YOUR FIRST STEP INTO AI

A little learning. A world of possibility.

New to AI? Start with the basics, try a better prompt, and then build your first useful agent. Learn the programming fundamentals, compare languages and LLMs, or choose a quick introduction, guided trail or practical exercise.

Explore AI programming ↓

Understand AI before you automate.

OUR BEGINNER STARTING PICK

Google’s AI Essentials introduction starts with everyday uses. No coding experience needed.

YouTube course preview. The full certificate program is separate and may be paid.

Browse the AI jargon index A–Z →

FROM AI USER TO AI BUILDER

Learn the foundations. Then build something useful.

You do not need to learn every language or train a model to start. Pick a small task, learn one programming language and practise checking the result.

01 / USE AI TO CODE

An assistant helps you program

Ask a coding assistant to explain code, draft a function or suggest a test. You still need to understand the inputs, outputs and errors, run the code, and review the changes.

02 / BUILD WITH A MODEL

Your app calls an LLM

Your program sends an instruction to a model API or a locally hosted model, receives a response and validates it. Learn APIs, data handling, evaluation and deployment.

03 / DEVELOP MODELS

You train or adapt a model

This adds machine learning, datasets, statistics, linear algebra and compute planning. Fine-tuning an existing model is a different project from training a foundation model from scratch.

The fundamentals, in a practical order

01 / FOUNDATION

Programming logic

Variables, strings, lists, dictionaries, conditions, loops, functions and exceptions. Learn to read an error message and break a task into small functions.

Try it: Build a script that reads sample tasks and counts them by category.

02 / FOUNDATION

Data and SQL

Work with CSV and JSON, check missing values and understand tables, joins and filters. Good input data is part of the AI application.

Try it: Load a fictional dataset, remove duplicate rows and explain the result.

03 / FOUNDATION

APIs and the web

Understand HTTP requests, JSON responses, API keys, SDKs, timeouts and rate limits. An SDK is a helper library; the API is the service interface.

Try it: Call a sample API, validate its response and handle a failed request.

04 / FOUNDATION

AI and model concepts

Learn training versus inference, prompts, tokens, context, embeddings and retrieval. Compare a rule-based solution with a model-based one.

Try it: Write five sample inputs and the outputs you would accept.

05 / FOUNDATION

Testing and evaluation

Use Git, small tests and a fixed evaluation set. Check correctness, unsupported claims, speed and cost. Rerun checks when the prompt or model changes.

Try it: Keep ten representative test cases, including a case with missing information.

06 / FOUNDATION

Math and deployment as needed

API apps need basic logic and measurement first. Training and fine-tuning add probability, statistics, vectors, matrices and optimization. Learn hosting and monitoring before sharing an app.

Try it: Explain where your data travels and how you would limit requests and spending.

Official starting points: Python tutorial (assumes basic programming) ↗ · JavaScript fundamentals ↗ · PyTorch tutorials ↗

Which language is useful for what?

There is no single best language for every AI project. Python is a practical first choice for many new AI learners. If you already work in Java, JavaScript or C#, you can build AI features in that ecosystem.

What would you like to build?

Start with Python fundamentals, then CSV/JSON, APIs and a small model-powered project.

Suggested uses and starter projects — guidance, not a performance ranking
Language / skill	A strong fit for	Why learn it / example tools	First small project
Python Learn more ↗	AI experiments, automation, data preparation and machine learning.	Readable scripting with libraries such as pandas, scikit-learn and PyTorch.	Summarise a sample text file, then validate the result.
JavaScript / TypeScript Learn more ↗	AI websites, interactive interfaces and Node.js services.	Use JavaScript for the web; TypeScript adds static type checking. Call a model through a backend that keeps credentials private.	Build a page that sends a question to your own server.
Java Learn more ↗	AI features in Java services and enterprise applications.	Reuse Java skills and Spring Boot services; Spring AI integrates models, tools and retrieval.	Add a draft-summary endpoint to a Java service.
C# / .NET Learn more ↗	AI features in existing .NET applications and services.	Use the .NET AI ecosystem and model client libraries alongside your existing application.	Add a text-classification feature to a .NET service.
SQL Learn more ↗	Finding, joining and preparing business data.	A complementary data skill, used with Python, Java or another application language. SQL alone does not supply an LLM.	Select and aggregate a small fictional support dataset.
R Learn more ↗	Statistics, research, analysis and visualisation.	A language and environment built around statistical computing and graphics.	Explore a dataset and plot its distribution before modelling.
C++ / Rust PyTorch C++ ↗ Rust book ↗	Specialised inference runtimes, systems integration and performance work.	Useful when you need low-level control. Usually a later step for an API-app beginner.	Inspect how an inference runtime loads and executes a small model.
Go Learn more ↗	Backend services and infrastructure around model APIs.	Useful for building and operating services; it does not replace the model itself.	Build a small HTTP service that calls a model endpoint.

Java and JavaScript are different languages. You do not need both to start. A coding assistant can help in many languages; using an LLM API does not require a GPU on your laptop.

What is an LLM?

Large language model: a trained model for working with language

An LLM learns patterns from training data. During text generation, it predicts tokens—pieces of text—using the input and what it has generated so far. This supports tasks such as drafting, summarising, coding and answering questions. Some model versions also accept images, audio or other inputs.

A convincing answer can still be wrong. A model does not automatically have access to live websites, your company database or your files: an application must supply that information or connect tools with permission.

Read the technical introduction: Hugging Face text generation guide ↗

Terms you will meet while building an AI application
Term	Plain-language meaning	Why it matters
Model vs application	A model produces outputs; an app adds an interface, prompts, tools, storage and controls.	A chat-app subscription and an API account can have different access and billing.
Training vs inference	Training adjusts model parameters. Inference uses a trained model to produce an output.	Most first applications use inference instead of training a new model.
Tokens and context window	Tokens are pieces of input/output. Context is the information the model can work with in one request.	Large documents may need selection or chunking; more context does not guarantee correctness.
Parameters / weights	Learned numerical values in the model.	Parameter count affects storage and compute, but is not a complete quality score.
Prompt and instructions	The task, context, examples and output constraints you give the model.	Make the task measurable and treat untrusted document text as data.
Embeddings	Numerical representations used to compare or retrieve related content.	Often part of semantic search; an embedding model is not the same as a chat model.
RAG	Retrieval-augmented generation: retrieve relevant source content and include it in the model request.	Useful for document questions; it does not retrain the model or guarantee a correct citation.
Fine-tuning	Further training an existing model on selected examples.	Can adapt behaviour or format; it is not a substitute for retrieving frequently changing facts.
Tool calling and agents	A model proposes a tool call; application code decides whether and how to execute it.	Validate arguments and permissions. Require review for consequential actions.
Open weights vs hosted API	Open weights can be downloaded under a licence; a hosted API runs a model for you.	Self-hosting needs memory, operations and licence checks. Downloadable does not mean cost-free.
Quantization	Representing model numbers with fewer bits to reduce resource needs.	Memory use can drop, but quality, speed and runtime support can change.
Evaluation	Checking outputs against representative examples and acceptance criteria.	Measure your own task; a leaderboard is not a guarantee for your application.

Explore retrieval in the Cohere RAG guide ↗. Use each provider’s model card and API documentation for model-specific behaviour.

Meet the LLM families

A starting map of the ecosystem, not every model or a “best model” ranking. Check the exact model’s version, language support, input types, licence, availability and pricing in its official documentation.

Family guide checked 20 October 2026. This is a published guide, not a live availability feed. Routes shown are examples; other hosts may offer additional options.

Search models or providersExplore a route

Model families, providers and example access routes
Family / provider	Access route	What to explore	Before choosing
GPTOpenAIOfficial catalogue / model card ↗	Hosted models	General language, coding and reasoning applications.	Choose the exact model ID from the API catalogue; ChatGPT is an application, not an API subscription.
ClaudeAnthropicOfficial catalogue / model card ↗	Hosted models	Document work, coding and assistant applications.	Opus, Sonnet and Haiku are model lines; compare the current versions and their limits.
GeminiGoogleOfficial catalogue / model card ↗	Hosted models	Text and supported multimodal applications.	Check the chosen model’s input types, stable/preview status and regional availability.
GrokxAIOfficial catalogue / model card ↗	Hosted models	Conversational and tool-connected applications.	Tool access, media support and pricing depend on the selected model and API.
CommandCohereOfficial catalogue / model card ↗	Hosted models	Enterprise text workflows and retrieval-based assistants.	Command generates text; embedding and reranking models serve different parts of a retrieval pipeline.
LlamaMetaOfficial catalogue / model card ↗	Weights available	Experiments with hosting and adapting language models.	Read the licence and model card for the exact release; open weights do not mean unrestricted use.
GemmaGoogleOfficial catalogue / model card ↗	Weights available	Local or managed experiments with downloadable models.	Gemma and Gemini are separate model families. Choose a variant that fits your device and task.
QwenAlibaba / Qwen teamOfficial catalogue / model card ↗	API & weights	Language, coding and supported multimodal tasks.	Capabilities, sizes and licences vary across the Qwen releases; inspect the individual model card.
DeepSeekDeepSeekOfficial catalogue / model card ↗	API & weights	Language, reasoning and code-related applications.	Hosted API names and downloadable checkpoints may differ; compare the exact versions.
MistralMistral AIOfficial catalogue / model card ↗	API & selected weights	Hosted AI applications and selected self-hosted models.	Some models have downloadable weights and others are hosted products; check each model’s terms.
PhiMicrosoftOfficial catalogue / model card ↗	Weights available	Experiments with smaller language models and adaptation.	Size alone does not guarantee suitability; check memory, model format and task results.
KimiMoonshot AIOfficial catalogue / model card ↗	API & selected weights	Language and supported reasoning or agent workflows.	The publisher’s model cards describe available checkpoints and their deployment requirements.
gpt-ossOpenAIOfficial catalogue / model card ↗	Weights available	Self-hosted reasoning-model experiments.	A separate open-weight family from hosted GPT models. Check the model card and supported runtime.

