Sara Altman
Simon Couch
Roundups

2026-03-13 AI Newsletter

2026-03-12
A collection of hexagons, each containing a different icon. On the left are three outlined heads: two with glasses, one a robot. The hexagons include a mall, a bridge, a raccoon, a Viking, and a goose.

External news

Anthropic’s dispute with the Pentagon

On March 4, the Pentagon designated Anthropic as a supply chain risk after a dispute over Anthropic’s refusal to allow Claude to be used for mass domestic surveillance or fully autonomous weapons systems. The supply chain risk designation is historically reserved for foreign adversaries, making this move an unprecedented action against a US company. 

On March 5, Anthropic released a statement about the designation, stating both that they do not think the supply chain risk designation was legitimate and that they do not expect it to affect the majority of their customers. On Monday, they filed two lawsuits against the Department of Defense. 

We’re keeping an eye on this story as Anthropic serves as the primary model provider for Posit AI and many of our developers use Claude models, but the designation’s broader implications for the tech industry remain unclear.

Model releases

Since our last newsletter edition, there have been several model releases:

  • OpenAI released GPT 5.4. Broadly, the model has a more natural and predictable feel than its recent predecessors. This release was also accompanied by an experimental 1M context window (up from 400,000 with GPT-5.3-Codex).
  • Google Gemini released Gemini 3.1 Flash-Lite. The Flash-Lite line, their name for a very small and cheap set of models, hadn’t received an upgrade since the 2.5 family.
  • The Qwen team released Qwen3.5, a series of models at the open-weights frontier. For those interested in running models on their own machine, the Qwen models are the top-of-the-line. That said, several lead researchers on the Qwen team stepped down shortly after the Qwen3.5 release. It’s unclear how these changes will affect future Qwen models and how long Qwen will maintain its lead in the open-weights space.

 

Posit news

AI in RStudio

Last Thursday, we announced AI functionality in RStudio, available through the Posit AI service. There are two components: Posit Assistant, a data science and coding agent that lives directly within RStudio, and Next Edit Suggestions (NES), advanced autocomplete functionality built for working with data. Simon has two blog posts about these features on his personal blog: How Posit AI’s Next Edit Suggestions work and Introducing Posit AI. 

Many people and teams across Posit have put a lot of work and thought into both Posit Assistant and NES to make them useful and trustworthy for people who work with data. If you’d like, you can try out Posit AI with a free trial.

nesevals

The NES evals that the team used to develop the NES functionality mentioned above are open-source! You can learn how it works in Simon’s blog post and check out the eval here. 

Wes McKinney’s agentic engineering projects and workflow 

In a recent blog post, Rich Iannone walks through five recent projects from Wes McKinney and explores how his use of coding agents and workflow setup helps him to be so productive. 

 

Learn more

  • A recent study from Anthropic found that junior engineers who used AI assistance to learn a new coding skill understood the skill worse than those who coded “by hand.” Relatedly, Dr. Cat Hicks released a learning-opportunities Claude Skill that helps coders pause for “deliberate skill development during AI-assisted coding.”
  • Anthropic and OpenAI both recently announced programs providing free premium subscriptions for maintainers of large open source projects.
  • The company Taalas hard-wires individual AI models into custom silicon for dramatically faster and more energy-efficient inference. A demo shows a deployment of Llama 3.1 8B responding at the (very fast) speed of 15k-20k tokens/second, nearly ten times faster than the current state-of-the-art for that model on a GPU.
Tags: AI AI Newsletter