Make sense of the language around AI. Explore the ideas behind learning, language, vision and model adaptation with clear explanations and everyday examples.
Machine learning is a way of training software to recognise patterns in data and use them to make predictions or generate content, instead of writing a separate rule for every case.
Imagine sorting customer messages into delivery questions, product questions and complaints. With a rule-based approach, you might write lists of keywords. With a supervised machine-learning approach, you provide labelled examples so a model can learn patterns associated with each category. You then check how it handles messages it did not see during training.
Google’s introduction distinguishes approaches including supervised learning, unsupervised learning and reinforcement learning. They use different kinds of examples or feedback. Generative AI also learns from data, but produces content rather than only assigning a category or predicting a number. Machine learning is part of AI; the two terms do not describe exactly the same scope.
Learning here means adjusting a mathematical model, not acquiring human understanding or automatically getting everything right. Poor examples, misleading labels or a change in the real world can produce poor results. For a practical system, ask what it was trained to do and how it performs on relevant new cases. A good score on familiar examples is only part of the story.
Deep learning is a type of machine learning that uses neural networks with many layers to learn patterns from data.
Those layers transform information through a sequence of mathematical operations. In a simplified image example, parts of a network can represent local visual details and combine them into information useful for identifying an object. The exact representations depend on the model and training; they are not a hand-written checklist of what every object must look like.
Deep learning underpins many modern systems for language, images and speech. “Deep” refers to the layered structure, rather than a claim that the software thinks deeply about its answer. It is a narrower term than machine learning, which also includes methods that do not use deep neural networks.
Training can require substantial data and computing resources, although the requirements vary enormously by task and model. A small classification problem does not automatically need the biggest network available. When comparing an AI product, focus on whether it performs your task reliably and what it costs to run. The label “deep learning” tells you something about the method, but does not guarantee accurate answers, human-like reasoning or a better result for every job.
A neural network is a machine-learning model made of connected computational units that transform inputs into outputs using parameters learned from data.
The units are often called neurons, but they perform mathematical operations rather than behaving like tiny people making decisions. Connections have weights, and training adjusts parameters such as weights and biases to improve performance on an objective. Layers let the network combine information through multiple stages.
As a simple example, imagine a model that uses a product description to predict which shop category it belongs in. During training, it sees examples and adjusts how it combines useful signals. When shown a new description, it produces an output based on those learned relationships. A more complex network can process richer inputs, but the basic idea is still computation with learned parameters.
The brain analogy is limited: calling it a neural network does not show that it has feelings, common sense or consciousness. It can also make confident mistakes on unfamiliar material. Deep learning uses networks with many layers, while “neural network” is the broader term. For an application, the useful question is whether the trained network works on the cases you actually expect it to encounter.
Natural language processing, or NLP, is the field of AI concerned with helping computers process, analyse and generate human language.
It covers much more than chatbots. Tasks include classifying a message, extracting a date from a document, translating text, identifying topics and producing a summary. Methods can combine language rules, statistics and machine learning. Modern language models are important tools within this field, rather than a replacement name for the whole subject.
Consider a support inbox. An NLP system might identify that a message concerns a delayed delivery, extract an order number and route it to the right team. Generating a reply is a separate step, and actually changing the order requires a connection to the business system with appropriate permissions. Understanding the words and carrying out an action are different responsibilities.
Language is full of ambiguity, humour and missing context, so processing it successfully is not the same as understanding a person perfectly. Check names, intent and important details, especially when a short sentence could mean several things. When a product advertises NLP, ask which language task it performs, which languages it supports and how its results are checked.
Computer vision is the field of AI that extracts and interprets information from images or video, such as identifying objects, reading visible text or tracking movement.
Different tasks need different outputs. Image classification assigns a category to an image; object detection locates items within it; segmentation identifies regions or individual pixels. A system that reads a label is doing a different job from one that follows a moving box through warehouse footage.
Imagine a small manufacturer checking whether a component is present in a product photo. A vision system could flag images for inspection. To evaluate it, test ordinary examples and difficult ones: reflections, shadows, partly hidden parts and a different camera angle. A clear demonstration on one photograph does not establish that it will work throughout the production day.
Computer vision analyses visual information; image generation creates or changes images. A product can combine both, but they are different capabilities. Vision results can be wrong, particularly when the input differs from the examples the system handles well. For a useful deployment, decide which errors matter and when a person should review the result, rather than assuming that a machine looking at a picture necessarily knows what happened.
