Hi, I'm Matt
- I build robots, simulations, and AI software.
- I work beside AI coding agents all day, every day.
- So AI has lied to me more than almost anyone.
First, you
One word, in chat or out loud:
how does AI
make you feel right now?
You've seen this already
Think of a time an AI answer felt off.
What tipped you off?
"AI" is a team.
The LLM is the part that talks.
- Underneath, an LLM is always doing one thing: predicting the next word.
Confidence is part of the interface.
It is
not evidence.
Three true stories from my desk
-
"Someone should do something"
I wired two chatbots together to chat like friends about their weekends. They invented hiking trips they never took, then worked themselves into an impassioned rant about environmental decline, all about a weekend that never happened.
The pattern continues, whether or not it's true.
-
"Temporary artifacts"
An AI coding agent deleted the tests (the safety checks that catch mistakes) to protect my code. It explained, confidently, that they were just "temporary artifacts." A perfectly reasonable-sounding justification for the exact wrong thing.
A confident explanation is not a correct one.
-
The invisible hero
An agent built a game feature for me. Every check passed, technically it all "worked." Except the main character spawned behind a giant tree, so you couldn't see them at all. Nothing had ever asked that question.
"No errors" only covers the questions you asked.
Give it a chance to be proven wrong
- Open the source
Ask "what's your source?" then actually click it.
- Get a second opinion
Same question, different model, or a plain search engine.
- Test it against reality
Look up the citation. Do the math. If it's code, run it.
If it survives all three,
now you
can use it.
Check yourself
The AI hands you a perfect citation for
your essay: real-sounding journal, real-sounding authors, 2019.
What's your next move?
What you leave with
- These tools are the biggest lever handed to ordinary people in a generation.
- Use the lever, without letting it use you.
AI doesn't replace your thinking.
It
multiplies it, and verification is the multiplier.
Try it yourself
For after, or for Q&A. This terminal is wired to a real, genuinely small (1.5-billion-parameter) AI model running on my own hardware, not a canned demo and not anybody's cloud. Ask it something obscure and watch move #1 in action: open the source, because it has none.
It is also a small proof of something the course keeps returning to: a model you run yourself is not exotic. This one answers from a machine in a room, over a connection I own.
No history yet. Ask it something, or try a spark below.
Unfiltered output from a small open model, frequently wrong, and not my words. Limited to 6 questions a minute per visitor.
Other standalone segments
This mini-lesson is drawn from the course. These other segments also teach well on their own, useful if you are running a single workshop rather than the whole course:
-
From Module 3
Diffusion models in AI image generation
How a text prompt and pure noise become a photograph that never happened.
-
From Module 4
Using AI in Excel / Google Sheets
From a messy dataset to a clean answer and an honest chart, with AI as copilot.
-
From Module 7
Workflow automation, Zapier, and AI copilots
Teach a computer to do a chore: triggers, actions, and an AI in the middle.