PaperTrace

Job Hunt

LeetCode resources and ML interview prep ā€” for SWE and research roles.

āœļø

Notes on the Job Hunt

Timing, information, connection & mindset

ā†’

LeetCode / åŠ›ę‰£

šŸ§©
NeetCodeEnglish

Curated 150-problem roadmap with video solutions. The best structured path for algo interview prep.

ā†—šŸµ
ēµčŒ¶å±±č‰¾åŗœBē«™

Top Chinese LeetCode explainer on Bilibili. Deep dives into patterns: sliding window, monotonic stack, trees, DP, and more.

ā†—šŸ’»
LeetCodePlatform

The primary platform. Use LeetCode CN (leetcode.cn) for Chinese UI and local contest.

ā†—

Job Openings / 岗位

Career pages for major AI labs ā€” US and China.

šŸ‡ŗšŸ‡øAnthropicā†—šŸ‡ŗšŸ‡øOpenAIā†—šŸ‡ŗšŸ‡øGoogle DeepMindā†—šŸ‡ŗšŸ‡øMeta AI (FAIR)ā†—šŸ‡ØšŸ‡³DeepSeekā†—šŸ‡ØšŸ‡³å­—čŠ‚č·³åŠØā†—šŸ‡ØšŸ‡³é˜æ里通义 / Qwenā†—šŸ‡ØšŸ‡³Moonshot AI (Kimi)ā†—

Community Posts

Submitted by the community. Want to post? Reach out.

+ Submit a post
UrgentDeepSeekĀ·Research Intern ā€” Reasoning & RL
InternshipHangzhoušŸ‡ØšŸ‡³

Working on LLM reasoning, post-training, and RL-based alignment. Background in math/CS preferred. 3ā€“6 months, potential for return offer.

šŸ“© å¾®äæ”ꊕ递ļ¼Œč”ē³»å…¬ä¼—号后台

2025-04
å­—čŠ‚č·³åŠØ Ā· SeedĀ·LLM Pre-training Intern
InternshipBeijingšŸ‡ØšŸ‡³

Joining the Seed foundation model team. Focus on data pipeline, training efficiency, and scaling experiments. Strong Python + distributed training background needed.

šŸ“© 邮件ē®€åŽ†č‡³ seed-intern@bytedance.com

Apply ā†—2025-04
AnthropicĀ·Research Engineer Intern
InternshipSan Francisco, CAšŸ‡ŗšŸ‡ø

Summer internship on the interpretability or alignment team. Strong ML fundamentals required. US work authorization needed.

Apply ā†—2025-04

ML Interview / 面ē»Ā·å…«č‚”

Core concepts with rigorous answers ā€” for LLM researchers and practitioners.

Transformer & AttentionTraining & OptimizationArchitectureTraining & AlignmentInference & Deployment
01

Transformer & Attention

When d_k is large, the dot products QĀ·Kįµ€ grow in magnitude ā€” their variance scales with d_k, pushing softmax into regions with very small gradients (saturation). Dividing by āˆšd_k normalizes the variance back to ~1, keeping softmax in a stable gradient regime.

Scaled dot-product attention
Attention(Q,K,V)=softmaxā€‰ā£(QKāŠ¤dk)V\text{Attention}(Q,K,V) = \text{softmax}\!\left(\frac{QK^\top}{\sqrt{d_k}}\right)V

Key insight: Concrete example: d_k = 64 ā†’ without scaling, dot products have std ā‰ˆ 8; after scaling by 1/āˆš64 = 1/8, std ā‰ˆ 1.

02

Training & Optimization

03

Architecture Design

04

Training & Alignment

05

Inference & Deployment

20 questions across 5 categories. More coming soon.