Week 1 Optional Weekend Assignment

This weekend, choose Quick Review or Extended Review if you want additional practice. You may also choose no homework without penalty.

Optional and ungraded: Complete one assignment route only if it is useful to you. If you complete a route, use the form below to create a PDF and submit it in eCampus for formative feedback. Choosing no homework does not affect your grade, Quiz 1 eligibility, or participation.

Choose Your Weekend Route

Quick Review · 8-Minute Video + One Question

3Blue1Brown: Large Language Models explained briefly

https://www.youtube.com/watch?v=LPZh9BOjkQs

Optional assignment task: watch the video, then complete the Quick Review Response below by writing one question you still have.

Extended Review · 83 Minutes of Video + Writing

Andrej Karpathy: Deep Dive into LLMs like ChatGPT

https://www.youtube.com/watch?v=7xTGNNLPyMI

Optional assignment task: watch 00:01–00:31, 00:59–01:42, and 03:21–03:31, then answer the six Extended Review prompts below. The 83-minute label covers video only; allow additional time for writing.

No Homework · No Penalty

You may stop after the Week 1 class activities. Bring your ungraded Week 1 evidence card back only if it helps your recap.

Choosing this route does not affect your course grade or participation in Week 2.

Deep Dive · Optional Enrichment

Watch the complete 3-hour-31-minute Karpathy lecture and use all 18 timestamped checkpoints.

This route is beyond the optional weekend assignment. The time label covers video only; allow additional time for checkpoints. You may split the work across sittings or change playback speed.

Assignment boundary: Quick Review means the 8-minute video plus one question. Extended Review means the three selected Karpathy segments plus six responses. The baseline, glossary, Deep Dive checkpoints, and self-checks are optional supports rather than additional assignment requirements.
Optional reference: Karpathy's MicroGPT article shows the same GPT ingredients in a tiny setting: documents, tokenizer, model, training loop, and inference loop. Use it only if you want a code-level preview after the class concepts.
https://karpathy.github.io/2026/02/12/microgpt/

Create Your eCampus Submission

Enter your response here, create a concise PDF, and upload that PDF to the Week 1 assignment in eCampus.

Assignment route

Your draft saves automatically on this device.

How to Save and Submit Your PDF

  1. Select Create PDF for eCampus.
  2. In the print window, choose Save as PDF as the destination.
  3. Use the suggested file name, which includes your name, and save the PDF to your device.
  4. Open the Week 1 assignment in eCampus, upload the saved PDF, and select Submit.
  5. Check that eCampus shows a successful submission confirmation.

Privacy: Your answers stay in this browser and are not sent to the course website. Clear your saved work when finished, especially on a shared device.

Optional Baseline

In one sentence, what do you think ChatGPT is doing when it answers a question?

Tiny Glossary

TermMeaning for this course
TokenA chunk of text the model processes. It can be a word, part of a word, punctuation, or code fragment.
Context windowThe current working memory: the text the model can see while answering.
PretrainingLearning broad patterns from massive text and code data.
SFTSupervised fine-tuning: training from example conversations or desired outputs.
RLHFReinforcement learning from human feedback: using human preferences to shape model behavior.
InferenceRunning the trained model to generate output, usually one token at a time.
HallucinationA fluent answer that is unsupported, misleading, or false.
ToolAn external action the system lets the model request, such as search, calculation, or code execution.
GuardrailA rule, approval step, or check that limits risky behavior.

Deep-Dive Route: 18 Optional Checkpoints

Timestamps are based on the current YouTube chapters for "Deep Dive into LLMs like ChatGPT." Use them as anchors, not as exact quiz boundaries.

00:01:00Pretraining data (internet)

What kinds of data help train an LLM? Why does data quality matter?

00:07:47Tokenization

What is a token? Give one example where tokenization might surprise a user.

00:14:27Neural network input/output

What goes into the model, and what comes out?

00:20:11Neural network internals

You do not need the math. What is the useful mental model for what the network has learned?

00:26:01Inference

Why does the answer appear one piece at a time?

00:31:09GPT-2: training and inference

What does a base model do? How is that different from a helpful chat assistant?

00:42:52Llama 3.1 base model inference

What did you notice about interacting with a base model?

00:59:23Pretraining to post-training

Why is pretraining not enough if we want a useful assistant?

01:01:06Post-training data (conversations)

How do conversation examples change the behavior of a model?

01:20:32Hallucinations, tool use, knowledge, working memory

Why can a model sound confident and still be wrong? How can tools help?

01:41:46Knowledge of self

Why might a model be confused about what it is, what version it is, or what it can access?

01:46:56Models need tokens to think

Why can writing intermediate reasoning or scratch work help a model perform better?

02:01:11Tokenization revisited

Why can spelling, counting letters, or manipulating strings be surprisingly hard?

02:04:53Jagged intelligence

What does it mean for a model to be brilliant at one task and weak at another?

02:07:28SFT to reinforcement learning

What changes when the system moves beyond examples and starts using feedback or rewards?

02:14:42Reinforcement learning

Explain reinforcement learning in one sentence using your own words.

02:48:26RLHF

Why do human preferences matter when building assistant-like models?

03:21:46Grand summary

Write three takeaways you want to remember before learning agents.

Bridge to Agents

Use this course definition:

Practical agent = LLM + goal + context + tools + loop + guardrails

Classify each example as chatbot answer, tool call, or agent loop.

ExampleYour classificationWhy?
ChatGPT explains the difference between pretraining and post-training.
A model asks a calculator tool to compute 438 * 27.
A coding system reads files, proposes a plan, edits code, runs tests, and asks before deploying.

Verification Habit

Pick one important claim from the video or from a chatbot answer. How would you verify it?

Exit Ticket

In 4-6 sentences, explain why an agent needs guardrails.

Optional Self-Check

Save Your Full Worksheet

This creates a PDF of every answer area on this page, including Optional Baseline, all 18 Deep-Dive checkpoints, Bridge to Agents, Verification Habit, Exit Ticket, and Self-Check. Unanswered areas are printed as (not answered) so nothing you typed is left out. Keep this for your own records; the eCampus submission still uses the Quick or Extended PDF above.

Every answer area on this page saves automatically on this device.

Feedback boundary: If you complete a route, submit the generated PDF in eCampus for formative feedback. No route is scored, recorded as extra credit, required for Quiz 1, or required for full participation in Week 2.