How to pick your first model

Choose a small task and compare two suitable models on the same ten examples. Record accuracy, unsupported claims, response time, total token cost and failure handling. For self-hosting, also check RAM/VRAM, model format and runtime support. Start with an API if operating a model is not your learning goal.

Understand processors and memory ↗ · See dated benchmark leaders ↗

YOUR FIRST AI APPLICATION

Build a small document-question helper

Use a short fictional handbook and ten sample questions. The goal is to answer from that handbook and say when information is missing.

Prepare: read the text file and check that it contains the expected content.
Connect: pick one model and follow its official SDK quickstart. Keep API keys in server-side environment variables.
Ask: send the relevant passage and question, and request an answer with a supporting quotation.
Validate: confirm the quotation exists in the source. Treat a quotation as evidence to review, not proof that the whole answer is correct.
Evaluate: try known answers, missing facts and misleading input. Record correctness, latency and usage.
Improve: add retrieval for larger documents only after the small version works. Add tool actions later.

A tiny Python preparation exercise

This example prepares a prompt. It runs locally with standard Python and does not call a model or spend API credits.

handbook = "The fictional club meets on Tuesdays at 6 pm."
question = "When does the club meet?"

def make_prompt(source, question):
    if not source.strip() or not question.strip():
        raise ValueError("Provide a source and a question")
    return (
        "Answer using only the source. If the answer is missing, say so.\n"
        f"Source: {source}\nQuestion: {question}"
    )

print(make_prompt(handbook, question))

Next: use this prompt in a provider’s official quickstart, then check the returned answer against the handbook.

A sensible starter toolkit: one language, an editor, Git, a small dataset and one model interface. You can add notebooks, pandas, scikit-learn, PyTorch, retrieval tools or agent frameworks as your project needs them.

Practise prompts and agent planning ↓ · Find your saved guides ↓

YOUR LEARNING STUDIO

Choose a small step. Make real progress.

Start with a one-minute introduction, follow a starter trail, or practise a skill. These are free AI Atlas learning activities; linked provider courses may have their own fees.

Bookmarks, exercise drafts and practice progress stay in this browser, including on shared devices. They do not sync to your account. Video completion uses the existing guest or member progress setting.

Three guided starter trails

Pick a route and follow its three steps. Time estimates are for the on-page introduction and practice; videos are optional extra learning.

15–20 min · beginner

AI for everyday tasks

Learn the basics, draft a better prompt and check the result.

Read the quick intro & check your understanding
Optional: explore the provider video
Write a useful prompt

0 of 2 on-page activities complete

20–30 min · beginner

Better prompts, better drafts

Give clear instructions, compare two outputs and reflect on what improved.

Read the quick intro & check your understanding
Optional: explore the provider video
Compare your prompts

0 of 2 on-page activities complete

20–30 min · beginner

Your first agent plan

Understand tool use and design a small workflow with a review point.

Read the quick intro & check your understanding
Optional: explore the provider video
Plan an agent

0 of 2 on-page activities complete

Just the introduction

Three short reads with a check to help the idea stick.

1 MIN READ / BEGINNER

AI in a minute

AI finds patterns to produce predictions, recommendations or new content. A language model generates text from learned patterns; a fluent answer can still contain mistakes.

Start with a small, low-stakes task. Check important claims against an independent source.

1 MIN READ / BEGINNER

A prompt that gives direction

A useful prompt explains the task, context, constraints and desired output. Give a short example if the format matters. Improve the result by saying what needs to change.

