Technical Interview Essentials A-Z

Author of Technical Interview Essentials A–Z: Coding, System Design, ML & LLMs. Ex-Google/Uber engineer and interviewer. Free interactive demos ↓

[Technical Interview Essentials A-Z @profxfang]

I wrote a 4-book series on technical interviews, one concept per letter A–Z:

🔹 Coding
🔹 System Design
🔹 Machine Learning
🔹 LLMs & Transformers

Many chapters have free interactive demos (consistent hashing, FlashAttention, N-Queens, speculative decoding…) 👇
https://t.co/Jv0X1FlJN4

[Technical Interview Essentials A-Z @profxfang]

Prepping for interviews? The LLM volume of my Technical Interview Essentials A–Z series is free on Kindle through Thu, Nov 12. 📘 Paperback also available.
https://t.co/gjRgLpeMOn

[Technical Interview Essentials A-Z @profxfang]

I just published in @gitconnected "Understanding Forward Checking with N-Queens" https://t.co/KXlUNd49mE

This is the second part in a three-part series on using the classic n-queens problem to explore variants of DFS.

[Technical Interview Essentials A-Z @profxfang]

@PythonPr Of course 11, but each of min and max costs O(n).

Far more elegant solution in one pass: 😜

from functools import reduce
print(
sum(
reduce(
lambda a, v: (min(a[0], v), max(a[1], v)),
x,
(x[0], x[0])))

[Technical Interview Essentials A-Z @profxfang]

Full walkthrough, from disk seeks to B-trees to the Bloom filter: https://t.co/bBrudj338H

Try it yourself: insert UUIDs and catch a false positive → https://t.co/HU0sqzSdDg

[Technical Interview Essentials A-Z @profxfang]

Interview question: 1 billion users, and most lookups are for keys that don't exist. How do you avoid a disk read for every "no"?

A Bloom filter. About 10 bits per key → 1% false positives.
1B keys = 1.12 GiB of RAM, k = 7 hashes.

Watch it work 👇

[Technical Interview Essentials A-Z @profxfang]

Start here: https://t.co/0zvpfHawlB

[Technical Interview Essentials A-Z @profxfang]

I wrote a 4-book series on technical interviews, one concept per letter A–Z:

🔹 Coding
🔹 System Design
🔹 Machine Learning
🔹 LLMs & Transformers

Many chapters have free interactive demos (consistent hashing, FlashAttention, N-Queens, speculative decoding…) 👇
https://t.co/Jv0X1FlJN4

[Technical Interview Essentials A-Z @profxfang]

Start here: https://t.co/0zvpfHawlB

[Technical Interview Essentials A-Z @profxfang]

Four interview books, one concept per letter, A–Z:

🏗️ System Design: Bloom filters, Consistent hashing, Quorum…
💻 Coding: BFS/DFS, DP, N-Queens…
🧠 ML: Adam, Backprop, XGBoost…
🤖 LLMs: Attention, KV cache, LoRA, MoE…

Plus free interactive demos. Map of all four 👇

[Technical Interview Essentials A-Z @profxfang]

Medium> This guy is pasting a bunch of code samples …
Medium> No matter what we auto-detect, he always sets the syntax highlighting to "Python."
Medium> Oh wait, he just pasted some more code. You think that's Python?
Also Medium> Fuck him, that's SCSS.

[Technical Interview Essentials A-Z @profxfang]

Excited to get my first article published in https://t.co/VjVo5s6hFF!

"Understanding Medusa and Tree Attention." https://t.co/VWT4emwR9c

This is part 4 of my series on accelerating LLMs:

Part 1: Intro to FlashAttention
Part 2: Speculative Decoding
Part 3: Multi-Token Prediction (MTP)
Part 4: Medusa and Tree Attention
Part 5: EAGLE & EAGLE-2 (pending)

Start here if you want a step-by-step walkthrough of LLM acceleration techniques.

[Technical Interview Essentials A-Z @profxfang]

Running out of Claude quota reminds me of when Netflix first started streaming but limited you to only 19 hours a month. I rationed those hours carefully.

Nowadays everyone is streaming 24/7 and the Internet doesn't blink. Eventually tokens will reach the same level of abundance.

[Technical Interview Essentials A-Z @profxfang]

After we repeat this sampling procedure 1,000 times, we see how it converges to generating a probability distribution that matches our original target model p(t).
https://t.co/bNykvm8L5J
https://t.co/WiCEZ8vhcD

[Technical Interview Essentials A-Z @profxfang]

Speculative decoding is a technique for accelerating decoding tokens from a target LLM. We find a smaller, similar model (the “draft” model) and extract several tokens in sequence, then we pass all these draft tokens through the target model in parallel to “verify” if they were generated by the target LLM.

But if the draft model generates token according to one distribution q(t) and the target model generates tokens according to a different distribution p(t), how can the tokens generated by the draft model behave as if they were generated by the target model? 🧵

[Technical Interview Essentials A-Z @profxfang]

But if the draft token t fails the r < p(t)/q(t) chance, then we reject it (and all following draft tokens). To replace it we resample a new token from the modified distribution max(0, p(t) - q(t)). This distribution has the effect of accounting for all the probability that we haven’t explored yet where the distribution p(t) is more likely than q(t):

[Technical Interview Essentials A-Z @profxfang]

In part 2 of "Accelerating LLMs" we look at speculative decoding: https://t.co/WiCEZ8uJn5

- Part 1: FlashAttention https://t.co/Xz6LS1KfZ7

[Technical Interview Essentials A-Z @profxfang]

My new favorite hobby is taking the figures from my books and having Claude Code make fully interactive web versions of them: https://t.co/EMIz1UKA3l

[Technical Interview Essentials A-Z @profxfang]

And solving "Perimeter of an Island" using a basic flood fill algorithm: https://t.co/EMIz1UKA3l

https://t.co/KiLJL36o4O