Chapterبەش 1 · Lectureوانە 1.1

What Is Generative AI?ژیریی دەستکردی بەرهەمهێنەر چییە؟

What makes AI “generative”: models that learn the patterns of many examples and create new text, code, pictures, sound and video; how chat assistants are made in three stages (pre-training, fine-tuning, human feedback); foundation and multimodal models; training versus inference; diffusion models for pictures; a short history from ELIZA to AI agents; what a model knows and does not know; and why Kurdish answers need extra care.

چی وا دەکات «بەرهەمهێنەر» بێت: کە شێوازی زۆر نموونە فێر دەبن و دەق، کۆد، وێنە، دەنگ و ڤیدیۆی نوێ دروست دەکەن؛ یاریدەدەری گفتوگۆ چۆن لە سێ قۆناغدا دروست دەکرێت (، ، هەڵسەنگاندنی خەڵک)؛ و ؛ ڕاهێنان بەرامبەر ؛ بۆ وێنە؛ مێژوویەکی کورت لە ELIZAـەوە تا ؛ چی دەزانێت و چی نازانێت؛ هەروەها بۆچی وەڵامی کوردی وردبینیی زیاتری دەوێت.

14 min read16 خولەک خوێندنەوە · 14 stepsهەنگاو

What you will learnچی فێر دەبیت

  • Explain the difference between predictive and generative AI with an example.
  • Name the kinds of content generative AI makes and say what a multimodal model is.
  • Describe the three stages that turn a language model into a chat assistant.
  • Explain what a foundation model is, why few companies train one, and how training differs from inference.
  • Describe how a diffusion model makes a picture from noise.
  • Say what a model knows and does not know, and check Kurdish output carefully.
  • جیاوازیی نێوان پێشبینیکەر و بە نموونەیەک ڕوون بکەیتەوە.
  • جۆرەکانی ئەو ناوەڕۆکەی دروستی دەکات بناسیت و بڵێیت چییە.
  • ئەو سێ قۆناغە باس بکەیت کە دەکەن بە یاریدەدەرێکی گفتوگۆ.
  • ڕوون بکەیتەوە چییە، بۆچی کۆمپانیای کەم ڕایدەهێنن و ڕاهێنان چ جیاوازییەکی لەگەڵ هەیە.
  • باس بکەیت چۆن لە خاڵخاڵییەوە وێنە دروست دەکات.
  • بڵێیت چی دەزانێت و چی نازانێت و کوردی بە وردی بپشکنیت.

From recognising to creatingلە ناسینەوەوە بۆ دروستکردن

Your phone's gallery can find every photo of your cat. It looks at each picture and decides: cat or not a cat. That is AI that recognises. Now ask an AI app to draw a cat asleep on a Kurdish carpet. A few seconds later you have a picture that never existed before. That is AI that creates.

گەلەریی مۆبایلەکەت دەتوانێت هەموو وێنەیەکی پشیلەکەت بدۆزێتەوە. سەیری هەر وێنەیەک دەکات و بڕیار دەدات: پشیلەیە یان نا. ئەمە کە دەناسێتەوە. ئێستا داوا لە بکە پشیلەیەک بکێشێت کە لەسەر قاڵییەکی کوردی خەوتبێت. چەند چرکەیەک دواتر وێنەیەکت هەیە کە پێشتر هەرگیز بوونی نەبووە. ئەمە کە دروست دەکات.

This course is about the second kind: generative AI. You meet it when a chatbot answers you, when your phone suggests a reply, or when a tool turns your words into a picture. This first lesson explains what it is, how it is made, and what it can and cannot do.

ئەم کۆرسە لەبارەی جۆری دووەمە: . کاتێک وەڵامت دەداتەوە، کاتێک مۆبایلەکەت وەڵامێکت بۆ پێشنیار دەکات، یان کاتێک ئامرازێک وشەکانت دەکات بە وێنە، ئەم جۆرە دەبینیت. ئەم یەکەم وانەیە ڕوون دەکاتەوە کە چییە، چۆن دروست دەکرێت و چی دەتوانێت و چی ناتوانێت بکات.

Two panels. Left, predictive AI recognises: a phone gallery of six small photos in which the three cat photos are ticked; output: a label, “cat” or “not a cat”. Right, generative AI creates: the request “Draw a cat asleep on a Kurdish carpet” becomes a new drawing of a sleeping cat on a carpet; output: a new picture that did not exist before.دوو بەش. چەپ، ژیریی دەستکردی پێشبینیکەر دەناسێتەوە: گەلەرییەکی مۆبایل بە شەش وێنەی بچووک کە سێ وێنەی پشیلەکان نیشانەیان لێدراوە؛ دەرخستە: نیشانەیەک، «پشیلە» یان «پشیلە نییە». ڕاست، ژیریی دەستکردی بەرهەمهێنەر دروست دەکات: داوای «پشیلەیەک بکێشە کە لەسەر قاڵییەکی کوردی خەوتبێت» دەبێتە وێنەیەکی نوێی پشیلەیەکی خەوتوو لەسەر قاڵییەک؛ دەرخستە: وێنەیەکی نوێ کە پێشتر بوونی نەبووە.
Recognising picks an answer about things that exist; creating makes something new.ناسینەوە وەڵامێک لەبارەی شتە هەبووەکان هەڵدەبژێرێت؛ دروستکردن شتێکی نوێ دروست دەکات.CêIT

What is generative AI?ژیریی دەستکردی بەرهەمهێنەر چییە؟

Generative AI is AI that creates new content: text, computer code, pictures, sound or video. A model learns the patterns in a huge number of examples, then uses those patterns to make something new of the same kind. A US definition used by NIST (2024) says these models copy the structure and characteristics of the data they learned from in order to produce new, synthetic content.

