LLM Primer (1 page)

📖 3 min read referencecheatsheetprimer Key Info

One-page LLM primer covering what LLMs are, how they work, and when to use each major model - Claude, GPT, Gemini, and DeepSeek.

On this page 14

OverviewWhat is an LLM? (one sentence)Key TermsMay 2026 Model ComparisonReal Capabilities vs Hype✅ What LLMs Actually Do Well⚠️ Improving (But Still Imperfect)❌ Can’t Do (Yet)3 Ways to Adapt Models1. Prompting2. Retrieval-Augmented Generation (RAG)3. Fine-TuningCommon MisconceptionsNext Steps

What is an LLM? (one sentence)

A Large Language Model is software that predicts the next word by learning patterns from billions of examples.

---

Key Terms

Term	Definition
Token	A piece of text (word, punctuation, subword). Models process tokens, not letters.
Parameter	A “dial” the model uses to predict. More parameters = more nuance.
Context	The conversation history the model can see (usually 4K–400K tokens).
Hallucination	When a model generates plausible-sounding but false information.
Temperature	Controls randomness (0 = deterministic, 1.0 = creative).
Top-p sampling	Limits predictions to the most likely options for quality control.

---

May 2026 Model Comparison

Model	Best For	Context	Cost	Training Data
Claude Opus 4.8	Writing, reasoning, analysis	1M tokens	$5/$25 per 1M input/output	Jan 2026
GPT-5.5	Speed, all-purpose	1M tokens	$5/$30 per 1M	Dec 2025
Gemini 3.1 Pro	Long documents, research	1M tokens	$2/$12 per 1M (free tier available)	Jan 2025
Claude Sonnet 4.6	Balanced, coding	1M tokens	$3/$15 per 1M	Jan 2026
DeepSeek V4	Cost-conscious teams	256K tokens	10–50x cheaper	Late 2024

---

Real Capabilities vs Hype

✅ What LLMs Actually Do Well

• Writing and editing (often better than most people)
• Explaining concepts (clear, accessible)
• Coding assistance (saves significant time)
• Analysis of existing text (pattern finding, extraction)
• Brainstorming and ideation (fast, creative)

⚠️ Improving (But Still Imperfect)

• Math (reasoning models like o3 are better, but still limited)
• Factual accuracy (hallucinations remain; add retrieval to fix)
• Real-time information (not built-in; Perplexity/Gemini add it)
• Domain expertise (fine-tuning helps)

❌ Can’t Do (Yet)

• Access the internet on their own
• Know anything after their training date
• Reliably do complex math
• Understand context beyond 200K tokens
• Guarantee truthfulness (they generate plausible text, not facts)
• Replace domain experts (they’re tools for experts)

---

3 Ways to Adapt Models

1. Prompting

Give better instructions. Works for most tasks.

• Time: Instant
• Cost: Free (uses existing model)
• When to use: Writing, analysis, coding, learning

2. Retrieval-Augmented Generation (RAG)

Feed the model your own data so it can answer questions about your documents.

• Time: Hours to days (build once, use forever)
• Cost: Cheap (just vector DB storage)
• When to use: Customer support, internal docs, proprietary knowledge

3. Fine-Tuning

Train the model on your examples so it learns your style/domain.

• Time: Days (one-time)
• Cost: Moderate ($100–$10K depending on scale)
• When to use: Specialized domains, specific format/tone, repeated use at scale

---

Common Misconceptions

• “LLMs understand language” → They pattern-match. Understanding is human projection.
• “LLMs are general intelligences” → They’re narrow: good at text, bad at reasoning under uncertainty.
• “ChatGPT is always right” → No. Verify important facts. They hallucinate.
• “LLMs will replace humans” → No. They’re tools. Humans + LLMs > either alone.
• “Training new models is cheap” → No. Billions in compute. Fine-tuning is accessible.

---

Next Steps

• Try one: Go to Claude.ai, ChatGPT, or Gemini
• Ask follow-up questions naturally; iterate
• Learn more: How LLMs Work
• Build something: Builder Path