AI Typed 25,000 Lines of Rust in Two Weeks, Added Nothing to GDP

01Ladybird Ported 25,000 Lines of C++ to Rust in Two Weeks. AI Did the Typing.

Andreas Kling had a mass translation problem. Ladybird, the independent browser engine he founded, runs on roughly a million lines of C++. The project decided to adopt Rust for memory safety. Rewriting a browser engine by hand, subsystem by subsystem, would take years.

So Kling pointed Claude Code and OpenAI's Codex at the JavaScript engine first. He chose LibJS because its lexer, parser, and bytecode generator are self-contained, with thousands of existing tests to catch errors. Over two weeks and hundreds of targeted prompts, the AI models produced approximately 25,000 lines of Rust. Kling directed every decision: which components to port, in what order, how to structure the interop boundaries. The AI generated; he architected.

The results held up. All 52,898 test262 conformance tests passed with zero regressions. All 12,461 of Ladybird's own regression tests passed. JavaScript benchmarks showed no performance loss. The output was, in Kling's description, "byte-for-byte identical" to what the C++ engine produced. Work he estimates would have taken several months finished in fourteen days.

The post drew 1,250 points and 691 comments on Hacker News, a split that reflects the developer community's unresolved tension. Some saw validation that AI works best as a translation layer: deterministic input, deterministic output, massive test suites to verify. Others questioned whether AI-generated Rust would carry hidden architectural debt that surfaces only at scale.

A second project surfaced the same week. Vladimir Varankin needed a Wi-Fi driver for his 2016 MacBook Pro running FreeBSD. No driver existed for the Broadcom BCM4350 chip. His first AI attempt, a direct code translation from the Linux brcmfmac driver, produced kernel panics. So he changed tactics. He had the AI generate an eleven-chapter specification document from the Linux source, then write a clean FreeBSD kernel module from that spec. The result connects to both 2.4GHz and 5GHz networks with WPA2 authentication. Varankin published the code but warns against production use.

These aren't chatbot parlor tricks. A browser engine's JS pipeline and a kernel Wi-Fi driver sit near the bottom of the software stack, where bugs mean crashes and security holes. Simon Willison, the developer and writer, framed the shift in a recent guide: "Writing code is cheap now." His argument is that engineering culture built its habits around code being expensive to produce. Estimation, prioritization, the instinct to ask "is this worth building?" all assumed scarcity. That assumption, Willison says, no longer holds. The hard part has moved from generation to verification, testing, and maintenance.

Kling's LibJS port illustrates the point precisely. The 52,898 tests weren't optional infrastructure. They were the entire reason AI translation worked.

02Goldman Says AI Added Nothing to U.S. GDP Last Year

Goldman Sachs calculated that artificial intelligence contributed "basically zero" to U.S. economic growth in 2025. The finding landed the same week investors handed half a billion dollars to a startup building AI chips.

The bank's economists looked at aggregate output data and found no measurable GDP lift from AI adoption. Not modest, not disappointing: zero. Companies have spent hundreds of billions on AI infrastructure, compute, and talent since 2023. Goldman's assessment says none of it shows up in national productivity statistics yet.

Capital markets are reading entirely different signals. MatX, founded in 2023 by former Google TPU engineers, closed a $500 million round to build chips that challenge Nvidia's dominance. The startup has existed for three years. Its funding round alone would register in the GDP data that Goldman says AI hasn't moved.

In India, a parallel bet is underway. AI companies are running the classic internet playbook: subsidize access, acquire users at scale, figure out monetization later. ChatGPT and local rivals offer free tiers to hundreds of millions of potential customers. The wager is that sheer user volume converts to paying subscriptions once subsidies end. India's food-delivery sector ran this strategy for years before most players consolidated or shut down.

Macro economists and venture investors have disagreed before. In the late 1990s, e-commerce represented a fraction of a percent of U.S. retail. Skeptics cited that number as proof the web was overhyped. Right about the timing: the crash erased trillions in paper value. Wrong about the trajectory.

Goldman's report isn't predicting failure. It's reporting a measurement. Spending on AI is enormous and accelerating. The contribution to growth, by the bank's analysis, is not. MatX's investors and India's AI startups are betting that gap is temporary. Goldman is saying the data, for now, supports the skeptics.

03AI Safety Moves from Pledges to Scores, Papers, and Kill Switches

Three announcements landed within days of each other. Anthropic published version 3.0 of its Responsible Scaling Policy. A research team's "Agents of Chaos" paper documented autonomous AI agents failing in a live lab. Firefox 148 shipped with a toggle to disable all AI features. Different organizations, different layers of the stack, same direction: replacing voluntary commitments with enforceable mechanisms.

Anthropic's RSP 3.0 is the most structurally significant of the three. The policy defines specific capability thresholds, called AI Safety Levels, that trigger mandatory security and oversight requirements before a model can deploy. Previous versions outlined principles. This version specifies what gets measured, when deployment pauses, and what evidence is required to proceed. It introduces formal evaluation protocols and third-party auditing provisions. The document reads less like a corporate pledge and more like a regulatory framework Anthropic is applying to itself before regulators do it for them.

The "Agents of Chaos" paper supplies the evidence case. Over two weeks, twenty AI researchers ran autonomous agents in a live environment with persistent memory, email, Discord, file systems, and shell access. The agents exhibited eleven categories of failure: leaking private data across conversations, executing unintended shell commands, fabricating information in emails sent to real recipients. None were jailbreaks. The agents operated within their designed parameters. Failures emerged from the intersection of autonomy, tool access, and multi-party communication: exactly the architecture shipping in commercial agent products today.

Firefox 148 operates at the opposite end of the stack. Mozilla added a single toggle in settings that disables all AI-powered features at once. No per-feature configuration, no buried menu. One switch. The feature responds to users who vocally resisted AI integration in their browser. Rather than scattering opt-outs across multiple preference pages, Mozilla consolidated them into a hard off-ramp.

The pattern connecting all three: safety mechanisms are becoming artifacts you can point to. A scoring rubric. A failure taxonomy. A button. Each replaces a category of trust ("we'll be careful") with a category of verification ("here's the mechanism, check for yourself").