Skip to content

Free AI learning hub

Learn AI: from fundamentals to modern systems

Start with how models work, practise checking their answers, then learn retrieval, tools and evaluation. Understand the technology behind the headlines with original lessons and official sources.

What is AI? Artificial intelligence is the field of building systems that perform tasks such as recognizing patterns, generating language and making predictions. For a developer, learning AI means understanding these capabilities, connecting them to an application and testing where they fail.

Last updated:

Model catalog facts checked on October 1, 2026. This page is a dated snapshot; availability and product terms can change.

What should you learn, and in what order?

For a beginner, a useful order is foundations → prompting → modern technology → evaluation → agents. Progress by completing a task you can check, rather than memorizing model names.

  1. 1. Understand the foundations

    Learn how AI differs from machine learning, what tokens and context mean, and why fluent answers can be wrong.

  2. 2. Practise prompting and verification

    Give a clear goal, useful context, constraints and a checkable output. Verify claims against the supplied evidence.

  3. 3. Learn modern AI technology

    Explore current models, SI / ASI, reasoning, multimodal inputs, retrieval and tool connections in seven practical lessons.

  4. 4. Build and evaluate an agent

    Understand the plan, act and check loop. Add tools with limited permissions and test failure cases before relying on automation.

  5. 5. Add programming when you need it

    Use Python for working with data and building small applications. Keep a record of test cases and what you learned.

How do you learn to build AI applications?

First get comfortable with functions, lists, dictionaries, HTTP and JSON. Then follow the topics below. Each step links to existing detailed teaching and gives you one concrete way to check your understanding.

  1. Step 1

    Python and machine learning basics

    Machine learning fits patterns from data. Learn to separate training data, validation and testing before interpreting a model's result.

    Check your understanding: Read a small dataset, identify its target and explain what a held-out example tests.

    Python for AI →
  2. Step 2

    LLMs and AI APIs

    A large language model generates language from learned patterns. An API lets your application send inputs and receive outputs; context limits, errors and timeouts still need handling.

    Check your understanding: Sketch a request, response and failure path before choosing a provider.

    HTTP and HTTPS →
  3. Step 3

    Prompt engineering

    Prompt engineering means designing instructions, context and examples for a task, then checking how reliably the output meets your requirements.

    Check your understanding: Compare two prompts on the same self-written inputs, including one ambiguous request.

    Prompt examples →
  4. Step 4

    Embeddings and vector databases

    An embedding is a numeric representation produced by a model. A vector database stores and searches these representations; similarity is useful evidence, not a truth score.

    Check your understanding: Explain why a paraphrase might match semantically and why an unrelated passage might still rank highly.

    Vector databases →
  5. Step 5

    Retrieval-augmented generation (RAG)

    RAG retrieves relevant source material and supplies it to a model before generation. It connects an answer to available evidence without guaranteeing that retrieval or the answer is correct.

    Check your understanding: Trace one question through retrieval, supplied passages, generated answer and checked citations.

    RAG and grounding lesson →
  6. Step 6

    LLM evaluation

    Evaluation tests an AI system against defined success criteria. Check retrieval, factual support, refusal when evidence is missing, response time and failures separately.

    Check your understanding: Write expected results before running the model, then keep a few cases out of prompt tuning.

    AI interview questions →
  7. Step 7

    AI agents, agentic AI and MCP

    An agent uses a model and tools in a repeated action loop. Agentic AI describes systems that pursue tasks through such actions. MCP standardizes connections to external tools and data; it does not decide whether an action is safe.

    Check your understanding: Design a read-only tool with a narrow input, log its calls and stop the loop after a fixed number of steps.

    MCP and permissions →
  8. Step 8

    AI safety, privacy and responsible reuse

    Treat model outputs and retrieved instructions as untrusted inputs. Limit permissions, keep credentials outside browser code and confirm you can use the source material and model license.

    Check your understanding: Add missing-evidence, hostile-document and permission-denied cases to your test set.

    Agent risk and controls →

For a longer career plan, use the AI engineer roadmap. It identifies learning that needs external tools as well as what runs on this site.

AI application brief: answer questions from your own handbook

Build a small document question-answering prototype that shows the evidence for each answer. Use three short, self-written documents about a fictional club: opening hours, booking rules and equipment. This is a project brief, not a hosted chatbot or a finished implementation.

