Skip to content
LearnAIZero → AI Engineer

YouTube companion · 31 videos + 8 for later

Videos, placed in the week they help

Every link was checked on 11 Nov 2026. Videos support the sessions – they don’t replace them. Pause and type along, and never watch a code video faster than 1.25×.

Week 1 · Set up your Mac, learn the Terminal and Git, and write your first Python.

Open week

Mon–Tue: Two 15-min halves. It says ‘master’; GitHub now uses ‘main’ and a token or SSH key instead of a password.

Wed: Your setup already uses uv; this shows what uv init, uv add and uv run do.

Fri–Sat: The same lecture your CS50P Week 0 page embeds. Pause and type every example.

Sat: CS50P Week 1, before Problem Set 1.

Week 4 · Classes and three small tested projects: the Phase 1 capstone.

Open week

Mon: Pause around the middle and write a small class of your own.

Tue: Alternative constructors (@classmethod), which Pydantic and many libraries use.

Wed: Modern type hints (list[str], str | None), needed for Pydantic and FastAPI later.

Thu: The lecture your CS50P Week 8 page embeds.

Week 5 · Start daily interview problems; HTTP, APIs, pandas and SQL.

Open week

Mon: Watch before Problem #1: the Python idioms interviews expect (deque, heapq, sets, dicts).

Week 6 · SQL in Python, Pydantic, then your first LLM calls.

Open week

Sat: Pairs with Saturday’s Pydantic work. Watch the first half, then code.

Week 7 · Catch up on Weeks 1–6, consolidate Python and SQL, and rest over New Year.

Open week

Any light slot: A short big-picture view before LLM engineering starts. New Year-week friendly.

Week 8 · Prompting, structured output, tool use, and Project 1.

Open week

Thu + flex: First hour on Thursday (pretraining, tokens, neural net internals); the rest in flex or a buffer week.

Week 10 · Evaluate your RAG system properly.

Open week

Thu: Practical companion to the AI Evals FAQ: metrics, unit tests, LLM-as-judge.

Flex: Stanford lecture on LLM evaluation and judge biases. Watch in 30-min chunks.

Week 11 · Polish P2 (hybrid search, local models) and start agents.

Open week

Mon–Wed: Why vector search alone misses, and how a cross-encoder reranker fixes it.

Thu + Sat: Code-along for BM25 + embeddings + fusion + reranking. Matches Friday’s hybrid search build.

Mon–Wed: Anthropic’s workflows-vs-agents talk; pairs with Saturday’s reading.

Flex: Agent patterns in plain Python (tools, memory, chaining, routing) before any framework.

Week 12 · MCP and agent evaluation: finish Project 3.

Open week

Mon–Wed: The MCP team explaining what it is and how to start.

Sat: First hour while you build your MCP server; the rest is optional.

Week 15 · PyTorch on Apple Silicon, transformers and attention.

Open week

Flex: Stanford lecture on the transformer, as a second explanation.

Week 16 · LoRA fine-tuning on your Mac (Project 4).

Open week

Mon–Wed: Stanford lecture on fine-tuning, LoRA and quantisation, in 30-min chunks.

Thu: QLoRA explained with code; the theory behind Saturday’s LoRA run.

Flex: Needs an NVIDIA GPU: do the fine-tune on a free Colab T4, then run the model on your Mac with Ollama.

Week 18 · FastAPI, JavaScript and TypeScript.

Open week

Mon–Tue: A January 2026 FastAPI series that uses uv. Part 1, then continue with the series if it helps.

Wed: Async basics for FastAPI and parallel LLM calls.

Flex: Adding a database with SQLAlchemy (FastAPI series, Part 5).

Week 19 · React front-end for P5 and your first agentic coding task.

Open week

Mon–Wed: Anthropic engineers on Skills (SKILL.md) for agents.

Flex: Anthropic’s Claude Agent SDK workshop: gather context → act → verify.

Week 20 · Scoping, Docker, deployment and the FDE capstone.

Open week

Thu + Sat: Watch the project-structure and Docker parts at 1.25× while you package P5; skip what you’ve already built.

Flex: Optional: how Anthropic designs agents that run for hours.

After Week 21 · Karpathy’s makemore series and Stanford CME295

How to use the videos

  • Every link was opened on 11 Nov 2026 (title, channel, upload date). Lengths appear only where YouTube showed them.
  • Videos support the sessions; they don’t replace them. Your ‘done when’ checks and commits still decide when a week is finished.
  • Code along or skip it. Pause and type; never watch a code video at more than 1.25×.
  • Fast-moving topics need recent videos. For APIs, agents, MCP and tooling, check the upload date and the top comments for ‘this no longer works’.
  • 3Blue1Brown (Weeks 14–15), Karpathy’s micrograd (Week 14) and Let’s build GPT (Week 15) are already in your sessions, so they aren’t repeated here.

What wasn’t taken from the Gemini review, and why

  • Claude Corps: it’s real, but it’s for people authorised to work in the US and places fellows at US non-profits, so it doesn’t fit you (S85).
  • Unverified claims: a dated MCP spec version, the ‘Agentic AI Foundation’ donation, details of specific London team postings, UK salary bands and quotes about lab hiring couldn’t be confirmed from primary sources.
  • Videos that couldn’t be opened: the CS50P intro, FastAPI Part 3, a Stanford Fall 2026 ‘LLM Systems’ lecture, Dave Ebbelaar’s context-engineering, agentic-RAG and PydanticAI videos, Shaw Talebi’s RLHF/DPO video, two NeetCode overviews and a freeCodeCamp NeetCode 150 compilation. They may be fine – they just couldn’t be confirmed.
  • Fixes applied: the micrograd link pointed to a different Karpathy video (makemore Part 4); the correct link is used. Some lengths were wrong (CME295 Lecture 2 is 1 h 47, not 1 h 20).
  • The suggested restructure (deep learning in Weeks 7–11 ahead of RAG and agents, with an ‘agent oversight’ capstone) wasn’t adopted: applied LLM work first is what most UK AI-engineer adverts ask for, and the capstone stays built for a real user.

Tutorials to avoid

Out of date, or the wrong kind of practice.

  • OpenAI Assistants API: shut down on 26 Aug 2026. Tutorials using client.beta.assistants or threads no longer work – use the Responses API.
  • LangGraph before v1: create_react_agent is deprecated (replaced by LangChain’s create_agent); older tutorials often break on import.
  • No-code agent builders: fine for automation, but they don’t build the Python skills these roles test.
  • ‘Learn X in 10 minutes’ summaries and sales-funnel ‘full courses’: they feel productive but build no coding muscle.
  • Re-uploads of CS50 by other channels: use the official CS50 lectures embedded on the course pages.