Clusterflick

Classify by similarity, not by neural network.

Most AI literacy training stops at GenAI. But "machine learning" covers a whole family of algorithms - K-Nearest Neighbors, decision trees, clustering, regression, support vector machines - most of which look nothing like a transformer. Clusterflick puts you inside one of the simplest and most transparent: K-NN. Place the labelled training data. Flick a new sample onto the board. Where it lands - and which neighbours it lands closest to - decides what it is.

• 🌐 Beyond the Black Box

  When people say "AI" today they almost always mean GenAI - large neural networks like ChatGPT or image generators. But machine learning is a whole family of algorithms, and most of them aren't black boxes at all. K-Nearest Neighbors, decision trees, clustering, linear regression - these are all "AI" too, and many of them produce predictions you can actually inspect and explain. Clusterflick lets you experience one of those algorithms directly. Once you see how K-NN works, the GenAI hype starts to feel like one tool in a much larger toolbox.

• 📍 Distance, Not Probability

  K-NN doesn't "learn" anything in the neural-network sense. There's no training run, no gradient descent, no weights. To classify a new input, K-NN just measures the distance to every labelled example it knows about, looks at the k closest ones, and takes a vote. That's it. In Clusterflick, that "distance" is literal - physical distance on the board between your flicked token and the sample squares around it. Understanding this distance-based logic is half of intuitively understanding most classical ML.

• 🧪 Data IS the Model

  In a neural net, the training data gets compressed into millions of weights and effectively disappears - you can't point at any one number and say "this is what the model learned about cats." K-NN is the opposite extreme: the training data is the model. Every labelled sample stays visible, and predictions are made by direct comparison against them. That makes the consequences of bad data immediately obvious - mislabel one sample and you can watch its bad influence spread to everything that lands near it. It's a powerful, tactile lesson about why data quality matters more than algorithmic cleverness in most real ML projects.

• 🧭 You are AI-Fluent

  Real AI fluency isn't being able to use ChatGPT. It's knowing which kind of model fits which kind of problem - and pushing back when someone reaches for GenAI on a problem that a 30-line classical ML algorithm would solve faster, cheaper, and more reliably. After Clusterflick, "AI" stops meaning "the magic chatbot" and starts meaning a family of techniques with very different tradeoffs. That's the difference between being talked at by AI vendors and asking the right questions.