AI, Actually Explained · Chapter 01

What this word actually means

इस शब्द का वास्तव में क्या मतलब है

8:04 · Narrated with a synthetic voice. Captions are available in the player.

Read the transcript

An answer appears. It sounds confident. Its references look convincing. But before we trust it, we need to look behind it. Welcome to Addaly. AI, Actually Explained. Chapter one: What this word actually means. Think of this chapter as a map. We will open it, follow its paths, and discover which questions are worth carrying. Artificial intelligence is the name of a broad field. It is not one machine. A chess program, a spam filter and a chatbot can work in different ways. The phrase appeared in a nineteen fifty-five proposal for a research project the following summer. It named an ambition, not a finished technology. So ask three questions. What goes in, and what comes out? Where did the behaviour come from? And when it is wrong, who notices, and who bears the cost? The label cannot answer those questions. Here is the map. AI includes search, written rules and learned models, often combined. Machine learning fits behaviour from data or experience. People still choose the procedure, examples and objective. Deep learning uses neural networks with many layers. Despite their name, artificial networks are not biological brains. Generative describes a task: producing content, such as words, images or sound. Many familiar generators use deep learning, but generation is not confined to that method. A chat box tells you how to interact. It does not tell you everything working behind it. Imagine two email filters. This is a teaching example. One follows a rule someone wrote: block messages containing free money. The other is fitted from examples labelled spam or not spam. It learns a decision function rather than receiving a separate instruction for every message. Their difference is how the behaviour was obtained. A small learned decision tree can have readable rules too. A large neural network is harder to interpret. If either blocks an important message, investigate the cause. The fix might involve a rule, an input, a threshold, or retraining. Then test it. How does training happen? In one common neural-network setup, begin with numerical parameters. Picture adjustable dials. The model predicts an answer for an example. The chosen error measure, called a loss, compares prediction with target. An optimiser makes small changes to trainable parameters to try to reduce that loss. Predict. Compare. Adjust. Repeat. That is one method. A decision tree is fitted differently, by selecting branches from data. Improving on training examples is not enough. Use separate validation data to guide choices, then evaluate on held-out tests. Check relevant groups and conditions. More examples or more updates do not guarantee a better result. For a neural language model, the trained artifact includes numerical weights used with model software. These parameters influence how inputs become outputs. They are different from tokens: the chunks of text entering and leaving. The context window limits how much the model can consider at once. Picture its working desk, not its permanent memory. The application adds other parts: instructions, stored notes, document search and tools. A weight file is not a dependable source archive. Yet neural models can sometimes reproduce training passages. Neither novelty nor repetition proves that an answer is verified. Prediction connects many of these systems. A system that detects spam predicts a category. Many text generators predict the next token from context. But predicting a chosen target is different from achieving our real goal. Imagine a picture classifier trained where every leaf photo has a blue corner mark, and every flower photo has a red one. It might learn the marks instead of the plants. Change the marks, and the shortcut can fail. Ask what it was rewarded for predicting. Where did those targets come from? What happens when the pattern changes? Now a chatbot makes a mistake. You correct it. Its next answer improves. Has it permanently learned your correction? Not necessarily. With frozen weights, ordinary answering does not update those weights. The correction can help because it is present in this conversation’s context. A product might also save a memory or retrieve a document later. Those are different mechanisms. Some systems deliberately update through further training, including incremental learning. So, which part changed: the current input, stored information, the surrounding software, or model parameters? A fresh conversation alone cannot reveal every process behind a product. Does any of this amount to understanding? Researchers disagree about what that requires. Smooth, convincing language alone does not settle the question. A failure alone does not settle it either. In a controlled Othello study, researchers trained a model on game-move sequences. They recovered information about the board from its internal activity. Changing that representation changed its move predictions. That is evidence about one model and task, not proof of human-like understanding. For practical decisions, ask what the system does reliably, which variations expose failures, and how costly those failures are. Words can hide these distinctions. An algorithm is a procedure. A product may combine many procedures and fitted models. Researchers use the abbreviation A G I for artificial general intelligence. It has different proposed definitions. Ask which capabilities are meant, under which conditions. How were they measured? The AI is equally vague. Ask which model and version. And trained on your data needs unpacking. Using a message to answer now, storing it, letting people review it, and using it for future training are separate questions. Check the particular product’s current policy and settings. Sending text as context does not itself retrain frozen weights. Try an imaginary school decision. One dashboard uses written attendance rules. Another predicts recorded school departures from past data. The second reports eighty-two percent accuracy. Suppose one fifth of pupils leave. Predicting that nobody leaves would already score eighty percent, while flagging none of the pupils who leave. The headline is not enough. Ask about missed cases, false alarms, local testing and who reviews a flag. Neither method’s name tells you which support would help a child. Return to our convincing answer. A reference is a route to evidence, not evidence by itself. Follow it to an independent source. Read whether the source supports this particular claim. If you cannot verify it, pause. A failed search does not prove fabrication. Check another reliable route before relying on the answer. Your map now has landmarks: rules and learning; weights and context; predictions and evidence. Choose one AI feature you use. Explain what goes in, what comes out, where its behaviour came from, and how you would check an error. If something is unknown, name the question you would ask. This is Addaly. Keep your curiosity. Carry the questions into the chapter activities.

Sources and further reading (4)