Question 1

What are reasoning models and when should you use them?

Accepted Answer

**Target Audience:** Engineers who know how a regular LLM API call works and need to decide where a "thinking" model belongs in a product.
**Difficulty:** Medium
**Verified with:** Node 22.22.0 and Python 3.12.3 (the example is deterministic arithmetic and stub answers; no real model is called). Product facts are cited inline with the source and the date checked (October 2026).

> **How to read this doc:** Every idea is explained in plain language first, then tagged with `📌 Interview term:` so

Question 2

What is the difference between fine-tuning, LoRA and prompt engineering?

Accepted Answer

**Target Audience:** Engineers who can call an LLM API and have heard "just fine-tune it", but have never trained a model or seen what LoRA actually changes.
**Difficulty:** Medium
**Verified with:** Node 22.22.0 and Python 3.12.3 (a toy linear layer trained with plain gradient descent; no real model is trained or called). Paper and library facts are cited inline with the date they were checked (October 2026).

> **How to read this doc:** Every idea is explained in plain language first, then tag

Question 3

How do temperature, top-p and other sampling parameters affect output?

Accepted Answer

**Target Audience:** Developers who call LLM APIs and want to set sampling parameters on purpose rather than by copying defaults.
**Difficulty:** Medium
**Verified with:** Node 22.22.0 and Python 3.12.3 (the sampling pipeline in the example runs on a fixed logits vector; no real model is called). API parameter rules are cited inline with the source and the date checked (October 2026).

> **How to read this doc:** Every idea is explained in plain language first, then tagged with `📌 Interview ter

Question 4

What techniques reduce LLM latency at each stage of a request?

Accepted Answer

**Target Audience:** Engineers whose LLM feature works but feels slow, and who want a method for finding and removing latency rather than a list of tricks.
**Difficulty:** Hard
**Verified with:** Node 22.22.0 and Python 3.12.3 (a latency-budget model and a speculative-decoding simulation). The per-stage numbers in the budget are **stated assumptions** used for the arithmetic, not measurements of any provider; vendor claims (prompt caching latency, predicted outputs, speculative decoding speedups

Question 5

What is quantization and when would you use a quantized model?

Accepted Answer

**Target Audience:** Engineers about to self-host or run an open-weight model who keep seeing names like "Q4_K_M", "AWQ" or "4-bit" and want to know what they trade away.
**Difficulty:** Hard
**Verified with:** Node 22.22.0 and Python 3.12.3 (a hand-written absmax quantizer on 4,096 synthetic weights, plus memory and bandwidth arithmetic; no real model is loaded). Paper and library facts are cited inline with the date they were checked (October 2026).

> **How to read this doc:** Every idea is e

AI ENGINEERINGDECODED.

Medium

What are reasoning models and when should you use them?

What is the difference between fine-tuning, LoRA and prompt engineering?

How do temperature, top-p and other sampling parameters affect output?

Hard

What techniques reduce LLM latency at each stage of a request?

What is quantization and when would you use a quantized model?