جۆرێکە لە کە ناوەڕۆکی نوێ دروست دەکات: دەق، ، وێنە، دەنگ یان ڤیدیۆ. شێوازەکانی ناو ژمارەیەکی زۆر زۆری نموونە فێر دەبێت، پاشان بەو شێوازانە شتێکی نوێ لە هەمان جۆر دروست دەکات. پێناسەیەکی ئەمریکی کە ⁦NIST (2024)⁩ بەکاری دەهێنێت دەڵێت ئەم پێکهاتە و تایبەتمەندییەکانی ئەو لێی فێر بوون لاسایی دەکەنەوە، بۆ ئەوەی ناوەڕۆکی نوێی دەستکرد بەرهەم بهێنن.

Definitionپێناسە

AI that creates new content, such as text, images or sound, after learning the patterns of a very large amount of existing examples. Chatbots built on are generative AI.

جۆرێک لە کە ناوەڕۆکی نوێ دروست دەکات، وەک دەق، وێنە یان دەنگ، دوای فێربوونی شێوازەکانی بڕێکی زۆر زۆری نموونەی هەبوو. ئەو لەسەر دروست کراون .

An apprentice carpet weaverقاڵیچنێکی شاگرد

A young apprentice in a carpet workshop in Sanandaj studies thousands of Kurdish carpets: their colours, borders and repeating shapes. One day he weaves a carpet nobody has seen before, yet everyone can tell it is Kurdish. He did not copy one carpet; he learned the style. A generative model does the same with text or pictures.

شاگردێکی گەنج لە کارگەیەکی قاڵیچنی لە سنە سەیری هەزاران قاڵیی کوردی دەکات: ڕەنگەکانیان، چوارچێوەکانیان و ئەو شێوانەی دووپاتە دەبنەوە. ڕۆژێک قاڵییەک دەچنێت کە کەس پێشتر نەیبینیوە، بەڵام هەموو کەس دەزانێت کوردییە. هیچ قاڵییەکی کۆپی نەکردووە؛ شێوازەکەی فێر بووە. بەرهەمهێنەر هەمان کار بە دەق یان وێنە دەکات.

Where the analogy stops workingسنووری ئەم نموونەیە

Unlike the apprentice, a model has no eyes, hands or taste: it learns from numbers. It can also repeat parts of its examples too closely, which raises copyright questions.

بە پێچەوانەی شاگردەکە، نە چاوی هەیە، نە دەست و نە سەلیقە: لە ژمارەوە فێر دەبێت. هەروەها دەشێت بەشێک لە نموونەکانی زۆر لە نزیکەوە دووپاتە بکاتەوە و ئەمەش پرسیاری مافی لەبەرگرتنەوە دروست دەکات.

Left: nine small carpets, the examples. An arrow leads to the apprentice, labelled “the model”, who learns colours, borders and shapes. Another arrow leads to one new carpet in the same style, not a copy. Bottom: generative AI learns the patterns of many examples and makes new content of the same kind.چەپ: نۆ قاڵیی بچووک، نموونەکان. تیرێک دەچێت بۆ شاگردەکە کە نووسراوە «مۆدێل» و ڕەنگ، چوارچێوە و شێوەکان فێر دەبێت. تیرێکی تر دەچێت بۆ قاڵییەکی نوێ بە هەمان شێواز، نەک کۆپی. خوارەوە: ژیریی دەستکردی بەرهەمهێنەر شێوازی زۆر نموونە فێر دەبێت و ناوەڕۆکی نوێ لە هەمان جۆر دروست دەکات.
Learn the style from many examples, then make something new.شێوازەکە لە زۆر نموونەوە فێر بە، پاشان شتێکی نوێ دروست بکە.CêIT

What it can makeچی دروست دەکات

Generative models make five main kinds of content. Some models make one kind; many of the newest handle several. The table gives examples as of November 2026; the names change often, the kinds do not.

بەرهەمهێنەر پێنج جۆری سەرەکیی ناوەڕۆک دروست دەکەن. هەندێک یەک جۆر دروست دەکات؛ زۆرێک لە نوێترینەکان چەند جۆرێک. خشتەکە نموونەی تشرینی دووەمی ⁦2026⁩ دەدات؛ ناوەکان زوو زوو دەگۆڕێن، بەڵام جۆرەکان نا.

KindجۆرWhat it makesچی دروست دەکاتExamples (November 2026)نموونە (تشرینی دووەمی ⁦2026⁩)
Textدەقanswers, summaries, translations, storiesوەڵام، کورتە، وەرگێڕان، چیرۆکchat models such as Claude, GPT and Gemini گفتوگۆ وەک Claude، GPT و Gemini
Codeکۆدprograms, fixes, explanations of errorsبەرنامە، چاککردن، ڕوونکردنەوەی هەڵەکانthe same chat modelsهەمان گفتوگۆ
Picturesوێنەdrawings, posters, edited photosوێنەکێشان، پۆستەر، وێنەی دەستکاریکراوGPT-Image-2.6 (OpenAI), Nano Banana 2 (Google)⁦GPT-Image-2.6 (OpenAI)⁩، ⁦Nano Banana 2 (Google)⁩
Voice and musicدەنگ و مۆسیقاspeech from text, spoken chat, short musicدەنگ لە دەقەوە، گفتوگۆی دەنگی، مۆسیقای کورتGPT-Realtime-1.5 (OpenAI), Gemini 3.8 Flash TTS (Google)⁦GPT-Realtime-1.5 (OpenAI)⁩، ⁦Gemini 3.8 Flash TTS (Google)⁩
Videoڤیدیۆshort clips from a descriptionکلیپی کورت لە وەسفێکەوەseveral companies offer them; check their model pagesچەند کۆمپانیایەک پێشکەشی دەکەن؛ پەڕەی ببینە

A multimodal model takes in more than one kind of data, for example a photo and a question about it, and often answers in more than one kind. Most big chat models of 2026 read pictures and documents as well as text: every current Claude model takes text and images, and Gemini 3.8 Flash also takes audio, video and PDF files.

زیاتر لە یەک جۆر وەردەگرێت، بۆ نموونە وێنەیەک و پرسیارێک لەبارەی. زۆرجاریش بە زیاتر لە یەک جۆر وەڵام دەداتەوە. زۆربەی گەورەکانی گفتوگۆی ساڵی ⁦2026⁩ وێنە و بەڵگەنامەش وەک دەق دەخوێننەوە: هەموو ئێستای Claude دەق و وێنە وەردەگرێت و ⁦Gemini 3.8 Flash⁩ دەنگ، ڤیدیۆ و PDFـیش وەردەگرێت.

