Stage 2: Training vs. Inference
Training prepares a model using data. Labels are used in supervised learning, but not all learning requires labeled examples. Inference uses a prepared model to produce an output. Testing measures performance on separate examples; it is not another name for inference.
This real, local toy classifier stores fictional labeled fruit examples. It predicts the label of the nearest example—there are no neural-network weights to update.
Training Dataset
The labeled examples the model learns from.
Model memory is empty.
Click "Load examples" to prepare the nearest-neighbor classifier. This stores examples rather than optimizing neural-network weights.
Model Not Trained
You cannot run inference (predictions) until the model has processed the training data.
Inference: make a prediction
Adjust sliders to classify a new, unseen item.
Model Prediction
How it works: distance = √((weight difference)² + (redness difference)²). The scales chosen for these two features affect the result. Distance is not a confidence score. A fruit's true identity cannot reliably be determined from just weight and redness.
Stage 3: The Generative Shift
Compare the goal of each task: prediction or classification outputs a forecast or label, while generation produces content such as text, images, or audio. This is about the output, not a strict boundary between technologies: language models can classify text too, and generative output is not guaranteed to be original or correct.
"Detecting credit card fraud"
Stage 4: Limits & Narrow Capabilities
Not All Chatbots Are AI
Historically, chatbots were rule-based conversational interfaces. They used explicit decision trees (e.g., "If user says 'refund', show 'refund policy'"). They did not understand language; they just matched keywords. Today's generative LLM-backed bots formulate responses dynamically, but both types are still broadly called "chatbots."
Context Window ≠ Core Learning
A typical chatbot uses messages included in its context without updating its core model weights during that conversation. Some products also store memories or retrieve earlier information across sessions. That saved information is separate from model training; availability and privacy depend on the product's settings and policies.
Hallucinations and "Retrieval"
Generative AI can produce fluent but false statements. Even when augmented with search tools (Retrieval-Augmented Generation), retrieval does not guarantee accuracy: a model can misinterpret a document or cite an irrelevant source. Verify important claims against reliable sources and keep human oversight for high-stakes decisions. Fluent output alone does not establish human-like understanding or consciousness.
Frequently Asked Questions
Is all AI just Machine Learning?
No. AI also includes symbolic reasoning, search, planning, and rule-based expert systems. Machine learning is widely used, and a single product can combine several approaches.
Does the AI 'know' what it's saying?
A language model can learn complex representations and perform useful language tasks, but a confident response is not evidence of factual knowledge or human-like experience. Evaluate the output and its evidence rather than treating fluency as proof.
Why can't we just fix the bias?
Bias can enter through data collection, labels, objectives, evaluation, and deployment. Better data, testing across relevant groups, and human review can reduce some harms, but no single adjustment guarantees fairness in every context.
Is Deep Learning the same as Generative AI?
No. Deep learning describes a family of neural-network methods; generative AI describes systems that produce content. Many modern generative systems use deep learning, but deep learning is also used for classification, and generative modeling includes other methods.