One Year from Big Tech SWE to AI Infra: Reading Source Code Is the Dividing Line, and Switching Careers Isn't Worth It for Everyone

Yichi Zhang posted on X a complete timeline of his first year moving from Big Tech SWE to AI Infra. The original is in English; he notes that the initial Chinese version was handwritten, as he wanted to preserve an honest and personal perspective as much as possible.

SGLang is an open-source LLM inference framework maintained by the LMSYS org community. For readers without background, AI infra can be roughly understood as an engineering layer that makes models run efficiently on real hardware, involving inference serving, scheduling, KV cache, parallelism, quantization, etc.

One-year timeline

He graduated in May 2025 and joined Big Tech in July. Life at Big Tech was different from what he had imagined. Tasks were mostly assigned from above, the technical work was repetitive, and there was little room for creativity. One evening three months after joining, he began to doubt the path he had laid out for himself.

In October 2025, he started learning AI infra fundamentals. His ML background was limited, but math and lab experience helped. There was no ready-made course, and it took him a long time to judge what was worth learning. Early learning notes were posted on Xiaohongshu; later he felt they weren't good enough and hid them, not wanting to mislead people.

In January 2026, he felt he had read enough introductory materials and blogs, but understanding concepts and writing code are two different things. So he opened the SGLang codebase and read it deeply. Reading source code exposed him to many concepts and optimization techniques he had never heard of: CPU-GPU overlap in the scheduler, CUDA Graphs, HiCache, etc. For newcomers without context, these terms can feel overwhelming.

From March to April 2026, early small tasks helped him build trust in the community. The Guitar Cat LLM account known to many in the community brought him into the SGLang-Omni project. After that, he began doing substantial development.

He calls the year after October 2025 the highest-ROI year of his life. It was also the year he became more aware of the limits of his personal knowledge.

Differences between Big Tech and startups

When a netizen asked what surprised him most after changing jobs, Yichi replied: Big Tech is standardized, with clear boundaries of responsibility, and you know exactly what to do; startups are the opposite—agile, fast, and broad in scope.

He also doesn't recommend that everyone switch. His previous job allowed him to work 9-to-5 / 965, play board games or hike after work. AI is highly competitive and changes quickly, and every minute feels fully occupied. For him, the meaning of work and its value to humanity are what drive him.

Why open source matters

He acknowledges that AI, as a productivity tool, is widening the gap between those with resources and those without. People who can access the most advanced models and supporting tools can do the work previously done by a hundred people; those who can't can only use a lower tier, or may not even know how much AI can raise productivity. That's why open source matters. He cites Arasaka from Cyberpunk 2077, saying that if powerful tools remain only in the hands of a few companies or the top 0.1%, Arasaka may not be confined to fiction.

A few questions in the comments

A netizen with 15 years of infra experience said that Kubernetes has been foundational since 2016, and now they work on serving engines, orchestrators, models, KV cache, and inference. Someone summarized the path as jumping into the codebase. An undergraduate also asked whether to read SGLang/vLLM first or practice backend first. There is no one-size-fits-all answer to this question, but the path given in the original is very clear: after fundamentals, source code and community tasks are the dividing line.

Original link: Yichi Zhang's tweet

Article link: From Big Tech SWE to AI Infra (Part 2)