Five tiles: text (a page: a reply, a story), code (the symbol </>: a program), pictures (a mountain: a poster), voice and music (a sound wave and a note: a song, a voice) and video (a film frame with a play button: a short clip). Below: a multimodal model takes in more than one kind and often gives more than one kind back, for example a photo of a broken printer and a question about it in words.پێنج خانە: دەق (پەڕەیەک: وەڵامێک، چیرۆکێک)، کۆد (هێمای </>: بەرنامەیەک)، وێنە (چیایەک: پۆستەرێک)، دەنگ و مۆسیقا (شەپۆلی دەنگ و نۆتەیەک: گۆرانی، دەنگ) و ڤیدیۆ (چوارچێوەیەکی فیلم بە دوگمەی لێدان: کلیپێکی کورت). خوارەوە: مۆدێلێکی فرەجۆر زیاتر لە یەک جۆر وەردەگرێت و زۆرجار زیاتر لە یەک جۆریش دەداتەوە، بۆ نموونە وێنەی چاپەرێکی تێکچوو و پرسیارێک لەبارەی بە وشە.
Five kinds of output; a multimodal model mixes them.پێنج جۆر ؛ تێکەڵیان دەکات.CêIT

Predictive or generative?پێشبینیکەر یان بەرهەمهێنەر؟

Both kinds learn from data, as in the three kinds of machine learning. The difference is the output. A predictive model chooses from a fixed set of answers: a label such as spam, or a number such as tomorrow's temperature. This is classification or regression. A generative model writes or draws something new, so the number of possible answers is almost endless.

هەردوو جۆرەکە لە فێر دەبن، وەک لە سێ جۆرەکەی فێربوونی ئامێرـدا. جیاوازییەکە لە . پێشبینیکەر لە نێو کۆمەڵەیەکی جێگیری وەڵامدا هەڵدەبژێرێت: وەک ، یان ژمارەیەک وەک پلەی گەرمیی سبەی. ئەمە یان . بەرهەمهێنەر شتێکی نوێ دەنووسێت یان دەکێشێت، بۆیە ژمارەی وەڵامە گونجاوەکان نزیکەی بێکۆتایە.

Predictive model پێشبینیکەرGenerative model بەرهەمهێنەر
Outputa label or a number from a fixed set یان ژمارەیەک لە کۆمەڵەیەکی جێگیرnew text, code, pictures or soundدەق، کۆد، وێنە یان دەنگی نوێ
E-maila spam filter: spam or not spamفلتەری : یان نییەa chatbot writes a reply to the e-mail وەڵامی دەنووسێت
Shilan's bookshopکتێبخانەکەی شیلانsorts each message: order, question or complaintهەر نامەیەک پۆل دەکات: داواکردن، پرسیار یان سکاڵاwrites a polite answer to the messageوەڵامێکی ڕێکوپێک بۆ نامەکە دەنووسێت
How you check itچۆن دەیپشکنیتcount right and wrong labelsلەیبڵە ڕاست و هەڵەکان بژمێرەread it: is it true, clear and safe?بیخوێنەرەوە: ڕاستە، ڕوونە و بێ مەترسییە؟

The two often work together: Shilan's helper first sorts a message, then writes the reply.

زۆرجار هەردووکیان پێکەوە کار دەکەن: یاریدەدەرەکەی شیلان سەرەتا نامەیەک پۆل دەکات، پاشان وەڵامەکە دەنووسێت.

Top: a customer writes to the bookshop, “Do you have the new Kurdish dictionary? How much is it?”. Left, a predictive model picks one label from a fixed list (order, question, complaint) and answers “question”. Right, a generative model writes new text: “Yes, we have it. It costs 15,000 dinars. Shall we keep a copy for you?”, a reply that did not exist before.سەرەوە: کڕیارێک بۆ کتێبخانەکە دەنووسێت «فەرهەنگە کوردییە نوێیەکەتان هەیە؟ نرخی چەندە؟». چەپ، مۆدێلێکی پێشبینیکەر یەک نیشانە لە لیستێکی جێگیر هەڵدەبژێرێت (داواکردن، پرسیار، سکاڵا) و دەڵێت «پرسیار». ڕاست، مۆدێلێکی بەرهەمهێنەر دەقی نوێ دەنووسێت: «بەڵێ، هەمانە. نرخەکەی 15,000 دینارە. دانەیەکت بۆ هەڵبگرین؟»، نامەیەک کە پێشتر نەبووە.
One model picks a label; the other writes something new. هەڵدەبژێرێت؛ ئەوی تر شتێکی نوێ دەنووسێت.CêIT

How a chat assistant is madeیاریدەدەرێکی گفتوگۆ چۆن دروست دەکرێت

A chat assistant such as ChatGPT, Claude or Gemini is made in three stages:

یاریدەدەرێکی گفتوگۆ وەک ChatGPT، Claude یان Gemini لە سێ قۆناغدا دروست دەکرێت:

  1. Pre-training. The model reads a huge library of text and code. Again and again it guesses the next small piece of text (a token, lesson 3) and is corrected; choosing the next piece is a probability question. Afterwards it can continue any text, but it is not yet a helpful assistant.
  2. Fine-tuning. People write thousands of good example answers to instructions, and the model trains a little more on them. Now it follows instructions.
  3. Human feedback (RLHF). People compare two answers to the same question and choose the better one. Thousands of such choices train the model, with reinforcement learning, to give the kind of answers people prefer.
  1. . کتێبخانەیەکی زۆر گەورەی دەق و کۆد دەخوێنێتەوە. دیسان و دیسان پارچە دەقێکی بچووکی داهاتوو (، وانەی ⁦3⁩) مەزەندە دەکات و هەڵەکەی بۆ ڕاست دەکرێتەوە؛ هەڵبژاردنی پارچەی داهاتوو پرسیارێکی ئەگەرە. دوای ئەمە دەتوانێت هەر دەقێک درێژە پێبدات، بەڵام هێشتا یاریدەدەرێکی بەسوود نییە.
  2. . خەڵک هەزاران نموونەی وەڵامی باش بۆ ڕێنمایی دەنووسن و کەمێکی تر لەسەریان ڕادەهێنرێت. ئێستا پەیڕەوی ڕێنمایی دەکات.
  3. ⁦(RLHF)⁩. خەڵک دوو وەڵامی هەمان پرسیار بەراورد دەکەن و باشترەکەیان هەڵدەبژێرن. هەزاران هەڵبژاردنی لەو جۆرە بە فێربوونی بەهێزکردنەوە ڕادەهێنن، تاکو ئەو جۆرە وەڵامانە بداتەوە کە خەڵک پێیان باشترە.

