or Codex, it’s not difficult to rack up a hundred dollars of usage in the course of a work week. Besides the price, if you’re working with sensitive IP or confidential data, you need to really believe that providers like Anthropic and OpenAI can be trusted with your data. And then, there’s evil billionaires. What if you…
Analyzing my music listening data with Databot
It’s Spotify Wrapped season, which means that everyone I follow on instagram is posting screenshots on their stories and I need to export my iTunes Library metadata as an .xml file and analyze it with the tidyverse. (If you’re new here, I do a little group_by() %>% summarize() on my own music listening data each year .)
When plotting, LLMs see what they expect to see
This post is a cross-post of a post on the Posit Blog, co-written with Sara Altman.
side::kick(), a coding agent for RStudio
I’m excited to share side::kick() , an experimental coding agent for RStudio users, built entirely in R. It can interact with your files, talk to your active R session, and run code.
I’m… writing a newsletter?
Between Positron Assistant, Databot, ellmer, chatlas, and their offshoots, there’s been a lot of LLM-related news coming out of Posit in 2025. Many folks across the organization felt that it was hard to keep up with new developments in the space, both internally and in the wider world. With this in mind, my colleague Sara Altman and I started working on an internal newsletter in June; released…
I was wrong about tidymodels and LLMs
One of my most visceral memories of my first interactions with LLMs was asking that first release of ChatGPT in late 2022 to write code to fit a linear regression with tidymodels. The model hallucinated a function tidymodels::install_tidymodels() again and again. That function does not exist. If it did, there’d be some serious chicken and egg happening. This thing was goofy.
R and the Model Context Protocol
This is a cross-post from tidyverse.org.
How I’m using Claude Code to write R code
A couple months ago, I wrote a bit about how I was using Claude Code to help me write R code. At the time, I mostly just shared my impressions of working with the tool and some prompting tips. In the month or two after I wrote the post, my usage waned; I was mostly back to using LLMs only for shorter, more narrowly-scoped tasks. A few weeks ago, though, we put together some tooling that has helped…
Kimi K2 and R Coding
It was a hoot and a half of a weekend in the LLM world. A company I hadn’t heard of called Moonshot AI released a model called Kimi K2 . From 30,000 feet:
Claude 4 and R Coding
Claude 4 dropped on Thursday! Given that Claude 3.7 Sonnet is my daily driver LLM for R coding, I’ve been excited to poke at it.
Evaluating Gemini 2.5 Flash on R coding tasks
Google’s preview of their Gemini 2.5 Pro model has really made a splash . The model has become many folks’ daily driver, and I’ve started to see “What about Gemini?” in the comments of each of these blog posts if they don’t explicitly call out the model series in the title. Yesterday, Google announced an update of the preview for Gemini 2.5 Flash, a smaller and cheaper version of 2.5 Pro.
Evaluating the new Gemini 2.5 Pro update on R coding
The title line of Google’s release post on the newest Gemini 2.5 Pro release is “even better coding performance.” Reading this, I was curious whether we’d see a notable increase in performance compared to the last generation on R coding tasks; in an earlier post , I saw that the March release of Gemini 2.5 Pro was a contender with Claude 3.7 Sonnet on An R Eval , a dataset of challenging R coding…
Evaluating o3 and o4-mini on R coding performance
48 hours after the drop of the GPT 4.1 series of models, a trio of non-reasoning models focused on “real-world developer needs,” OpenAI dropped another set of models, o3 and o4-mini. These two models are the latest generation of thinking models from OpenAI, and they form the backbone of Codex , a new Claude Code competitor from OpenAI. In short, OpenAI wants market share among developers.
How good are the GPT 4.1 models at writing R code?
Yesterday, OpenAI dropped a new series of models called GPT 4.1, 4.1 mini, and GPT 4.1 nano. This line from their release post, specifically, caught my eye: