If you are wondering how to learn AI from scratch, start by choosing a path. Path 1 is the AI user: prompts and tools, about 30 days. Path 2 is the AI builder: assistants, automation and agents without code, 2–4 months. Path 3 is the AI engineer: Python, maths and machine learning, 6–12 months or more. All three can start for free.
Most people need paths 1 and 2; path 3 is a profession. New to the basics? Read what is artificial intelligence.
To be clear: at Tbilisi Digital School (TDS) we teach paths 1 and 2 to adults. We do not teach machine learning engineering, so for path 3 we list verified free resources.
How to learn AI: three paths in one table
| Path | 1. AI user | 2. AI builder | 3. AI engineer |
|---|---|---|---|
| Who it is for | Anyone who works at a computer | Owners, managers, marketers, freelancers | Future developers and data specialists |
| What you learn | Prompts, checking, documents | Assistants, automation, agents | Python, maths, ML, deep learning, LLM apps |
| Time | ~30 days, 20–30 min a day | 2–4 months, 4–6 h a week | 6–12+ months, 8–10 h a week |
| Tools | ChatGPT, Gemini, Claude, Gemini Notebook | Projects, Gems, Make, n8n | Python, Jupyter, scikit-learn, PyTorch |
| First project | A 5-prompt library for your job | A FAQ assistant and an automation | A prediction model on open data |
| Where to learn | This plan, free courses, TDS | This plan, a TDS course | Free online courses, not TDS |
Do you want to use AI, have it do work for you, or build it? The paths stack, and path 1 alone is already a complete skill.
How to learn AI from scratch: a 30-day plan (path 1)
Spend 20–30 minutes a day on your real work with one assistant, ChatGPT or Gemini (both list Georgian as a supported language). For week 1 you can follow our 7-day plan for beginners.
Week 1: account, privacy and first tasks
In ChatGPT, go to Settings, Data controls and turn off “Improve the model for everyone”. In Gemini, with Keep Activity off, chats are not used for training (unless you send feedback) and are kept for 72 hours. No ID numbers, customer data or passwords in a chat, ever.
Exercise: list 10 weekly tasks, give five to AI and log minutes before, minutes with AI and quality from 1 to 5.
Week 2: prompting and checking
A good prompt has a goal, context, format and limits. Add: “Ask me three clarifying questions first.” See our 10 rules for a good prompt.
Exercise: write three versions of one prompt and compare. Then ask five questions from your field whose answers you know and find at least one confident mistake (a hallucination).
Week 3: documents and research
Gemini Notebook (formerly NotebookLM) answers from the sources you upload, with citations, and makes Audio Overviews in about 80 languages, Georgian included.
Exercise: take a document you must learn, build a study guide and a 10-question quiz, and check every citation.
Week 4: your own routine
Collect 5–10 prompt templates and create a Project (ChatGPT or Claude) with standing instructions for your most frequent task. Run one prompt in a second assistant; our AI apps guide shows which tool suits which job.
Exercise: repeat the week 1 log on the same tasks and compare.
Theory alongside, no code: Elements of AI (University of Helsinki and MinnaLearn, about six weeks, free certificate), Claude Academy’s “AI Fluency” (4 hours) and Andrew Ng’s “AI Prompting for Everyone” (about 7 hours of video). More in our guide to free AI courses.
How to learn to build AI assistants, automations and agents (path 2)
On path 2, AI answers from your documents and does tasks with your approval. No code, but plenty of testing.
| Month | Topic | Exercise | Pass mark |
|---|---|---|---|
| 1 | Prompts as instructions | Instructions for one repeated task | 9 of 10 examples right |
| 2 | An assistant with your documents | A FAQ assistant with prices and policies | 18 of 20 questions right |
| 3 | No-code automation | Form, sheet, AI category, alert | One week without errors |
| 4 | Your first AI agent | A lead agent that drafts replies for approval | 10–20 test cases |
Month 1: system prompts, examples and strict output formats (a table, JSON), tested on 10 real cases.
Month 2: one knowledge base with prices, policies and FAQs, used by a Claude or ChatGPT Project or Gemini. This is RAG in plain words. As of October 2026, Gemini Gems on personal accounts become skills from November, and OpenAI retires custom GPTs on 11 December 2026 in favour of plugins, so do not build a new GPT. Steps: our no-code chatbot guide.
Month 3: Make (free plan: 1,000 credits a month) or n8n (self-hosted Community Edition is free): triggers, filters, errors and logs. See our AI automation guide.
Month 4: an agent is a model plus instructions, tools and memory. The n8n AI Agent node needs at least one tool; Make AI Agents does the same in Make. Rules: minimum access, human approval for money, customer messages and data changes, and a log. Walkthrough: how to build an AI agent.
After path 1, month 1 takes two weeks and the whole path 2–3 months.
Course at TDSAI Business Co-Pilot: From Assistants to AI Agents4 weeks · 20 h · Starts 16 NovDetailsLearning AI as an engineer: Python, maths, machine learning (path 3)
A career path: 6–12 months or more at 8–10 hours a week, coding daily. Resources are free unless noted; times are mostly the providers’ own estimates as of October 2026.
| Stage | What you learn | Free resource | Time |
|---|---|---|---|
| 1. Python | Variables, functions, lists, libraries | Kaggle Intro to Programming and Python; Harvard CS50P | 5 + 5 h; CS50P 10 weeks |
| 2. Maths | Matrices, derivatives, probability, statistics | 3Blue1Brown; Mathematics for Machine Learning | 16 linear algebra lessons; a reference book |
| 3. Machine learning | Regression, classification, overfitting | Kaggle Intro and Intermediate ML; Google ML Crash Course; Andrew Ng’s ML Specialization | 3 + 4 h; ~14 h; ~95 h |
| 4. Deep learning | Neural networks, PyTorch | fast.ai; Andrej Karpathy’s Zero to Hero | 9 × ~90 min video; 8 lectures, ~14 h |
| 5. LLM apps | APIs, RAG, agents | Hugging Face LLM Course and Agents Course | 12 chapters, a week each at 6–8 h |
Python: Kaggle Learn runs in the browser with free certificates; Harvard CS50P adds problem sets and a final project.