Try: “Draft a 100-word welcome email for new club members. Use a friendly tone and end with three next steps.”

1 MIN READ / BEGINNER

From an answer to an action

An agent combines a model with instructions and tools to work toward a goal. A workflow may read information, decide what to do next and call a tool. Tool access also means mistakes can affect other systems.

For a first project, use sample data and review the plan before any action that changes or sends information.

Learn by doing

Use fictional or non-sensitive examples. Draft here, try the task in an AI tool if you wish, then assess your own work. Your writing is never sent to a model by this page.

PRACTICE / SELF-ASSESSMENT

Write a useful prompt

Create a prompt that asks an assistant to plan a short study session. Include your topic, time available and the output you want.

See an example approach

Plan a 30-minute study session for a beginner learning spreadsheets. Include three activities in a table with minutes and a way to check understanding.

One possible approach, not a graded answer.

PRACTICE / SELF-ASSESSMENT

Compare two prompts

Choose a simple task. Write a vague prompt and an improved prompt, then compare their results. Note one improvement and one claim you would verify.

See an example approach

Vague: Explain solar energy. Improved: Explain solar energy to a 12-year-old in 120 words, with one example and two limitations. Check the technical claims against an educational source.

One possible approach, not a graded answer.

PRACTICE / SELF-ASSESSMENT

Plan an agent

Sketch a draft-only assistant that gathers three sample articles and prepares a summary. Describe its goal, sources, tool permissions and human review point.

See an example approach

Goal: draft a weekly reading summary. Sources: three approved article links. Tool: read those pages. Output: three summaries with links. Human review: check the sources and approve the draft before sharing. No sending tool is connected.

One possible approach, not a graded answer.

Your learning bookmarks

Bookmark a trail, intro, exercise or video to find it here.

Manage browser practice data

Clear removes only learning bookmarks, intro checks, exercise drafts and practice completion in this browser. Video completion and member account data are managed separately.

WATCH. TRY. MAKE IT YOURS.

Your AI starter playlist.

Curated for different starting points. Videos open on YouTube.

Mark lessons as you complete them.

01 / Start hereGrow with Google · Google AI Essentials

Beginner · no code · English

Introduction to Artificial Intelligence (AI)

Course preview

A YouTube introduction to AI and everyday workplace uses. This is a course preview, not the complete Google AI Essentials certificate course; the separate course is offered through Coursera.

You’ll learn: Recognize useful AI tasks and know where human judgment is needed.

PUT IT INTO PRACTICE

Choose one everyday task, ask an AI assistant for a first draft, and check its output against the facts you know.

Watch on YouTube ↗ Mark complete

Course / tutorial source ↗ · Checked 21 Oct 2026

02 / Understand & buildMicrosoft Developer · Microsoft Cloud Advocates

Beginner → developer · English

Full Series [Part 1–18] | Generative AI for Beginners

Full video series

The combined 18-part Microsoft video series introduces generative AI application development. The linked Microsoft Learn course provides companion resources; technical exercises require development skills.

You’ll learn: Understand language models, prompting, and the building blocks of AI applications.

PUT IT INTO PRACTICE

Write a prompt with a task, useful context, and a requested format. Compare the answer with a vague version of the same prompt.

Watch on YouTube ↗ Mark complete

Course / tutorial source ↗ · Checked 20 Oct 2026

03 / Understand agentsMicrosoft Developer

Beginner · concepts first · English

What are AI agents?

Introductory lesson

The opening Microsoft lesson introduces AI agents, their components, use cases, and a code demonstration. Later course exercises involve programming.

You’ll learn: Explain the difference between answering a question and completing a task using tools.

PUT IT INTO PRACTICE

Sketch an agent with a goal, source, and tool. Decide where a person must review its output before any external action.

Watch on YouTube ↗ Mark complete

Course / tutorial source ↗ · Checked 20 Oct 2026

04 / Try your first agentTutorial linked by n8n

Beginner · visual builder · English

Build your first AI agent — template walkthrough

Hands-on tutorial

Follow the video linked from n8n’s first-agent template. Connect a model and try tools for weather or RSS news. Account or model usage charges may apply.

You’ll learn: See how a chat trigger, model, tools, and short-term memory fit together.

PUT IT INTO PRACTICE

Use the template with sample requests. Ask for five RSS headlines with source links, then inspect the workflow execution.

Watch on YouTube ↗ Mark complete

Course / tutorial source ↗ · Checked 18 Oct 2026

Ready to try an agent? ↗Explore practical uses and guided first steps.Find a tool for your task ↗Browse the AI websites directory.