Bun Rewrites Its Entire Codebase from Zig to Rust with Claude Fable 5 — More Than 1 Million Lines in 11 Days
Bun, the JavaScript runtime and toolkit, has fully migrated its programming language from Zig to Rust. What makes this remarkable is that the bulk of the work was carried out by Anthropic's Claude Fable 5. Reportedly, more than one million lines of code were rewritten in just 11 days.
Fable 5 is a coding-focused agent offered by Anthropic, capable of autonomously executing large-scale codebase transformations. By leveraging the tool, the Bun team dramatically shortened a migration that could have taken human engineers months.
This case shows that AI coding agents are evolving from mere assistive tools into "executors of large-scale software engineering projects." That said, humans led the quality verification of the generated code and the architectural design, making this a notable milestone in one model of AI-human collaboration.
Fed Appoints Marc Andreessen as AI Economy Advisor — Conflict-of-Interest Concerns Arise
Kevin Warsh, chair of the U.S. Federal Reserve Board, has appointed venture capitalist Marc Andreessen as an advisor to assess the impact of AI on the economy. Warsh has positioned AI as a "significant deflationary force."
At the same time, Andreessen Horowitz (a16z), which Andreessen co-founded, has made substantial investments in AI companies, and concerns over conflicts of interest have been raised. Whether AI suppresses or accelerates inflation is an issue directly tied to investors' financial stakes.
It is noteworthy that the Fed has begun seriously examining the macroeconomic impact of AI, but the choice of advisor is likely to keep facing questions about whether independence can be ensured.
Grok Now Handles Conversations Inside Tesla Vehicles — Available in Japan
On July 10, Tesla Japan announced on its official X account that the conversational AI "Grok" is now available in Tesla vehicles in Japan. Users can configure navigation and check routes by voice.
Using it requires software version 2026.20 or later as well as a subscription to "Premium Connectivity" (a paid in-car connectivity plan). As AI assistants continue to spread into car interiors, attention is focused on how Grok will be put to use within Tesla's ecosystem.
Seeing Through AI-Written Code — 5 Techniques for Detecting "Confident Lies"
A practical approach to reviewing AI-generated code has been drawing attention on Qiita. The article points out a trap: because AI-generated code often looks cleaner than sloppy human-written code, reviewers' guard actually drops.
The article introduces five techniques grounded in AI-specific mistake patterns: (1) verify existence — check that APIs and functions actually exist; (2) probe boundaries — confirm behavior on edge cases; (3) ask "why" — question the design intent; (4) run it — verify runtime behavior; (5) try deleting — check for unnecessary code.
The argument that traditional code review criteria need to be updated for AI, because human mistakes and AI mistakes follow different distributions, offers practical insights for workplaces where AI coding tools are becoming widespread.
GPT-5.6 Usage Limits Hit "Too Soon" — The Problem Isn't the Plan, It's Context
Regarding the GPT-5.6 family (Sol, Terra, Luna) released by OpenAI on July 9, developers on the $200-per-month Codex plan have been reporting that they hit the usage limit after just one or two agent runs on Sol.
An article on Qiita argues that the root cause lies not in plan restrictions but in "context consumption." With agentic task execution, context accumulates with every tool call and intermediate reasoning step, so you reach the cap in no time. Rather than benchmark scores, usability in real-world development workflows has emerged as the real pain point.
---