The networks inside these models, and how training changes their numbers, are taught in Neural networks and deep learning.

ئەو تۆڕانەی ناو ئەم و چۆنیەتیی گۆڕینی ژمارەکانیان لە کاتی ڕاهێناندا، لە تۆڕە دەمارییەکان و فێربوونی قووڵـدا دەخوێنرێن.

Three boxes joined by arrows. 1 Pre-training: reads a huge library of text and learns to predict the next token, for weeks or months on thousands of GPUs (the apprentice looks at thousands of carpets). 2 Fine-tuning: studies thousands of good example answers that people wrote and learns to follow instructions (a master weaver shows a few good samples). 3 Human feedback (RLHF): people compare two answers and choose the better one again and again, so it learns the answers people prefer (customers say which carpet they like more). Below: a helpful assistant (a chatbot), fluent and polite, but it can still be wrong.سێ چوارگۆشە کە بە تیر بەیەکەوە بەستراون. 1 پێشڕاهێنان: کتێبخانەیەکی زۆر گەورەی دەق دەخوێنێتەوە و فێری پێشبینیکردنی تۆکنی داهاتوو دەبێت، بۆ چەند هەفتە یان مانگێک لەسەر هەزاران یەکەی چارەسەرکردنی وێنەیی (شاگردەکە سەیری هەزاران قاڵی دەکات). 2 ڕێکخستنی ورد: هەزاران نموونەی وەڵامی باش دەخوێنێت کە خەڵک نووسیویانە و فێری پەیڕەوکردنی ڕێنمایی دەبێت (وەستایەک چەند نموونەیەکی باشی نیشان دەدات). 3 هەڵسەنگاندنی خەڵک (RLHF): خەڵک دوو وەڵام بەراورد دەکەن و دیسان و دیسان باشترەکە هەڵدەبژێرن، بۆیە فێری ئەو وەڵامانە دەبێت کە خەڵک پێیان باشترە (کڕیاران دەڵێن کام قاڵییان پێ باشترە). خوارەوە: یاریدەدەرێکی بەسوود (چاتبۆت)، ڕەوان و ڕێکوپێک، بەڵام هێشتا دەکرێت هەڵە بکات.
Pre-training gives knowledge of language; fine-tuning and human feedback give manners. زانینی زمان دەدات؛ و هەڵسەنگاندنی خەڵک ڕەفتاری باش دەدەن.CêIT

Foundation models: one model, many usesمۆدێلە بنچینەییەکان: یەک مۆدێل، زۆر بەکارهێنان

A model trained this way is so general that one model can serve many apps: a chat app, a translator, a coding helper, a bookshop assistant. In 2021 researchers at Stanford named such models foundation models: models trained on broad data at scale that can be adapted to a wide range of tasks. The EU AI Act calls them general-purpose AI models and gives their makers duties (rules for general-purpose AI models).

کە بەم شێوەیە ڕاهێنرابێت ئەوەندە گشتییە کە یەک دەتوانێت خزمەتی زۆر بکات: گفتوگۆ، وەرگێڕێک، یاریدەدەرێکی ، یاریدەدەرێکی کتێبخانە. لە ساڵی ⁦2021⁩دا توێژەرانی زانکۆی ستانفۆرد ناوی ئەم جۆرە نا : کە لەسەر فراوان و بە قەبارەیەکی گەورە ڕاهێنراوە و دەکرێت بۆ ئەرکی زۆر جیاواز بگونجێنرێت. پێیان دەڵێت گشتی و ئەرکیان دەخاتە سەر دروستکەرەکانیان (یاساکانی مۆدێلە گشتییەکان).

Definitionپێناسە

A very large trained on huge and varied , so that the same model can be adapted to many different tasks: chatting, translating, summarising, writing code. are the best-known examples.

زۆر گەورەیە کە لەسەر بڕێکی زۆر زۆری جۆراوجۆر ڕاهێنراوە، بۆ ئەوەی هەمان مۆدێل بۆ چەندین ئەرکی جیاواز بگونجێنرێت: گفتوگۆ، وەرگێڕان، کورتکردنەوە، نووسینی کۆد. ناسراوترین نموونەی ئەم جۆرە مۆدێلانە .

Large language models are the best-known foundation models. An app adapts the same model with its own instructions (prompts, lesson 5), sometimes with fine-tuning, instead of training a new model from the start.

ناسراوترین . هەمان بە ڕێنماییەکانی خۆی دەگونجێنێت (، وانەی ⁦5⁩)، هەندێک جاریش بە ، لە جیاتی ئەوەی نوێ لە سەرەتاوە ڕابهێنێت.

A circle in the middle, “one foundation model, trained once”, with lines to six apps: a chat app, translation, a coding helper, summaries of reports, Shilan's bookshop helper and a college help desk. Bottom: trained once at great cost; adapted for each app with prompts or fine-tuning.بازنەیەک لە ناوەڕاستدا، «یەک مۆدێلی بنچینەیی، یەک جار ڕاهێنراوە»، بە هێڵ بە شەش ئەپەوە بەستراوە: ئەپی گفتوگۆ، وەرگێڕان، یاریدەدەری کۆدنووسین، کورتەی ڕاپۆرت، یاریدەدەری کتێبخانەکەی شیلان و مێزی یارمەتیی کۆلێژ. خوارەوە: یەک جار بە تێچوویەکی زۆر ڕادەهێنرێت؛ بۆ هەر ئەپێک بە هەوێنەکان یان ڕێکخستنی ورد دەگونجێنرێت.
One expensive model at the centre, many cheap adaptations around it.یەک گران لە ناوەڕاست، زۆر گونجاندنی هەرزان لە دەوری.CêIT

Train once, use many timesیەک جار ڕاهێنان، زۆر جار بەکارهێنان

Why do only a few companies train foundation models? Because pre-training is enormous. Meta's 2024 report on its largest Llama 3 model (405 billion weights) says it read 15.6 trillion tokens of text and was trained on up to 16,000 GPUs (what a GPU is) in data centres (inside a data centre). That costs a great deal of money and electricity.