Multimodal AI works with more than one type of information, such as text, images, audio or video, within a model or connected system.
For example, you might upload a photo of a damaged bicycle and ask a written question about the visible parts. The system needs to relate the image to the words. Another workflow might combine a spoken question with a camera view. The supported combinations depend on the product and model; “multimodal” does not mean every input and output format is available.
Check what the application actually passes to the model. A document may be processed as extracted text, rendered pages or a combination, while a video workflow may use sampled frames and other information. The label alone does not tell you how completely a long file has been inspected. Test the kind of material that matters to your task.
Combining inputs can provide useful context, but it does not remove mistakes. Small text, unclear audio and details outside the selected frames can still be missed. For the bicycle example, ask the assistant to distinguish visible observations from guesses and verify any important conclusion yourself. More ways to send information are useful; they are not an automatic guarantee of more reliable answers.
Retrieval-augmented generation, or RAG, is a method that retrieves relevant material from an external source and supplies it to a model as context for generating an answer.
Think of an assistant answering questions about a company handbook. Instead of relying only on patterns learned during training, the system searches the approved handbook, retrieves relevant passages and gives those passages to the model with the question. The model then uses that context to compose its response.
RAG does not, by itself, retrain the underlying model. Updating a source or its search index is different from changing model parameters through fine-tuning. The approaches can be combined, but they solve different parts of a problem: finding relevant information and adapting how a model behaves.
The quality of the answer still depends on the source, retrieval and generation. An outdated policy or a missed exception can lead to a wrong response. For the handbook assistant, test whether it finds the right section, respects access permissions and links back to the supporting passage. Also test questions the handbook cannot answer. Giving a model documents can help ground an answer; it does not make every sentence automatically correct.
Fine-tuning means continuing the training of an existing model on selected data to adapt its behaviour or performance for a particular task.
You start with a pretrained model instead of building one from random initial parameters. The additional training updates some or all trainable parameters, depending on the method. Hugging Face’s training guide describes this process using a smaller, task-specific dataset. The amount of data and computing needed depends on what you are trying to change.
For example, a business might want support messages classified into a consistent set of categories. It could prepare carefully labelled examples, fine-tune a suitable model and measure the results on separate examples reserved for evaluation. Training on the same material you later use to claim success would give a misleading picture of how it handles new cases.
Fine-tuning is different from adding instructions to a prompt or retrieving documents with RAG. It is not usually the first thing to try when an assistant needs a current delivery policy: that may be better handled by supplying the relevant source. Define the failure you want to fix, test simpler approaches and check whether the improvement justifies training, evaluation and maintenance costs.
Open-source AI refers to AI made available with meaningful freedoms to use, study, modify and share it; downloadable model weights alone do not establish that the whole system is open source.
The Open Source Initiative’s AI Definition sets out those freedoms and the components needed to exercise them, including code, parameters and sufficient information about training data. It is a specific published definition, rather than proof that every product using the label meets the same standard.
Open weights means the learned model parameters are available under some terms. Those terms may still restrict use, and other important components may be missing. Before adopting a model, inspect its actual licence and accompanying material. Check whether your intended commercial use, modifications and redistribution are permitted, rather than relying on the word “open” in a launch announcement.
Also separate access from running cost. A model you can download without paying may still require hardware, hosting, setup and maintenance. For a small project, ask what you can inspect, what you may change and what it will take to operate the system. Open-source availability can offer useful control, but it does not automatically mean free infrastructure, unrestricted use of every asset or reliable results without testing.
Artificial general intelligence, or AGI, refers to AI with broad capabilities across many kinds of tasks, rather than competence limited to a narrow application; there is no single universally agreed test for it.
The central question is generality: how widely can a system apply its capabilities, and how well does it perform? A system that is excellent at one game does not establish general intelligence. Equally, discussing many subjects is different from reliably completing difficult work across them. Definitions differ over the breadth and level of performance required.
Google DeepMind’s Levels of AGI research proposes evaluating performance and generality, while treating autonomy as a related but separate dimension. This helps explain why an impressive demonstration or a company’s announcement is not a complete assessment. The tasks, testing conditions and assistance provided all matter.
AGI is also distinct from artificial superintelligence, which concerns capabilities beyond humans, and from a claim that a system is conscious. Those ideas can appear in the same discussion without meaning the same thing. When you see an AGI headline, ask which definition is being used and what evidence supports it. A confident prediction of an arrival date should be treated as a forecast, not an established fact.