Anthropic, with a budget of $20,000 (Claude Opus 4.6), directed the Agent team to build a Rust compiler capable of compiling the Linux kernel. This signifies that AI programming has transcended the realm of auxiliary tools, possessing the ability to build complex systems at a very low cost, replacing experienced teams. If a human team were to implement such an engineering feat from scratch, the cycle would typically be measured in years. This experiment directly established a new benchmark for Coding Agents in handling highly complex and tightly coupled tasks. However, this engineering marvel is not solely due to the model's programming capabilities; its core lies in encapsulating the act of "writing code" within a rigorous automated testing and CI/CD system. To ensure the effective operation of this system, the experiment introduced GCC as an "oracle," locating errors by comparing standard output, transforming the open-ended creative challenge into a closed-ended verification task. Complementing this external verification, the choice of Rust constitutes an internal constraint, leveraging its strict type system to intercept errors during the compilation phase, compensating for the instability of the model-generated code. It is precisely this dual guarantee of "external comparison + internal constraints" that allows LLM to leverage the massive amount of compiler knowledge absorbed during its pre-training phase, completing the refactoring on the shoulders of human wisdom. This not only demonstrates capability but also reveals an irreversible shift in the cost structure of software engineering: the marginal cost of implementing details is approaching zero. When generated code becomes so cheap and massive, our focus will be forced to shift from "logical elegance" to "verification completeness." Future software deliverables may no longer be lines of code that are difficult for humans to maintain, but rather the Prompt strategies that drive it all and the test sets that define boundaries. Perhaps one day, the unit for measuring the workload of software development will no longer be the traditional "man-months," but rather "token consumption" and "prompt complexity." More perspectives👇

Thoughts after reading this: The experience of redirecting to external links is so much better now! 😂