بۆچی تەنیا چەند کۆمپانیایەکی کەم ڕادەهێنن؟ چونکە زۆر گەورەیە. ساڵی ⁦2024⁩ی Meta لەبارەی گەورەترین ⁦Llama 3 (405⁩ ملیار ) دەڵێت ⁦15.6⁩ تریلیۆن دەق خوێندووەتەوە و لەسەر تا ⁦16,000⁩ (یەکەی چارەسەرکردنی وێنەیی چییە) لە (ناو ناوەندێکی دراوە) ڕاهێنراوە. ئەمە پارە و کارەبایەکی زۆری تێدەچێت.

Using a trained model is called inference. Training happens once and costs a lot; inference happens every time someone asks, and each request costs a little. The prices on model pages (next lesson) are inference prices.

بەکارهێنانی ڕاهێنراو پێی دەوترێت . ڕاهێنان یەک جار ڕوو دەدات و زۆر گرانە؛ هەموو جارێک ڕوو دەدات کە کەسێک پرسیار دەکات و هەر کەمێک تێدەچێت. نرخەکانی سەر پەڕەی (وانەی داهاتوو) نرخی .

TrainingڕاهێنانInference
How oftenچەند جارonce, plus later updatesیەک جار، لەگەڵ نوێکردنەوەی دواترevery requestهەموو
How longچەند دەخایەنێتweeks or monthsچەند هەفتە یان مانگێکsecondsچەند چرکەیەک
Who paysکێ پارە دەداتthe company that makes the modelئەو کۆمپانیایەی دروست دەکاتthe user or the app, per request or by subscriptionبەکارهێنەر یان ، بۆ هەر یان بە
Machinesئامێرەکانthousands of GPUs in data centresهەزاران لە GPUs in data centres, or your own computer for small models لە ، یان خۆت بۆ بچووک

Pictures from noise: diffusion modelsوێنە لە خاڵخاڵییەوە: مۆدێلەکانی بڵاوبوونەوە

Many picture generators use a diffusion model. In training, the model sees many pictures with random speckles (noise) added to them, a little more at each step, and it learns to remove the noise again. To make a new picture, it starts from pure noise and removes noise step by step, guided by your prompt, for example a snowy mountain at sunrise. A 2020 paper by Ho, Jain and Abbeel showed that this gives high-quality pictures; it used 1,000 small steps.

زۆرێک لە دروستکەرانی وێنە بەکار دەهێنن. لە کاتی ڕاهێناندا، زۆر وێنە دەبینێت کە خاڵخاڵیی هەڕەمەکییان تێکراوە (لە هەر هەنگاوێکدا کەمێک زیاتر) و فێر دەبێت خاڵخاڵییەکە دیسان لاببات. بۆ دروستکردنی وێنەیەکی نوێ، لە خاڵخاڵیی پەتییەوە دەست پێدەکات و هەنگاو بە هەنگاو خاڵخاڵی لادەبات، بە ڕێنمایی ، بۆ نموونە چیایەکی بەفرین لە خۆرهەڵاتندا. توێژینەوەیەکی ساڵی ⁦2020⁩ی Ho، Jain و Abbeel نیشانی دا کە ئەمە وێنەی کوالیتی بەرز دەدات؛ ⁦1,000⁩ هەنگاوی بچووکی بەکار دەهێنا.

Each run starts from different noise, so the same prompt gives a different picture each time. Companies do not always say which method their newest picture models use. The same power can make fake photos of real people: see deepfakes.

هەر جارێک لە خاڵخاڵییەکی جیاوازەوە دەست پێدەکات، بۆیە هەمان هەر جارە وێنەیەکی جیاواز دەدات. کۆمپانیاکان هەمیشە ناڵێن نوێترین وێنەیان چ ڕێگایەک بەکار دەهێنن. هەمان توانا دەتوانێت وێنەی ساختەی خەڵکی ڕاستەقینە دروست بکات: دیپفەیکەکان ببینە.

An animation in six steps beside the prompt “a snowy mountain at sunrise”. Step 0 of 1,000 shows only random coloured speckles. By step 200 the speckles are a little calmer; by step 400 big shapes appear, sky and mountain; by step 600 the sun and the snow are clear; by step 800 fine details are added; at step 1,000 there is a new picture of two snowy mountains and an orange sun that nobody drew. A progress bar fills as the steps go on.
An animation in six steps beside the prompt “a snowy mountain at sunrise”. Step 0 of 1,000 shows only random coloured speckles. By step 200 the speckles are a little calmer; by step 400 big shapes appear, sky and mountain; by step 600 the sun and the snow are clear; by step 800 fine details are added; at step 1,000 there is a new picture of two snowy mountains and an orange sun that nobody drew. A progress bar fills as the steps go on.
A diffusion model removes noise step by step until a new picture appears. هەنگاو بە هەنگاو خاڵخاڵی لادەبات تا وێنەیەکی نوێ دەردەکەوێت.CêIT

A short historyمێژوویەکی کورت

Generative AI did not appear overnight. Each idea on the timeline grew out of the one before: from hand-written rules, to counting which word comes next, to turning words into numbers, to one huge model that you steer with words.

لە شەوێکدا دەرنەکەوت. هەر بیرۆکەیەک لەسەر هێڵی کاتەکە لە بیرۆکەی پێشترەوە سەری هەڵدا: لە یاسای دەستنووسەوە، بۆ ژماردنی ئەوەی کام وشە دواتر دێت، بۆ گۆڕینی وشە بە ژمارە، بۆ یەک زۆر گەورە کە بە وشە ئاڕاستەی دەکەیت.

