Free preview7 min

The Three Things AI Can and Can't Do

Reviewed by Human · Updated July 13, 2026

An AI can write a decent sonnet in four seconds. Ask it to multiply two 8-digit numbers, and it face-plants. Why?

Same tool, two very different jobs

✗ Using it as a database

What was the exact box office revenue of every movie released in 1994?

This demands precise recall of thousands of facts. The model doesn't look anything up — it predicts plausible-sounding numbers. Some will be right. Many will be confidently wrong. You can't tell which.

✓ Using it as a pattern engine

You are a film buff. Here are five 1994 movie summaries I pasted below. Group them by theme and write one sentence per group explaining the pattern. Keep it under 100 words.

Grouping, summarizing, and explaining are pure pattern work. You supplied the facts; the model supplies the synthesis. This is the sweet spot.

What Are LLMs Good At (and Bad At)?

Three genuine strengths: 1. **Pattern recognition**: Spotting structure in text you give it — tone, themes, style, errors. 2. **Synthesis**: Combining and reshaping provided information — summaries, comparisons, rewrites. 3. **Generation**: Producing fluent new text on demand — drafts, options, variations. Three real limits: 1. Exact facts. The model can state false things fluently and confidently. **Hallucination**: A confident AI output that is factually wrong or invented. Researchers have documented this failure mode extensively (arxiv.org/abs/2311.05232). 2. Precise calculation and long chains of reasoning. Word-guessing is not arithmetic. 3. Memory. As you saw last module — blank slate every chat, by default. Rule of thumb: bring your own facts, let it do the pattern work.

Anthropic's own docs describe model capabilities and limits in detail — worth a skim: docs.anthropic.com

Your turn

Sort these six tasks. Mark each one GOOD FIT or RISKY for an LLM working alone, and write three words explaining why.

Reflect

Odd numbers were pattern work. Even numbers leaned on exact recall or calculation — hallucination territory. Feel how fast the sorting gets once you know the trick underneath?

Which task plays directly to an LLM's actual strengths?

You now know the engine: prediction, not magic. You know the blind spots: facts, math, memory. Next up, you'll write your first real prompt — and you'll write it like someone who knows exactly what machine they're talking to.

📚

Go deeper

end of module

You've finished this module.

Mark it complete to earn your XP and keep your streak alive.

Progress saved locally · Sign up to earn XP

←Why It Sounds So Smart (But Isn't)