本文由 AI 分析生成
建立時間： 2026-06-25 來源： https://alisawuffles.github.io/blog/job-search/
Summary
A PhD graduate’s (NLP, UW, 6-year PhD on tokenization) retrospective on her industry job search for Research Scientist / Member of Technical Staff roles: timeline mechanics, interview category breakdown, preparation strategy (including Stanford’s CS336 course and from-scratch transformer implementation), and the harder, less-discussed negotiation and emotional-management phase.
一位 PhD 畢業生（華盛頓大學 NLP 領域，六年博士，研究分詞 tokenization）回顧其應徵 Research Scientist / Member of Technical Staff 職位的求職歷程：時程安排機制、面試類型分類、準備策略（包含 Stanford CS336 課程與從零實作 transformer），以及較少被討論、但更困難的議價與情緒管理階段。
Key Points
- Interview categories: ML coding (most common, PyTorch fluency required), general coding (LeetCode-style), technical discussion (depth of field knowledge), research discussion (own work + papers on CV), behavioral, math, and a job talk focused on a single line of work.
- Technical skills/knowledge are evaluated more heavily than research experience in interviews, though research experience is what gets you the interview in the first place.
- Recommends Stanford’s CS336 (Language Modeling from Scratch) as a way to organize scattered ML knowledge into one coherent picture before deep-diving individual topics.
- Practicing without AI assistance is explicitly recommended to mimic real interview conditions — “you will underestimate your reliance otherwise.”
- Negotiation is structurally asymmetric (candidate vs. recruiter market knowledge/skill) and offer timing has more flexibility than candidates assume; companies often explicitly invite negotiation.
- The emotional and social cost of the search (comparison to peers, unsolicited opinions, decision-making under incomplete information) is flagged as understated relative to the technical preparation advice typically given.
Insights
The explicit warning to “turn off AI assistance” while practicing for ML/coding interviews is a useful, non-obvious calibration point for anyone using Claude/ChatGPT heavily day-to-day — reliance is easy to underestimate precisely because the tool is good. The piece is adjacent to this vault’s secondary Claude-Code-ecosystem interest mainly through its meta-point about AI-assisted skill atrophy, not through robotics/VLA content.
Connections
Raw Excerpt
Make sure you are practicing coding with AI assistance completely off to mimic interview settings (you will underestimate your reliance otherwise)!