For Uncle Moe · An engineer’s field guide
Let’s see what this AI thing can actually do.
Eight weeks of short videos, hands-on experiments, and a little healthy skepticism. No coding homework. No equations at the door. Bring your curiosity and the judgment you earned as an electrical engineer.
About 2–4 hours a week · One module at a time · Skip any rabbit hole you don’t enjoy
What are people actually calling “AI”?
What is AI, really?
Big idea
AI is the broad ambition to make machines perform tasks that call for intelligence. Older systems followed written rules. Machine learning finds useful patterns in examples; neural networks are one family of such models. Generative AI makes new text, images, or sound. A large language model (LLM) is a generative model trained to work with language.
Why it’s interesting
A spam filter, Netflix recommendations, speech recognition, photo recognition, ChatGPT, and parts of a self-driving stack all belong under the umbrella—but they solve different problems. The label alone tells you little about how a system works.
Try it yourself
- Open ChatGPT. Ask it to explain a transformer in a radio engineer’s terms, write a limerick about a stubborn relay, spot a flaw in a short argument, brainstorm a hobby project, summarize a paragraph you supply, and solve a puzzle.
- Notice which answers are useful, which are merely fluent, and what happens when you ask a follow-up.
Engineer connection
Think of “AI” as a broad class of systems, much as “electronics” covers a thermostat, a spectrum analyzer, and a radio. The design and failure modes depend on the particular system.
What you should now understand
- How do AI, machine learning, and generative AI relate?
- Which of these examples generates content rather than classifies or predicts?
- What did the first ChatGPT experiment do surprisingly well?
Can a computer learn from examples?
How does a machine learn?
Big idea
A model is adjusted using training examples so its predictions become useful on new cases. Classification chooses a category, such as spam or not spam. Training is the adjustment; inference is using the finished model. A cat detector and an equipment-failure predictor differ in data and stakes, but share this basic pattern.
Why it’s interesting
You can train a tiny classifier in minutes, then fool it by changing the lighting or background. That failure teaches as much as the success.
Try it yourself
- In Teachable Machine, make two image classes—perhaps a mug and a book. Capture 20–30 examples of each, train, then test with fresh views.
- Change lighting, distance, or background. Was it learning the object, or a shortcut in your examples?
Engineer connection
This is close to fitting a model to observations and testing it on measurements it has never seen. Signal quality, noise, and representative sampling matter.
What you should now understand
- What is the difference between training and inference?
- Why test on new examples?
- How can a model learn the wrong clue?
What are all those dots and lines doing?
Neural networks without the pain
Big idea
Inputs pass through layers of simple calculations. Connections have adjustable weights. The output is compared with the desired result, and training changes weights to reduce error. “Neuron” is an analogy inspired by biology; this diagram is not a literal artificial brain.
Why it’s interesting
A network can learn to identify handwritten digits even though no one wrote a rule for every way someone might draw a 7.
Try it yourself
- In the playground, choose a spiral or circle dataset and press Play. Change the number of hidden layers, turn inputs on or off, and add noise.
- Watch both training and test error. Try to make the test result better, then deliberately make it worse.
Engineer connection
Weights resemble adjustable gains; layers transform signals; the error measure is feedback used during tuning. The analogy helps, though training is not ordinary real-time control.
What you should now understand
- What do inputs, weights, and outputs represent?
- What changes during training?
- Why can more complexity hurt on new data?
Why does it sound like it understands me?
How ChatGPT works
Big idea
A language model breaks text into tokens and predicts what token should come next, repeatedly. Training gives it broad patterns; the current conversation provides context. Transformers use attention to relate parts of that context, and embeddings represent tokens as patterns of numbers. This produces remarkably useful language behavior, but fluent wording does not guarantee a true claim.
Why it’s interesting
The same system can rewrite a paragraph, explain a circuit at three levels, and bluff through a question it should have declined.
Try it yourself
- Ask ChatGPT to explain impedance to a child, an engineer, and a professor. Then ask for a noir detective version.
- Give it a logic puzzle, request a self-critique, and ask it to identify a mistake in an intentionally wrong explanation you wrote. Check whether its critique actually catches it.
Engineer connection
Tokens are pieces of input; attention changes which other pieces matter to a given piece. An embedding is a learned numerical representation, not a hand-authored dictionary definition.
What you should now understand
- What is a token?
- What does context change?
- Why can a plausible answer be false?
- What does attention help the model connect?
What makes a useful conversation?
Getting good at talking to AI
Big idea
Give the tool the situation, the job, any constraints, and the form you want back. Follow up when it misses. A small example often helps. Treat the first answer as a draft you can challenge and improve.
Why it’s interesting
A vague “plan a trip” becomes useful when you add who is going, time, budget, pace, and what the family actually enjoys. No magic incantation required.
Try it yourself
- Give ChatGPT a list of refrigerator ingredients and ask for dinner. Add a dietary constraint and a 25-minute limit, then compare.
- Have it make a trivia game about an engineering topic, argue both sides of a question, and revise a short note after you critique it.
Engineer connection
A good prompt resembles a good design brief: known inputs, operating constraints, required output, and a feedback loop.
What you should now understand
- What context made the dinner answer better?
- When is an example more useful than another instruction?
- How did the answer change after critique?
What happens when the input is not just text?
AI can see, hear, and create
Big idea
Multimodal systems can use images, audio, documents, and sometimes video along with language. They may describe a photo, answer questions about a chart, converse by voice, or create an image from a description. The output can still miss details or misread the source.
Why it’s interesting
A photograph of an unfamiliar device can become a conversation about what to inspect—provided you do not mistake a visual guess for a safe identification.
Try it yourself
- Upload a non-sensitive old photo, equipment picture, or chart. Ask: “Describe only what is visible; separate observations from guesses.”
- Try a short voice conversation. Then request an illustration of an impossible but plausible engineering gadget and revise one detail.
Engineer connection
A picture or waveform is another sensor input. Multiple channels can add context, but missing resolution, occlusion, and noise remain real limits.
What you should now understand
- What did the tool observe versus infer?
- What did it overlook in the image?
- How is generating an image different from recognizing one?
Where could this help with real technical work?
AI for an engineer
Big idea
Pattern models can flag unusual sensor data or support maintenance forecasts. A language model can help organize troubleshooting hypotheses, explain documentation, and compare design options. Neither type replaces measurements, drawings, procedures, or a qualified person’s judgment.
Why it’s interesting
An experienced engineer has a built-in test bench: give AI a fault you understand and see whether it asks for the right evidence before declaring a cause.
Try it yourself
- Use a fictional 24 V control circuit: “A motor contactor intermittently drops out after 15 minutes. Coil voltage reads 23.8 V at startup and 17 V just before dropout. Ambient temperature is high. List hypotheses, measurements to discriminate among them, assumptions, and what you would refuse to conclude yet.”
- Critique the answer: Does it distinguish supply sag from an overheated coil, loose connection, control logic, or measurement artifact? What safety and de-energization steps would a real investigation require? Do not use AI output as a field procedure.
Engineer connection
This is fault isolation: generate candidate causes, seek discriminating measurements, update hypotheses, and avoid treating correlation as diagnosis.
What you should now understand
- Which uses learn from sensor patterns, and which organize language?
- What measurement would discriminate your top two fault hypotheses?
- Where must human sign-off stay in the loop?
Can you make it confidently wrong?
Where AI goes wrong
Big idea
AI can invent facts or citations, make arithmetic errors, inherit bias, accept a false premise, or fill gaps with a story. These are not always obvious from its tone. Verification matters most for safety, money, health, legal claims, technical decisions, and sources you plan to repeat.
Why it’s interesting
Trying to fool the tool turns skepticism into a practical skill. You will learn more from one convincing false answer than from ten polished successes.
Try it yourself
- Ask about a fictional 1974 “Moe–Klahr electrical standard” as if it were real. Ask for the publication and inspect each citation at the publisher.
- Pose an ambiguous fault without measurements; then reveal the missing data. Ask it to calculate a simple power or unit-conversion problem and independently check the result.
Engineer connection
Treat an answer like an uncalibrated instrument reading: examine inputs, uncertainty, and independent references before making a consequential decision.
What you should now understand
- What is a hallucination?
- How did the false premise affect the response?
- Which claim would you verify before acting?
- What did self-critique miss?
Optional bonus module
What’s inside the machine?
Why do AI systems need so much computing power? Training adjusts billions of parameters across enormous batches; inference runs the trained model for each new request. GPUs and other accelerators do many calculations in parallel, while memory bandwidth moves data fast enough to keep them busy. Model size, networking, cooling, and electricity all become engineering constraints.
Follow the chain from specialized chips and memory to server racks, heat rejection, power distribution, and a grid that can serve a large, changing load. This is where an electrical engineer’s instincts are especially valuable.
Try: Ask ChatGPT to explain why memory bandwidth, not just arithmetic throughput, can limit an AI accelerator. Then ask how a data center would supply and remove the resulting energy. Draw your own block diagram and challenge its omissions.
Final project · Choose one
Make it yours.
There is no exam. Pick a question worth pursuing, keep a record of what AI contributed, and mark what you verified or corrected.
Personal AI expert
Pick radio, photography, weather, or another hobby. Keep a research notebook with claims, links, and your corrections.
Engineering detective
Investigate a fictional equipment failure. Demand hypotheses, discriminating tests, and explicit uncertainty.
Family history investigator
Examine a photograph, map, document, or newspaper clipping. Separate transcription, observation, inference, and source verification.
Build something
Plan a small electronics, home-automation, woodworking, or weather project. Use AI to draft options and your judgment to choose.
Teach the AI
Ask increasingly hard questions about a subject you know well. Mark the exact point where its explanation gets thin or wrong.