The author challenges the assumption that large language models (LLMs) automatically lead to more software optimizations. While LLMs can reduce the cost of writing code, they do not necessarily change the underlying business priorities or the complexity of maintaining production systems.
In many organizations, software development is viewed as a cost center rather than a profit driver. Management often demands faster project delivery due to LLM adoption, which leaves little room for performance-focused tasks that were previously deprioritized.
Implementing optimizations in production environments remains difficult. The author cites the example of a Bun fork of the Zig compiler, where parallel semantic analysis was implemented without necessary language changes, leading to non-deterministic behavior and errors.
The author also notes that performance gains in some projects, such as pgrust, appear to be heavily tuned for specific benchmarks like ClickBench. This raises concerns about whether such optimizations are truly architectural or merely tailored to specific test cases.
Ultimately, the author argues that the decision to optimize software depends on priority rather than just the cost of implementation. If a task was not a priority before the advent of LLMs, it is unlikely to become one simply because coding speed has increased.





