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.
Posit AI is composed of two pieces, Posit Assistant and Next Edit Suggestions. Posit Assistant is a data science and coding agent, sort of Claude Code meets Databot. Next Edit Suggestions are souped-up autocomplete.
It’s launching first in RStudio, but it’ll be coming to other platforms soon.
For the full overview, I’d recommend you read the announcement post on the Posit Blog from Joe, Nick, and Sara. I’ll share a few notes about each that stick out to me, though.
Posit Assistant
At this point, it’s relatively straightforward to get “LLM calling read/write/bash tools in a loop” working in a variety of frameworks. That “last 10%”, though—permissions and approvals, tool UI, context management and compaction, interruption, conversation forking—is a huge engineering lift. It’s also what makes the difference between a cool demo and something you want to reach for daily. We’ve really focused on making sure there aren’t any gaps compared to the coding agents we otherwise use day-to-day, and it’s become a really nice experience.
I’m maybe 50/50 between Posit Assistant and Claude Code at this point. I didn’t realize how much I was missing a “real UI” when using Claude Code until I had access to one. Easily copying + pasting, some nice markdown CSS, reactive tool UI… it’s a much better experience.
Next Edit Suggestions
I’ve got a post drafted right now that delves much deeper into our NES system and how it works; that will come out in the coming days. In short, we deployed an 8B model on a GPU so that it can respond with an edit prediction in 100-200ms. It can see the edits you’re making, some metadata about your R/Python environment, and where your cursor is. From that, it guesses what you’ll do next. It’s like other NES/supercomplete systems you may have used in other IDEs, but it doesn’t hallucinate dataset and column names and doesn’t f*** up Quarto documents. It was really fun to figure out how to make that work.
Broadly, I’m very interested in the mode of “I need to type a short snippet of text that is very obvious given the state of my IDE right now, and a reasonably smart model could guess what needs to be there 80% of the time.” Beyond just writing code in the editor, I think there are a few use cases for this sort of system that I’m excited to make happen in future releases of the service.