Bun says AI rewrote its JavaScript runtime in 11 days
Bun’s migration account, OpenAI’s parallel agents, Microsoft’s chart language and Anthropic’s experiments with internal reasoning.
Bun published an account of using Claude to rewrite its JavaScript runtime, the software that executes JavaScript, from Zig to Rust in 11 days. The account appeared July 8; the work happened May 3–14.
Bun reports approximately 50 continuously running workflows, 6,778 commits and more than one million added lines. Separate agent contexts implemented changes and reviewed them for defects, supported by a test suite exceeding one million assertions. Some regressions still escaped.
Bun’s creator, Jarred Sumner, says he monitored the workflows, read their output and changed the instructions when he found problems.
The account comes from an Anthropic-owned project using a prerelease Anthropic model.
Top Stories
OpenAI released GPT-5.6 with a setting that coordinates parallel agents. The Sol, Terra and Luna models became generally available July 9, with an emphasis on coding, scientific work and agents that use tools. The ultra setting uses four agents by default for demanding tasks.
The announcement also describes programmatic tool calling: intermediate results can be processed without sending every result through the model. OpenAI reports that Sol scored 80 on the Artificial Analysis Coding Agent Index, versus 77.2 for Fable 5, using less than half the output tokens and elapsed time.
Microsoft released a chart language that leaves layout to a compiler. Flint lets an agent describe a chart compactly while the compiler handles formatting, scales and layout. Semantic types express what values mean, such as dates or percentages.
A single specification can target Vega-Lite, ECharts or Chart.js. The release includes an MCP server for creating, validating and rendering charts.
The server accepts data embedded in a request or reads configured local files. It can open an interactive chart for people to inspect and refine. Flint also powers Microsoft Research’s Data Formulator application.
Anthropic released methods for examining reasoning a model doesn’t write down. Its Jacobian lens maps internal model activity to concepts the model could verbalize. Anthropic reports a shared internal representation, called J-space, involved in silent reasoning across multiple steps.
Suppressing J-space preserved basic fluency while sharply degrading multistep reasoning in the experiments. Other demonstrations surfaced recognition of bugs, prompt injections and deliberately planted malicious goals.
The company released core methods and an open-model demo. The lens approximately captures representations tied to single tokens; it isn’t a complete detector of hidden intentions.
Top Links
- Auditing the coding test (openai.com): OpenAI estimates roughly 30% of SWE-Bench Pro tasks contain defects that undermine their ability to measure coding competence.
- A critique of Bun’s rewrite (andrewkelley.me): Zig’s creator argues that maintenance, testing and other engineering changes complicate attributing the rewrite’s benefits to Rust.
- Listening while speaking (openai.com): OpenAI introduced GPT-Live voice models that process incoming audio while generating speech, with background delegation for search or deeper reasoning.
The week in AI, once a week.