Before you begin: AI without an LLM

We will build a small program for short questions about an imaginary school club. It analyzes Kazakh words, finds an approved record, and answers from that record. Without a record it says “білмеймін” (“I don’t know”). You need a computer and Python 3.10 or newer; Lesson 2 explains setup. The course is free and needs neither a paid API nor a dedicated GPU.

How languages express the same idea

Consider “to our houses”: a house, plurality, possession, and direction. Languages place these meanings differently. Grammar is the rules for combining them; morphology concerns word structure. The labels below describe common patterns, not every word of a language.

• Analytic structure is like separate cards on a table: English to our houses puts direction and possession in the separate words to and our. Strong analytic traits occur in English, Mandarin Chinese, Vietnamese, Thai, and Khmer. A program must examine neighboring words to recover the whole meaning.
• Agglutinative structure is like food placed in order on a skewer: distinct parts attach to a meaningful stem. For analysis only, write Kazakh үйлерімізге as үй-лер-іміз-ге: house, plural, ours, toward. Ordinary spelling contains no hyphens.
• Fusional structure is like one card carrying two instructions: Russian столом uses -ом to express both singular number and instrumental case (“with what?”). Fusional traits are prominent in Russian, Polish, German, Spanish, and Lithuanian.

The skewer image does not mean that every stem letter stays fixed: кітап (“book”) becomes кітабым (“my book”), changing п to б. A program must include that rule too.

Ten languages with agglutinative word forms

A word borrowed from Persian can take ordinary Turkic endings. Its origin does not give us a percentage of a language’s “agglutinativeness”. We chose Kazakh because its forms make the steps visible, we can check examples ourselves, and our qazaq-ir and adam projects offer practical reference points. Other agglutinative languages would also work.

Why build AI without an LLM?

A large language model (LLM) can freely generate text. This is useful for many tasks, but it can produce a plausible claim unsupported by a source: a hallucination. Our system gives a factual answer only when it finds an approved record. Think of a librarian showing a catalog card instead of inventing a missing book. Errors in analysis, records, or answer selection remain possible.

A small local program may use fewer resources and keep questions off external services. “A thousand times lighter, cheaper, and faster” remains a hypothesis: at the end we will measure specific systems on the same task at comparable answer quality. The first five lessons establish the foundation; 30 lessons are planned.

Sources: WALS, Kazakh grammar description, Turkic languages and language contact, survey of factual errors in LLMs.