The maths you actually need: Google’s course asks for algebra, matrix multiplication and basic statistics; Andrew Ng’s needs high-school maths. Learn it next to code with 3Blue1Brown and the free book Mathematics for Machine Learning. Andrew Ng’s videos are free to watch; graded work and the certificate need paid Pro or Coursera.
Deep learning and LLM apps: fast.ai expects about a year of coding, ideally Python, and its 13.5 hours of video take several times longer with notebooks. The Hugging Face LLM Course needs good Python; its Agents Course takes 3–4 hours a week, with a free certificate.
By our rough count, that is 350–450 hours. For a job, 2–3 finished projects on GitHub count for more than certificates.
How to know you are progressing: a checklist
- Path 1: you get a hard task right in 2–3 rounds and spot AI’s mistakes yourself.
- Path 1: you use 5–10 templates weekly and know what never goes into a chat.
- Path 2: your assistant gets 18 of 20 real questions right and never invents a price.
- Path 2: an automation runs a week without you, and your agent has written permission rules.
- Path 3: you write a script from scratch that reads a CSV and plots a chart.
- Path 3: you evaluate a model on test data and notice overfitting.
Common mistakes when learning AI
- Tool-hopping. Stay with one assistant for a month.
- No project. Skill comes from your own task, not from videos.
- Pasting private data. Anonymise customer lists and contracts first.
- Unchecked answers. Verify facts, numbers and quotes.
- The wrong start. Maths first when you need AI at work, or neural networks before Python.
- Collecting certificates. Clients and employers care about what you can do.
Keep a learning log: one row a week with the task, prompt or workflow, result, minutes saved and what failed. After four weeks you will see your progress, and the log becomes a portfolio.
Learning AI in Georgian: which tools answer well
- Who answers in Georgian. ChatGPT and Gemini list Georgian as supported, and Gemini is available in Georgia. Georgian is not in Claude’s benchmark table, but Anthropic says Claude works beyond it, so test it. Compare them in ChatGPT, Claude or Gemini.
- Use Georgian script. Anthropic recommends native script over transliteration, so avoid “xelovnuri inteleqti” and say “answer in Georgian”.
- Where English helps. Most courses and documentation are in English (Elements of AI has no Georgian version). Ask your assistant to explain them in Georgian, and note terms in both languages.
- Georgian text inside AI images is still unreliable: add it separately.
Where to learn AI in Tbilisi: TDS courses and what we do not teach
At TDS we teach paths 1 and 2 to adults, in Georgian, on your real tasks:
- [AI Business Co-Pilot: From Assistants to AI Agents](/courses/ai-agents-small-business). 4 weeks, 8 sessions of 2.5 hours, no code: path 2 from prompts to an agent.
- [AI Employees: Grok Bot and ChatGPT Dots](/courses/ai-employees). 2 weeks, 4 sessions of 2.5 hours, for beginners.
- [AI for Teachers](/courses/ai-teachers). Online, 8 sessions of 2 hours.
- A corporate program of 30–50 hours: see AI courses and training for business.
All our AI courses are on one page. We do not teach Python, model training or machine learning engineering; for that, start with the free courses above.
Frequently asked questions
How do I learn AI from scratch?
Start with one assistant and the 30-day plan, 20–30 minutes a day on your real tasks. Then decide whether you need path 2 (automation and agents) or path 3 (Python and machine learning).
How long does it take to learn AI?
Everyday use takes about a month, building assistants and automations 2–4 months, and machine learning engineering 6–12 months or more.
How can I learn AI for free?
Elements of AI, Kaggle Learn, fast.ai, Google’s Machine Learning Crash Course and the Hugging Face courses are free. On DeepLearning.AI and Coursera the videos are free, but graded work and certificates are paid.
Do I need maths and programming to learn AI?
Not for paths 1 and 2: basic computer and spreadsheet skills are enough. Path 3 needs Python, algebra, matrices, basic statistics and later derivatives.
Where can I take a machine learning course in Georgia?
The most accessible route is free courses in English: Andrew Ng’s specialization, Kaggle’s Intro to Machine Learning and fast.ai. TDS does not teach machine learning engineering.
Which AI assistant works well in Georgian?
ChatGPT and Gemini officially support Georgian, and Claude writes it too, though Georgian is not in its benchmark table. Run the same prompt in two assistants and compare.
Next step
Do the week 1 exercise today. Unsure which path or course fits you? Book a free consultation and we will pick a plan together.
Sources
- How long will the courses take?, Elements of AI
- Kaggle Learn, Kaggle
- Practical Deep Learning for Coders, fast.ai
- Machine Learning Specialization, DeepLearning.AI
- Machine Learning Crash Course: prerequisites, Google for Developers
- LLM Course, Hugging Face
- Claude Academy, Anthropic
- Where you can use the Gemini web app, Google Help
