← All of Kevin's Corner
Use It, Don't Build It
Nobody cares about chatbots anymore
The AI backlash is real, and a lot of it is fair. Here's my contribution to it, from the inside.
For the last three years, as a CIO, I've deployed enterprise AI chatbots. We followed the textbook RAG model (retrieval-augmented generation): point a large language model at the company's own documents, let people ask questions, and get answers grounded in what we actually wrote down. It was safe and easy, and it was somewhat useful.
That's the whole review: somewhat useful. People tried it, got a decent answer to a policy question, and went back to working the way they always had. Three years in, nobody cares about chatbots anymore, and I've stopped pretending they do.
Answering questions was never the job
A chatbot answers questions. But almost nobody's job is answering questions. The job is getting something done: a report filed, a schedule built, a customer helped, a product shipped.
A chatbot stops right where the work starts. It tells you what the policy says, then leaves you to do everything the policy requires. That's why the novelty wore off so fast.
What people want now is AI that does part of the work while a person stays in charge. You can see it in Camp-9, The 4-Hour Build. AI helpers wrote most of the code for DogsPlayingPoker.ai, five AI players ran a tournament overnight, and a human made every call that mattered. The site took about four hours at the keyboard. Nobody asked it a single policy question.
Nobody but a computer scientist wants to build an LLM
The other half of my soapbox: almost nobody wants to build a large language model. Computer scientists do, and the world needs them. A teacher, a student, a shop owner or a CIO needs to use one well.
We already live this way with everything else. Most people drive without building engines and cook without breeding wheat. They still learn enough about a car to tell when it's unsafe, and enough about food to read a label.
HoldemRobots AI Training Camp is built on that assumption. The course assumes you will use an AI, not build one. Campers play AI from five companies, coach a bot, catch one cheating, and price the result. None of them trains a model from scratch, and that's on purpose.
Using one well is the hard part
"Just use it" is not the easy option. Using AI well takes five skills that a chatbot never asked of anyone, and the camps teach each one:
- Choose. Which AI for which job, and whether the job needs AI at all.
- Test. Run it against something you can check before you trust it.
- Price it. Know what an answer cost, in dollars and in electricity.
- Catch it. Spot when it cheats, bluffs or makes things up.
- Know where its thinking comes from. A model behaves the way its training data taught it to.
The fifth skill is the guardrail on my own argument. You don't have to build a model to be fooled by one. If you never learn that a model's habits come from its examples, you can't tell when lopsided examples have made it lopsided. That's why the camp still plans to teach how AI learns from data. It won't use code or a GPU. It'll use a deck of cards and a paper bot that only knows the hands it was shown.
What this means for the standards
The course meets the standards about using and judging AI, and skips one standard about building it, on purpose. Our Nov 4 analysis rated CSTA's new Machine Learning subconcept as the camp's one clear miss. Seen through this lens, that miss splits in two:
| CSTA 2026 standard | What it asks | The course's stance |
|---|---|---|
| HS-ALG-ML-06 | Justify choosing a type of AI algorithm | Keep. Already partly met: rules versus AI. |
| HS-ALG-ML-07 | Evaluate training data: source, quality, bias, privacy | Keep. It's skill 5, and the Paper Bot lesson is how we teach it. |
| HS-ALG-ML-08 | Develop a machine learning model | Skip on purpose. It belongs in courses for future computer scientists, like CS50 AI. |
This gives schools a plain statement to work with. HoldemRobots is a course in using AI with judgment. For building models, pair it with a computer science course. That's the same "use the big courses for the concepts, use HoldemRobots as the lab" stance the analysis already takes.
Back to the backlash
Much of the backlash is aimed at two promises: that a chatbot would change everyone's work, and that you need to understand how models are built before you're allowed an opinion. The first didn't happen, and the second was never true.
The answer isn't more chatbots or more model-building classes. It's ordinary people who can choose an AI, test it, price it, catch it, and know where its thinking comes from. That's cheap to teach, and it's what HoldemRobots teaches.