A timeline with nine points. 1966: ELIZA, fixed rules, no learning. 1970s to 2000s: n-gram models count which word comes next. 2013: word2vec, words become lists of numbers. 2017: the transformer, attention lets words look at each other. 2020: GPT-3, 175 billion weights, learns from examples in the prompt. November 2022: ChatGPT, chat with a model for everyone. March 2023: GPT-4 reads pictures as well as text. September 2024: o1-preview “thinks” before it answers. 2025 to 2026: AI agents, models use tools in a loop.هێڵێکی کات بە نۆ خاڵ. 1966: ELIZA، یاسای جێگیر، بێ فێربوون. 1970 تا 2000: مۆدێلەکانی n-gram دەژمێرن کام وشە دوای کامیان دێت. 2013: word2vec، وشە دەبێتە لیستێک ژمارە. 2017: ترانسفۆرمەر، سەرنجی خۆیی وا دەکات وشەکان سەیری یەکتر بکەن. 2020: GPT-3، 175 ملیار کێش، لە نموونەی ناو هەوێندا فێر دەبێت. تشرینی دووەمی 2022: ChatGPT، گفتوگۆ لەگەڵ مۆدێل بۆ هەمووان. ئازاری 2023: GPT-4 وێنەش وەک دەق دەخوێنێتەوە. ئەیلوولی 2024: o1-preview پێش وەڵامدانەوە «بیر دەکاتەوە». 2025 تا 2026: ئەیجێنتەکان، مۆدێلەکان لە خولێکدا ئامراز بەکار دەهێنن.
Sixty years from ELIZA's rules to agents that use tools.شەست ساڵ لە یاساکانی ELIZAـەوە تا ئەو ئامراز بەکار دەهێنن.CêIT

Read more: ELIZA in the Turing test lesson, next-word counting in predictive text, the transformer in RNNs and transformers, today's chatbots in Chatbots today and the wider story of AI in What is AI?. ChatGPT was released on 30 November 2022 and reached 100 million users in about two months.

زیاتر بخوێنەوە: ELIZA لە وانەی تاقیکاریی تیورینگ، ژماردنی وشەی داهاتوو لە پێشبینیی دەق، لە RNN و ترانسفۆرمەرەکان، ئەمڕۆ لە چاتبۆتەکان ئەمڕۆ و چیرۆکی فراوانتری لە ژیریی دەستکرد چییە؟. ChatGPT لە ⁦30⁩ی تشرینی دووەمی ⁦2022⁩ بڵاو کرایەوە و لە نزیکەی دوو مانگدا گەیشتە ⁦100⁩ ملیۆن بەکارهێنەر.

What a model knows, and what it does notمۆدێل چی دەزانێت و چی نازانێت

A language model has never seen a computer, a car, the sun or a person. Everything it “knows” comes from its training data: what people wrote and, for multimodal models, pictures, sound and video. It has no feelings: it is not kind or angry, even when its words sound so. And it does not look up stored answers: it writes new text, token by token, from the patterns it learned. So it can sound sure and still be wrong. NIST calls this confabulation; most people say hallucination (limits and safety). Hidden bias is another risk: see chatbots and large language models: five questions.

هەرگیز ، ئۆتۆمبێل، خۆر یان مرۆڤێکی نەبینیوە. هەموو ئەوەی «دەیزانێت» لە هاتووە: ئەوەی خەڵک نووسیویانە و بۆ وێنە، دەنگ و ڤیدیۆش. هەستی نییە: میهرەبان یان تووڕە نییە، تەنانەت کاتێک وشەکانی وا دەردەکەون. هەروەها وەڵامی هەڵگیراو ناگەڕێتەوە: لە شێوازە فێربووەکانەوە، بە ، دەقی نوێ دەنووسێت. بۆیە دەکرێت دڵنیا دەربکەوێت و هێشتا هەڵە بێت. NIST پێی دەڵێت confabulation؛ زۆربەی خەڵک بە ئینگلیزی دەڵێن hallucination، بە کوردی (سنوور و سەلامەتی). لاسەنگیی شاراوە مەترسییەکی ترە: چاتبۆتەکان و مۆدێلە گەورەکانی زمان: پێنج پرسیار ببینە.

A model also knows little about anything after its training data ends (its knowledge cutoff, next lesson), unless an app gives it fresh information. That is why a clear, detailed request helps so much: it gives the model what it cannot guess (prompt engineering).

هەروەها کەم دەزانێت لەبارەی هەر شتێک کە دوای کۆتایی ڕوویداوە (، وانەی داهاتوو)، مەگەر زانیاریی نوێی پێبدات. بۆیە داوایەکی ڕوون و ورد ئەوەندە یارمەتی دەدات: ئەوەی پێدەدات کە ناتوانێت مەزەندەی بکات (ئەندازیاریی هەوێن).

Four panels joined by arrows. The world: a sun, a mountain and a car. People write and take photos: a person, a page and a camera. Training data: a pile of pages. The model: patterns of words and pixels. Below: it never saw the sun or a car, it read what people wrote about them; no eyes, no feelings, no stored answers, it writes new text from the patterns it learned.چوار بەش کە بە تیر بەیەکەوە بەستراون. جیهان: خۆر، چیا و ئۆتۆمبێلێک. خەڵک دەنووسن و وێنە دەگرن: کەسێک، پەڕەیەک و کامێرایەک. دراوەی ڕاهێنان: کۆمەڵێک پەڕە. مۆدێل: شێوازی وشە و پیکسڵ. خوارەوە: هەرگیز خۆر یان ئۆتۆمبێلی نەبینیوە، ئەوەی خوێندووەتەوە کە خەڵک لەبارەیانەوە نووسیویانە؛ نە چاو، نە هەست، نە وەڵامی هەڵگیراو، لە شێوازە فێربووەکانەوە دەقی نوێ دەنووسێت.
The model meets the world only through what people wrote and recorded. تەنیا لە ڕێگەی ئەوەی خەڵک نووسیویانە و تۆماریان کردووە جیهان دەناسێت.CêIT

Generative AI and Kurdishژیریی دەستکردی بەرهەمهێنەر و زمانی کوردی

Models learn mostly from text on the internet, and there is far less Kurdish text than English text. NIST's 2024 profile warns that a model “may perform less well for non-English languages or certain dialects”. In practice, a Kurdish answer may mix Sorani and Kurmanji, borrow Persian or Arabic words, misspell, or invent technical words. Kurdish text is also cut into more tokens, so it costs more and fills the model's working memory sooner (why Kurdish costs more).

