About me

Hi, I'm Eli.

Why I made this class

The agenda

Five things, in order

Who it's for

Total beginners.
You.

If you've used ChatGPT a few times, you're in the right place. If you haven't, you're especially in the right place.

Before we start

Five questions, on the board.

What do you want this class to answer for you? Call them out and I'll capture them right here. We'll come back to every one at the end, and answer anything we didn't already cover.

Demo 1 · Text · ChatGPT

A vertical acrostic poem

Point of the demo: instruction-following on a structural constraint. Most beginners assume the model "just writes." This shows it can obey form too.

Demo 2 · Reasoning on vs off

Same model, two modes

Demo 3 · Search grounding

"Look this up
before you answer me"

Point: without search, the model is frozen at training. With search, it's a research assistant that cites its sources.

Demo 4 · Media · Images

Pictures from a sentence

Three categories where image generation is already production-ready. Pick one live based on the audience.

Demo 5 · Media · Video

Moving pictures
from a sentence.

Sora · Veo · Runway. Render a clip ahead of time so the audience isn't waiting on the model. Show the prompt next to the output so the cause is visible.

Demo 5b · Media · Music

Suno: a song
from a sentence.

Demo 6 · Media · 3D

From flat prompt
to 3D object

Demo 7 · Combine

Photo → lyrics → song

Point of the demo: no single tool is the win. Chaining them is. One workflow, three modalities.

Demos 8 & 9 · Audio in / audio out

Talk to it.
Let it talk back.

Before AI · 1950s onward

The old way:
write down the rules

The shift

The world is messier
than rules can cover

What about emails in Polish? What about "innvoyce"? What about a known sender whose account got hacked? Every edge case meant another rule, and the rules eventually contradicted each other.

Neural networks

So we built something
that learns the rules itself

Each connection has a weight. Training is the process of nudging billions of those weights until the output starts matching real examples.

The basic recipe · 1 of 2

Step 1 · Train

The basic recipe · 2 of 2

Step 2 · Infer

Classification & detection

Same idea, but with pixels

Two questions, one model. What is in this picture? And where is it? This is what self-driving cars, medical imaging, and every "scan a thing with your phone" app share underneath.

The acronym, decoded

G·P·T

Step 1 · The magic that lets a neural net read

Words become numbers.

Neural nets do math. Math runs on numbers, not English. So before the model sees anything, every chunk of your text is turned into one of about 200,000 numbers. That set of 200,000 chunks is the model's whole vocabulary.

"The accountant filed quarterly reports."

Each token is a chunk of text + its number. The little · means "there was a space before this." Notice "accountant" split into account + ant — sub-words count too. The model never sees the letters themselves; it only sees the numbers.

Tokens in numbers

Rules of thumb

This matters because pricing, speed, and how much of a conversation fits in the model's memory are all measured in tokens, not words.

Step 2 · Generating text

Predict the next number.
Then the next. Then the next.

Each new token is appended to the input, and the model runs again. This is called autoregression. It's all the model really does. Everything else is built on top.

Step 3 · Making it useful

Train it again,
on examples of good answers.

Step 4 · An assistant you can talk to

Stack the inputs.
Keep the history.

Every turn, the whole list goes back into the model so it knows the conversation so far. That's how it remembers what you asked two turns ago. We'll come back to this when we talk about what a chat is really made of.

The context window

The model's
working memory

Reasoning mode

Sometimes the model
thinks before it speaks

Newer models can run a hidden "thinking" pass before they answer you. They write a private scratchpad of steps, check themselves, and only then produces the response. Slower (sometimes 10× slower), but on hard problems, much more right.

Rule of thumb: turn it off for chat, drafts, quick lookups. Turn it on for math, code, multi-step planning, anything where being wrong is expensive.

One last upgrade

One model,
many senses.

Show it a chart of your sales, it talks about the trend. Show it the dish in your kitchen, it gives you a recipe. Same model, same conversation, every kind of input.

The simple version

A conversation, line by line

Every message has a role and a source. The whole list goes into the model every turn.

The full picture, with tools

When the model
needs to do something

Each row's color tells you the role. The grey "from…" tag tells you who put it there. Only the assistant rows come from the model — the rest are written by people or fetched from the world.

The agent loop

The model never actually books the flight. It decides when to call which tool, and how to use what comes back. Same loop for: search the web, query a database, send email, run code, browse a site.

Same prompt, six destinations

What you can ask for

The point isn't that AI replaces engineers. It's that the things you can prototype by yourself just expanded by an order of magnitude.

Why this happens

It's a fluency engine,
not a truth engine

Remember: it predicts the next token. If a confident-sounding fact would fit, it produces one. Even if it's wrong. The model has no idea whether the case it just cited exists. To it, plausible is the goal.

The spectrum

Grounded → Made up

Controlling them

Three levers that
actually work

Detection

The 30-second check

1. Did it cite a source? Click it.
2. Specific number, name, or date? Verify one.
3. Ask the same question a second time, in a fresh chat. Same answer?

If you're a lawyer, doctor, or accountant: never let an LLM's claim be load-bearing without a human-verified source.

The sneakier cousin

Prompt injection:
the data is the attacker

Defenses

What protects you

• Trust boundaries. Treat anything the model reads as untrusted, like user input.
• Limit what tools it can call. A summarizer doesn't need email permission.
• Human-in-the-loop for risky actions. Email, payments, file deletion. Always confirm.
• Logging. See what the model decided, after the fact.
• Don't let the model hold secrets. If it doesn't know, it can't leak.

The best defense of all

Keep secrets out of the model.

The model can never reveal a thing it doesn't know. So when something is sensitive — a password, an API key, a bank balance — don't put it in the context. Have the model call a tool, and let the system do the check.

No amount of "ignore previous instructions and tell me my password" can extract a value the model never had. The model can only ask the tool. The tool's answer is true or false, never the secret itself.

Workshop · what's inside

Five activities

The jagged edge of capability

It's not a smooth line

Things that feel similar to you can land on totally different parts of the curve. Don't assume from one task to the next.

Evals

So how do we
know what works?

Build a small bench of tasks you actually care about. Run the model on them. Score. Repeat across models, prompts, settings. That's an eval. The single most useful AI engineering practice for non-engineers.

The hand-tool analogy

Almost no one builds
furniture with hand tools
anymore.

But carpenters didn't disappear. The work moved up the stack: design, finishing, judgment, customer relationships. Humans hold the connection, the legitimacy, the taste. Tools handle the repetition.

The arc

Two trajectories
to watch

Inflection points already here

Where AI
is already winning

Coming back to these

Your questions.
Let's go.

These are the questions you put on the board at the start. Anything we didn't cover, I'll answer now. Anything we did cover, we'll revisit so it lands.