AI Capabilities and Limitations
AI is brilliant one minute and confidently wrong the next. It is not random. It fails in predictable places.
Start with a test
Fifteen words, one at a time. Then write down every one you remember.
This test needs JavaScript switched on.
Most people keep the first few and the last few. The middle goes. AI does the same thing: researchers at Stanford found models use the start and end of a long input best, and the middle worst (Lost in the Middle, 2023).
What generative AI actually is
AI has been around for years, quietly sorting your spam. Generative AI is new. It writes, and it does it by predicting what comes next, one small piece at a time.
It is closer to autocomplete than to a search engine. It is not looking anything up.
The tools change every month. The four properties below do not. That is the point of this page: a framework that lasts.
How it gets its character
- 1
Pretraining. It reads a huge amount of text and learns what tends to follow what. Ask “What is the capital of France?” and it just carries on writing, like a document.
- 2
Fine-tuning. People teach it to be a helpful assistant. Now it answers: “Paris.”
Built to help, it picks up four habits. Tap each one.
The four properties
Next token prediction
where answers come from- Good at: well-worn paths. Summarising, reformatting, common ideas.
- Weak at: the niche and the new. It fills gaps with plausible filler.
- You: check names, dates, figures and citations.
Knowledge
what it actually knows- Good at: mainstream topics, well documented.
- Weak at: rare, local, and anything after its cutoff.
- You: verify anything recent, or give it search.
Working memory
what it is paying attention to- Good at: material that fits comfortably.
- Weak at: long chats and long documents. The middle gets lost.
- You: the right context, not more. Key things first. Start fresh.
Steerability
how much you are in control- Good at: short, checkable instructions. “A table.” “Under 100 words.”
- Weak at: vague asks and long chains of steps.
- You: give the goal with the steps. This is why SDS works.
A model only knows what it read before training ended. In September 2026 Claude Cowork and chat merged into one Claude. The model I was working with had no idea, and would have taught me the old version with total confidence. That is the cutoff.
How much should you check?
Two questions. Is this a well-worn path? And what happens if it is wrong?
Not sure where a task sits? Ask: is it recent? Does my context fit? Are my instructions concrete, or is there a gap between my words and what I mean?
Your four Ds meet its four properties
AI Fluency is about you. This page is about the model. Match each of your skills to the property it answers.
Pairs from the Anthropic course. The four Ds →
Based on the AI capabilities and limitations course by the Anthropic team. Thirteen lessons and a quiz. The rule of thumb and the examples are mine.
- Generative AI predicts. It does not look things up.
- Built to help, it agrees, rambles, hedges and bluffs.
- Four properties: prediction, knowledge, working memory, steerability.
- Check by stakes: well-worn and low stakes, use it. Niche and high stakes, verify it.
This week: take one answer that annoyed you and name which property caused it.
Why agreeing is the dangerous one: Dangers of AI → Close the gap between your words and your intent: How to Prompt →
Do it with someone. Everything here is free and you can run it alone. If you would rather work through it with me, one to one, book a free 30-minute chat or see lessons and prices. No pitch.