زیاتر لە دەقی سەر فێر دەبن و دەقی کوردی زۆر کەمترە لە دەقی ئینگلیزی. پرۆفایلەکەی NIST لە ساڵی ⁦2024⁩دا ئاگادار دەکاتەوە کە دەشێت بۆ زمانە نائینگلیزییەکان یان هەندێک شێوەزار خراپتر کار بکات. لە ڕاستیدا، دەشێت وەڵامێکی کوردی سۆرانی و کرمانجی تێکەڵ بکات، وشەی فارسی یان عەرەبی بخوازێت، هەڵەی ڕێنووسی تێدا بێت، یان وشەی تەکنیکی دابهێنێت. هەروەها دەقی کوردی بە زیاتر دابەش دەکرێت، بۆیە تێچووی زیاترە و زووتر بیرگەی کارکردنی پڕ دەکاتەوە (بۆچی کوردی گرانترە).

So use it, but check it. Name the dialect and script in your request, take technical words from a trusted list (such as CêIT's terms), and ask a teacher or a dictionary when a word looks strange.

کەواتە بەکاری بهێنە، بەڵام بیپشکنە. لە داواکەتدا ناوی شێوەزار و ئەلفوبێکە بنووسە، وشە تەکنیکییەکان لە لیستێکی متمانەپێکراو وەربگرە (وەک زاراوەکانی CêIT) و کاتێک وشەیەک سەیر دەردەکەوێت لە مامۆستایەک یان فەرهەنگێک بپرسە.

Left, an illustration that is not to scale: three bars for what a model read, a long bar for English, a shorter one for other big languages and a tiny one for Kurdish; companies rarely publish exact shares. Right, four checks for Kurdish: write “Central Kurdish (Sorani), Arabic script” in your request; expect mixed dialects and odd technical words; check spelling and terms with a teacher or a dictionary; Kurdish text uses more tokens (lesson 3).چەپ، وێنەیەکی نموونەیی کە بە پێوەر نییە: سێ هێڵ بۆ ئەوەی مۆدێل خوێندوویەتییەوە، هێڵێکی درێژ بۆ ئینگلیزی، کورتتر بۆ زمانە گەورەکانی تر و زۆر بچووک بۆ کوردی؛ کۆمپانیاکان کەم ڕێژەی ورد بڵاو دەکەنەوە. ڕاست، چوار پشکنین بۆ کوردی: لە داواکەتدا بنووسە «کوردیی ناوەندی (سۆرانی)، بە ئەلفوبێی عەرەبی»؛ چاوەڕێی تێکەڵبوونی شێوەزار و وشەی تەکنیکیی سەیر بکە؛ ڕێنووس و زاراوەکان لەگەڵ مامۆستا یان فەرهەنگێک بپشکنە؛ دەقی کوردی تۆکنی زیاتر دەبات (وانەی 3).
Less Kurdish text in training means more checking of Kurdish answers.دەقی کوردیی کەمتر لە ڕاهێناندا واتە پشکنینی زیاتری وەڵامە کوردییەکان.CêIT

Common mistakesهەڵە باوەکان

Exercisesڕاهێنانەکان

ExerciseڕاهێنانGenerative or predictive?بەرهەمهێنەر یان پێشبینیکەر؟

Sort these eight tools: (a) a keyboard that suggests your next word; (b) an app that writes a cover letter; (c) a camera that finds faces; (d) a tool that draws a logo from a description; (e) a filter that moves e-mails to spam; (f) a model that reads text aloud in a natural voice; (g) a system that predicts tomorrow's electricity use; (h) a helper that writes Python code from a request.

ئەم هەشت ئامرازە پۆل بکە: ⁦(a)⁩ کە وشەی داهاتووت پێشنیار دەکات؛ ⁦(b)⁩ کە نامەیەکی داواکاریی کار دەنووسێت؛ ⁦(c)⁩ کامێرایەک کە ڕووخسار دەدۆزێتەوە؛ ⁦(d)⁩ ئامرازێک کە لە وەسفێکەوە لۆگۆیەک دەکێشێت؛ ⁦(e)⁩ فلتەرێک کە دەگوازێتەوە بۆ ؛ ⁦(f)⁩ کە دەق بە دەنگێکی سروشتی دەخوێنێتەوە؛ ⁦(g)⁩ سیستەمێک کە بەکارهێنانی کارەبای سبەی پێشبینی دەکات؛ ⁦(h)⁩ یاریدەدەرێک کە لە داوایەکەوە کۆدی Python دەنووسێت.

Show answerوەڵامەکە پیشان بدە

Generative: (b), (d), (f), (h): each makes new content. Predictive: (c), (e), (g): each picks a label or a number. (a) sits in between: it predicts the next word from a fixed vocabulary, which is a tiny language model. Repeating that step word after word is exactly how chat models write (lesson 3).

بەرهەمهێنەر: ⁦(b)⁩، ⁦(d)⁩، ⁦(f)⁩، ⁦(h)⁩: هەریەکەیان ناوەڕۆکی نوێ دروست دەکات. پێشبینیکەر: ⁦(c)⁩، ⁦(e)⁩، ⁦(g)⁩: هەریەکەیان یان ژمارەیەک هەڵدەبژێرێت. ⁦(a)⁩ لە نێوانیاندایە: وشەی داهاتوو لە نێو وشەنامەیەکی جێگیردا پێشبینی دەکات، کە زۆر بچووکە. دووپاتەکردنەوەی ئەو هەنگاوە وشە بە وشە ڕێک ئەوەیە کە گفتوگۆ پێی دەنووسن (وانەی 3).

