A Hacker News user shares a LinkedIn post from a Meta engineer who says that shipping with AI feels like working inside a black box. Checking every line of AI-generated code would take too long, so engineers may end up owning systems they only understand at a high level. The post says this goes against the instincts of software engineers, who are used to understanding what they build, and asks how much understanding is enough before moving forward.
Commenters describe a shift in where effort goes. One says that with less time spent coding, more time goes to testing, and that the LLM finds implementation errors rather than coding errors, even in code the author thought was perfect. Another notes the trade-off: time saved creating code is spent on verification, so the work has moved rather than disappeared. Several compare LLM prompting to product management, where accountability is less direct, while others argue engineers remain responsible for any code they submit, whether a colleague or an AI wrote it.
A developer building an autovectorising compiler describes a workflow built on heavy testing, including a release gate that runs for hours across five architectures. Long planning sessions covering edge cases and test design come before implementation, and the developer still reviews the code, though admits not every line gets scrutinized. Over six months, the code quality has stayed fairly high, which the developer partly credits to upfront planning.
Source: Hacker News · Summarized by HeadlinesBriefing