@mylifcc: OpenAI just disclosed a pretty outrageous engineering case study. With just 2 engineers + Codex + GPT-5.5, they complet…

OpenAI just disclosed a pretty outrageous engineering case study. With just 2 engineers + Codex + GPT-5.5, they completely rewrote a core storage service underlying ChatGPT from Python to Rust. Results: CPU efficiency improved 6x Memory efficiency improved 15x Average latency and tail latency significantly reduced Rust now handles 95% of production requests Even more outrageous: The original Python version had already peaked at handling 20 million requests per second. This might be a key case study in how AI Coding could fundamentally change the cost structure of software engineering. The service is called Habitat. It's OpenAI's online storage access platform, relied on by products like ChatGPT and Codex. When users log into ChatGPT, start a conversation, or read settings, it all likely involves this system behind the scenes. How big is the entire platform? Over 500 PB of data, more than 70 million requests per second, serving over 1 billion weekly active ChatGPT users. Interestingly, OpenAI had long known about Python's performance bottlenecks but deliberately waited a year before rewriting. Because there were more pressing engineering issues at the time: asyncio event loop latency, tail latency from Feature Flags, uneven connection pool loads, and burst traffic overwhelming downstream databases First get the architecture and stability right, then think about language migration. The Python version holding up to 20 million RPS was itself a pretty successful engineering feat. Notably, this refactor involved only 2 engineers, aided by Codex and GPT-5.5. Previously, rewriting a large production system across languages would typically require massive human effort for code migration, interface adaptation, testing, and validation. Now, AI Coding has the chance to dramatically cut those implementation costs. Also, OpenAI didn't abandon Python from the start. They first used Python to support rapid business growth and fix architecture issues, then migrated to Rust once the performance gains were clearly worthwhile. This path is very worth emulating: Prioritize development efficiency early on, tackle architecture bottlenecks mid-stage, and optimize runtime efficiency once scale is massive enough. In the past, a stable-running Python service meant that even if Rust performed better, it might not justify the heavy manpower for a rewrite. Now, Coding Agents could shift that economic calculus. In the future, competition in software engineering might increasingly hinge on who can use Agents to safely handle large-scale refactors, testing, and migrations. But the premise remains: Performance gains must be real, correctness backed by evidence, and migration risks controllable. Do you think we'll see more cases of "a few engineers + Agent" rewriting large-scale infrastructure in the future?