ExerciseڕاهێنانMatch the stageقۆناغەکە بدۆزەوە

Which training stage is each description? (1) Reviewers see two answers and choose the clearer, more honest one. (2) The model guesses the next token in billions of sentences. (3) The model trains on 10,000 good answers that teachers wrote.

هەر وەسفێک کام قۆناغی ڕاهێنانە؟ ⁦(1)⁩ هەڵسەنگێنەران دوو وەڵام دەبینن و ڕوونتر و ڕاستگۆترەکەیان هەڵدەبژێرن. ⁦(2)⁩ داهاتوو لە ملیارەها ڕستەدا مەزەندە دەکات. ⁦(3)⁩ لەسەر ⁦10,000⁩ وەڵامی باش ڕادەهێنرێت کە مامۆستایان نووسیویانە.

Show answerوەڵامەکە پیشان بدە

(1) Human feedback (RLHF). (2) Pre-training. (3) Fine-tuning. The real order is (2), then (3), then (1).

⁦(1)⁩ ⁦(RLHF). (2)⁩ . ⁦(3)⁩ . ڕیزبەندیی ڕاستەقینە ⁦(2)⁩، پاشان ⁦(3)⁩، پاشان ⁦(1)⁩ـە.

ExerciseڕاهێنانTraining or inference?ڕاهێنان یان کارپێکردنی مۆدێل؟

Decide for each: (a) a company runs thousands of GPUs for two months on trillions of tokens; (b) Aram asks a chatbot to explain recursion; (c) Shilan's helper answers 300 messages a day.

بۆ هەریەکەیان بڕیار بدە: ⁦(a)⁩ کۆمپانیایەک بۆ ماوەی دوو مانگ هەزاران لەسەر تریلیۆنەها کار پێدەکات؛ ⁦(b)⁩ ئارام داوا لە دەکات ڕوون بکاتەوە؛ ⁦(c)⁩ یاریدەدەرەکەی شیلان ڕۆژانە وەڵامی ⁦300⁩ نامە دەداتەوە.

Show answerوەڵامەکە پیشان بدە

(a) Training (pre-training): huge, once. (b) and (c) Inference: the trained model answers each request, and each one costs a little.

⁦(a)⁩ ڕاهێنان (): زۆر گەورە، یەک جار. ⁦(b)⁩ و ⁦(c)⁩ : ڕاهێنراوەکە وەڵامی هەر دەداتەوە و هەریەکەیان کەمێک تێدەچێت.

Summaryپوختە

  • Generative AI creates new content (text, code, pictures, voice, music, video) from patterns it learned; predictive AI picks a label or a number.
  • A multimodal model takes in (and often gives back) more than one kind of data.
  • Chat assistants are made in three stages: pre-training (predict the next token on a huge library), fine-tuning (good examples), human feedback / RLHF (people choose the better answer).
  • A foundation model is trained once at great cost and adapted to many apps; using it is inference, paid per request.
  • A diffusion model turns random noise into a picture, step by step, guided by the prompt.
  • A model never saw the world: it learned from what people wrote, has no feelings, can be confidently wrong, and is weaker in Kurdish, so check its answers.
  • لەو شێوازانەوە کە فێری بووە ناوەڕۆکی نوێ دروست دەکات (دەق، کۆد، وێنە، دەنگ، مۆسیقا، ڤیدیۆ)؛ پێشبینیکەر یان ژمارەیەک هەڵدەبژێرێت.
  • زیاتر لە یەک جۆر وەردەگرێت (و زۆرجار دەیداتەوە).
  • یاریدەدەرەکانی گفتوگۆ لە سێ قۆناغدا دروست دەکرێن: (پێشبینیکردنی داهاتوو لەسەر کتێبخانەیەکی زۆر گەورە)، (نموونەی باش)، (خەڵک وەڵامی باشتر هەڵدەبژێرن).
  • یەک جار بە تێچوویەکی زۆر ڕادەهێنرێت و بۆ زۆر دەگونجێنرێت؛ بەکارهێنانی و بۆ هەر پارەی دەدرێت.
  • هەنگاو بە هەنگاو و بە ڕێنمایی خاڵخاڵیی هەڕەمەکی دەکات بە وێنە.
  • هەرگیز جیهانی نەبینیوە: لەوەوە فێر بووە کە خەڵک نووسیویانە، هەستی نییە، دەکرێت بە دڵنیاییەوە هەڵە بکات و لە کوردیدا لاوازترە، بۆیە وەڵامەکانی بپشکنە.

Next: the models you can use today, how to read their pages and how to choose one: AI models today.

داهاتوو: ئەو ئەمڕۆ دەتوانیت بەکاریان بهێنیت، چۆن پەڕەکانیان بخوێنیتەوە و چۆن یەکێکیان هەڵبژێریت: مۆدێلەکانی ژیریی دەستکرد ئەمڕۆ.

Lesson complete!وانەکە تەواو بوو!

Well done. Review the key terms, then go on to the next lesson.دەستخۆش. سەیرێکی زاراوە سەرەکییەکان بکەرەوە، پاشان بڕۆ بۆ وانەی داهاتوو.

Key termsزاراوە سەرەکییەکان

Next lessonوانەی داهاتوو

AI Models Today: Claude, GPT, Gemini and Open Modelsمۆدێلەکانی ژیریی دەستکرد ئەمڕۆ: Claude، GPT، Gemini و مۆدێلە کێشکراوەکان

References and further readingسەرچاوەکان و خوێندنەوەی زیاتر

This lecture is a first edition. The Kurdish text and terminology are waiting for review by Kurdish IT teachers.ئەم وانەیە چاپی یەکەمە. دەقە کوردییەکە و زاراوەکانی چاوەڕێی پێداچوونەوەی مامۆستایانی کوردی بواری IT ن.

Appearanceڕووکار

Modeدۆخ
Styleشێواز
Coloursڕەنگەکان

Saved on this device only.تەنیا لەسەر ئەم ئامێرە پاشەکەوت دەکرێت.