How ChatGPT Actually Works — No Math Required

Enroll

The belief that AI is complex math you cannot do keeps capable people out of rooms where they belong. The internals that matter for building are not mathematical: tokens, context windows, attention as a way of weighting relevance, sampling, and post-training. Once you hold those, hallucinations stop being mysterious, model launches stop feeling like whiplash, and you can answer the question every product review now asks, which is whether the model will do this reliably and at what cost.

The outcome

How does this prepare you for a job?

Product and platform interviews test judgment about capability: what the model can do, where it fails, and what it costs. This course is that judgment, practiced.

Explained LLM behavior to non-technical stakeholders, covering context limits, the causes of hallucination and the effects of sampling, in order to guide product decisions.
Set token, latency and cost budgets for LLM features, then validated them against real usage data.
Chose between frontier and small models for each use case using analysis of capability and cost, with evals rather than impressions.
Designed prompts and structured outputs informed by how models actually process context.
Anticipated failure modes in product specs, covering stale knowledge, injection and degradation over long context.

Tools

AI tools learnt

Mostly the products themselves, opened up and measured.

Hands-on skills

Read a model card and work out what actually changed
Explain tokens and context windows well enough to size a feature's budget
Name the reasons models hallucinate, and the product patterns that invite it
Use temperature and sampling deliberately rather than leaving the defaults in place
Estimate the cost of a request from tokens and price sheets, and find the change that saves 10x
Separate a genuine capability shift from marketing when a model launches
Frame reasoning models accurately: what changes, what does not, and what it costs

Syllabus

Modules

Level: Basic

Tokens and context

What a token is, and why context windows bound everything else
Run the tokenizer over your own documents
Cost arithmetic, taking you from tokens to dollars

Why it answers the way it does

Next-token prediction, and attention as a way of weighting relevance, with no equations
Sampling and temperature, understood by running the experiment
Where hallucination actually comes from

From base model to ChatGPT

Pretraining, then instruction tuning, then RLHF, told as one story
System prompts, and why the same model behaves differently in different products
What claims that the model got safer or dumber usually mean

Level: Advanced

Tool calling and structured outputs

How function calling works underneath
Schema adherence, and the ways it fails
Design patterns that survive a malformed output

Reasoning models and long context

What extended thinking changes, and what it costs
Degradation over long context, meaning the real limits behind the advertised number
Retrieval set against context stuffing, with the decision made by evals

Capability judgment as a job skill

The questions a product review asks: can it do this, reliably, and at what cost
Build a small eval that answers one of them for real
Frame the success metric for an AI feature the way a FAANG product review would

Students can take the full course (both levels) or either level on its own.

Register

Waitlist

Join the course waitlist

No cohorts are scheduled yet. Join the waitlist and we’ll email you when registration opens.