LLM Interview Prep

Hot take: Interview prepping actually gets you ready for the job.

• They don’t test for your leetcode

https://github.com/Devinterview-io/probability-interview-questions

ML Interview

Need to start doing ML interviews. Some good exercise questions:

Links

• https://mimansajaiswal.github.io/posts/llm-ml-job-interviews-fall-2024-process/
• https://huyenchip.com/ml-interviews-book/contents/chapter-5.-math.html
• https://github.com/Devinterview-io/llms-interview-questions
• https://github.com/Praveen76/LLMs-Interview-Prep-Guide

Use this site to practice various things:

• https://www.deep-ml.com/problems/1
  • This was referenced from this site https://mimansajaiswal.github.io/posts/llm-ml-job-interviews-resources/

The little book of deep learning

• https://fleuret.org/public/lbdl.pdf

https://goyalpramod.github.io/blogs/Transformers_laid_out/

• Again referenced from the blog

Probability practice:

• https://openquant.co/questions
• https://devinterview.io/questions/machine-learning-and-data-science/probability-interview-questions/

Let’s build something more targeted than “just CF”. Think of 3 pillars:

1. Parsing + communication
2. Jane-Street-style math / state / trading thinking
3. Clean implementation speed

https://www.reddit.com/r/csMajors/comments/1f55c47/jane_street_ml_intern_interview/ https://www.interviewquery.com/interview-guides/jane-street-machine-learning

Things I failed:

• Implementing a tree, the scheduling for a dag

1. Implement a decision tree from scratch (you will need to figure out how to split a tree effectively)
2. Implement backgrad from scratch

https://github.com/andrewekhalel/MLQuestions https://github.com/Devinterview-io/naive-bayes-interview-questions

Likelihood updates, vs. posterior

(c) Talk through confusion instead of freezing

Instead of going quiet when it doesn’t click:

“I’m stuck on X specifically. I see how to simulate, but I don’t yet see a clean formula or invariant. My guess is there’s a symmetry or a conservation law I’m missing, maybe something like ‘cash + position·price’ staying constant. Let me check that on the small example.”

This is huge. That’s exactly how a strong teammate sounds when stuck.

TODO:

• Read the deep learning book, and add more notes on how MLPs work
• Do all problems on https://www.deep-ml.com/problems/1

You are forced to learn the foundational knowledge required for the interview.

1) ML fundamentals (the “must-cover” list)

A) Core math + statistical thinking

• Probability: conditional probability, Bayes rule, expectation/variance, covariance/correlation, common distributions, CLT, concentration intuition
• Estimation: MLE/MAP, log-likelihood, bias/variance, regularization as priors
• Information: entropy, cross-entropy, KL, calibration intuition
• Optimization: gradients, convexity intuition, SGD, momentum/Adam, learning-rate schedules, numerical stability

B) Supervised learning fundamentals

• Linear regression, logistic regression (including interpreting log-odds)
• Losses: MSE/MAE, log loss, hinge; why/when each makes sense
• Regularization: L1/L2, weight decay, early stopping
• Bias–variance tradeoff, overfitting diagnosis, learning curves
• Feature scaling, collinearity, interactions, basic feature engineering

C) “Classic” models you should be fluent with

• Trees, random forests, gradient boosting (XGBoost/LightGBM-style ideas: depth, learning rate, subsampling)
• kNN (good for intuition), Naive Bayes (likelihood view)
• SVM (margin intuition), PCA (variance + linear algebra intuition)
• Clustering: k-means, GMM basics

D) Deep learning essentials (engineering + intuition)

• Backprop + autograd mental model
• MLP basics; attention/transformer basics at the level of “what it computes / why it works”
• Normalization (batch/layer norm), dropout, initialization
• Training dynamics: batch size, gradient clipping, mixed precision
• Monitoring: loss curves, metrics, overfitting, debugging “it doesn’t learn”

E) Evaluation + experiment design (very interview-relevant)

• Metrics: ROC-AUC vs PR-AUC, log loss, RMSE; classification thresholds; ranking metrics (NDCG) if it comes up
• Calibration and when accuracy is misleading
• Handling imbalance
• Data leakage patterns; train/val/test discipline
• Cross-validation vs time-series split / walk-forward validation (esp. for finance-y data) Jane Street
• Uncertainty + confidence intervals (bootstrapping is often enough)