# Skill: Explain what an AI language model is and read model size names correctly

> Lets an AI coding/design agent ground explanations of LLM behavior, hallucinations, and parameter counts in the correct mental model; applies whenever a user asks "why did the model do that?" or when selecting a model by size.

Source: sodigi·learn — models/chto-takoe-model · https://sodigi.io/learn/models/chto-takoe-model

## When to use
- A user asks why a model "makes things up" or why answers vary between identical prompts.
- A user encounters a model name with a B-suffix (7B, 70B, 405B) and needs to interpret it.
- Selecting or recommending a model and needing to match size to task complexity.
- Explaining LLM limitations before starting a project or demo.

## Core rules
- Always frame a language model as a next-token predictor, not a database or a mind.
- "Parameters" and "weights" are synonyms -- the billions of tuning numbers that encode everything the model learned.
- B in a model name stands for billion; the number is the parameter count (7B = ~7 billion parameters).
- More parameters = smarter AND pricier AND slower; match size to task, never default to "biggest is best."
- Fabrication (hallucination) is not a bug -- it is a natural consequence of predicting likely continuations rather than retrieving facts.
- Token != word; models operate on sub-word chunks; responses are non-deterministic by design.
- Training = automatic gradient-based tuning of weights; no human sets them by hand.
- Never claim the model "knows" or "thinks"; it predicts.

## Procedure
1. When someone asks why a model invented a fact, explain: the model predicts the most probable next token, not the truthful one -- it has no lookup table.
2. When someone asks what "7B" or "70B" means, decode: B = billion, the number is the parameter count; larger = smarter but also pricier and slower.
3. When recommending a model tier, ask first: is the task simple/routine, mid-complexity, or genuinely hard multi-step? Then map: small = fast/cheap/routine; large = complex/expensive.
4. When a user expects deterministic identical answers, explain the temperature/randomness mechanism: each token pick has a probability distribution, so answers naturally drift.
5. When someone blames the model for "being dumb," check whether they used an undersized model for a hard task before diagnosing a prompt problem.

## Ready-to-use prompt
```
You are an expert at explaining AI concepts to non-technical people.
Explain the following in plain English, using no formulas:
- What a language model actually does (one core idea only)
- What parameters / weights are and why their count matters
- Why the model sometimes makes things up
- What "7B" in a model name means

Use a concrete everyday analogy. Keep it under 150 words.
```

## Pitfalls
- Saying the model "knows" or "stores" facts -- it does not; it predicts likely text.
- Equating parameter count with RAM or internet speed -- parameters are the model's internal weights, independent of infrastructure.
- Assuming bigger is always better -- oversized models waste money on routine tasks and can "overthink" simple inputs.
- Expecting the same prompt to produce byte-for-byte identical output every time -- randomness in token sampling is normal and tunable.
- Treating hallucination as a solvable bug to wait for -- it is structural to the prediction mechanism.
- Confusing "training" with "memory" -- after training, weights are frozen; the model does not learn from your conversation by default.