  1. Write the source material and expected answers. Give each document an ID. Create five questions: two directly answerable, one needing two passages, one ambiguous and one whose answer is absent. Record the expected evidence before prompting a model.
  2. Build retrieval first. Start with keyword matching and show the selected document IDs and passages. The mini search engine project teaches an index you can inspect. Add embeddings only when your tests show a retrieval problem they might solve.
  3. Generate from the supplied passages. Ask a model to answer the question with document IDs and to say when evidence is insufficient. Keep retrieved text separate from your instructions. Inspect each citation yourself; a plausible ID alone is not proof.
  4. Test retrieval and generation independently. Was the right passage available? Did the answer follow it? Replace a passage with an irrelevant one, add a conflicting opening time, and include text that tries to change the task. Record each failure rather than averaging it away.
  5. Handle operations explicitly. If you choose a hosted API, call it from a local script or a server that protects the key. Set a timeout and request limit and check the provider's billing terms. This site's network-isolated browser playgrounds are for local coding practice; never paste an API key into them.
  6. Keep a small evaluation log. Record the model version, prompt version, source IDs, expected answer, actual answer and result. Keep a few questions unseen during prompt editing. Extend your test set after every failure you investigate.

What counts as a useful result? You can trace an answer to its supporting passage, distinguish missing evidence from a retrieval failure, reproduce a test and explain a remaining limitation. Passing five practice questions is an educational milestone, not proof of production reliability.

Design your first evaluation →

Current AI models: a source-checked snapshot

These are selected catalog entries, not a complete list or a benchmark ranking. Check model IDs, stable or preview status, input types, access and current terms on the linked official pages.

Selected model catalogs checked October 1, 2026
Family and sourceModel namesStatusWhat to learn
OpenAI GPTGPT-6 Astra, GPT-6.1 Sol, GPT-6 LunaListed in the current API catalogCompare reasoning quality, latency and cost on the same task. A ChatGPT subscription and API access are separate products.
Anthropic ClaudeClaude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 4.5Listed in the current platform overviewLearn model IDs and versioning. Confirm access on your chosen platform before designing a workflow around a particular model.
Google GeminiGemini 3.8 Flash, 3.5 Flash-Lite; Gemini 3.1 Pro previewFlash models stable; 3.1 Pro in previewCheck stable versus preview status, supported input types and retirement notices. Text, live voice and image generation use different endpoints.

For downloadable models, inspect the individual model card and license. Open weights, open source and unrestricted commercial use are different claims.

What is SI / ASI? How is it different from AGI?

In future-AI discussions, SI often means superintelligence. ASI stands for artificial superintelligence: a proposed system exceeding human ability across a broad range of intellectual tasks. AGI refers to broader general capability, with definitions and measurements still debated. Neither label follows automatically from releasing a more capable chatbot.

Judge specific evidence: tasks measured, failure rates, test conditions and independent verification. Avoid treating a forecast or marketing label as an established capability.

Read the AI, AGI and ASI lesson →

Seven practical lessons for modern AI literacy

Each lesson includes an explanation, an original practice task and sources or related reading. No paid subscription is required to read the course.

AI learning questions, answered directly

What should I learn first in AI?
Start with AI fundamentals, tokens, context and checking outputs. Then practise prompting, retrieval and evaluation before giving an agent permission to act. Python is useful for building applications, but you can learn the core ideas without coding.
What is SI or ASI in artificial intelligence?
In discussions about future AI, SI often means superintelligence; ASI means artificial superintelligence. It describes a proposed system exceeding human ability across a broad range of intellectual tasks. It is a research concept, not a capability you can infer from a chatbot's confident answer.
Which AI model is best for learning?
There is no universal best model. Choose one you can access, test it on your learning tasks, and compare accuracy, explanations, supported inputs, privacy requirements and cost. Verify current names and availability in the vendor's official catalog.
Are open-weight AI models automatically free for commercial use?
No. Downloadable weights do not establish permission for every use. Read the specific model license, acceptable-use conditions and any dataset or software licenses before reuse or distribution.
Does RAG prevent hallucinations?
No. Retrieval can supply relevant evidence, but the system can retrieve the wrong passage or misrepresent a correct one. Evaluate retrieval and answers separately, show source passages and include cases where the right response is that evidence is missing.

Sources, originality and corrections

Edited by SkillAIVibe. Explanations and the handbook project brief were written for this learning hub. Vendor documentation supports catalog facts; it does not mean we independently benchmarked the products. Research concepts are distinguished from available features. Product names identify their respective owners and imply no endorsement.

The developer roadmap's MCP, embedding, evaluation and risk explanations were checked against the linked primary documentation on October 7, 2026. The model table has its own October 1 catalog date. See our editorial policy, or report an error. Use fictional or self-written data for practice, and check permissions before publishing or automating work.

Last updated: