Lessons0 of 20 done

Module 2: Language models. Lesson 8 of 20, 7 minutes.

What they are good at, and what they are not.

The uneven shape of a model's abilities, which tasks to hand it with confidence, which to check, and why the same prompt can give two answers.

The same assistant that writes a graceful, well-structured cover letter in ten seconds can fail to count the words in it. It can translate a page of Portuguese and then get the sum of two numbers in that page wrong. If you expect a tool to be uniformly smart or uniformly dumb, this is baffling.

It stops being baffling once you know the shape of the thing. This lesson draws that shape, so you can hand a model the right work and keep an eye on the rest.

A jagged profile

With people, ability tends to come in bundles. Someone who can write a clear legal memo can certainly count the paragraphs in it. Language models break that expectation. Their abilities are jagged: extraordinary in some places and surprisingly weak in others, including places that look like they should be easy.

The reason, in one line, is that they learned by predicting text. They are strongest at tasks that are shaped like text (produce something that reads well and fits the request) and weakest at tasks that need exact, step-by-step tracking or facts they were never given. Once you know the profile, you stop being surprised and start placing work where it fits.

Where they are strong

These are tasks where producing text that fits is almost the whole job:

  • Drafting and rewriting. A first draft of an email, a report introduction, a job description. Or your rough paragraph made tighter, warmer or more formal.
  • Summarizing and explaining. A long document into five lines. A dense passage explained at the level of a smart fourteen-year-old.
  • Brainstorming options. Twenty names for a product, ten angles for an article, five ways to open a difficult conversation. Breadth is cheap.
  • Structuring messy notes. Paste a wall of meeting notes and ask for decisions, owners and open questions.
  • Translation. Between widely spoken languages, quality is high and tone survives.
  • Writing and explaining code. Code is text with strict patterns, which suits the model, and it can explain what an existing snippet does in plain words.
  • Adapting tone for an audience. The same facts for your CEO, your customers and your new hire.

Notice what these share. Each has many acceptable answers, and you are there to judge the result. The model proposes; you choose.

Where they are weak

  • Arithmetic and counting. A model does not calculate; it predicts the digits that usually follow. Small sums are usually right. Long multiplication, percentages of awkward numbers and counting the items in a list often go wrong. If the app can run a calculator or write code, ask it to.
  • Exact quotes. Unless the text is in front of it, the model reconstructs a quote rather than retrieving one, and reconstructions drift.
  • Recent events. Anything after the knowledge cutoff (the date its training text stopped) is unknown unless the app searches for it.
  • Knowing things about you it was never told. It does not know your company's naming conventions, your manager's preferences or what you meant by "the usual format." If it matters, say it.
  • Staying consistent over long tasks. Across a long document or a long conversation, details shift: a character's eye color, a column name, a rule you set in your second message. Early instructions get crowded out.
  • Judging its own confidence. As the last lesson covered, tone is not a reliability signal, and the model's own estimate of how sure it is only loosely tracks whether it is right.

Many of these look like they should be trivial, and that is the point. Counting is trivial for a calculator and hard for a text predictor.

The same prompt, a different answer

One more property surprises people. Ask the same question twice and you may get two different answers. This is by design. When the model picks the next token (the next piece of a word), it does not always take the single most likely one. It samples from the likely options with a little randomness, so that its writing does not become repetitive and flat. Developers call the dial that controls this randomness the temperature. For creative work, the randomness is useful. For factual work, it means a single answer is a single draw.

The practical rule: when it matters, ask twice. Open a fresh chat, ask again, and compare. If the two answers agree, your confidence can rise. If they disagree, you have learned something important before acting on either.

Putting work where it fits

Look back at the two lists and a pattern appears. The strong tasks have many acceptable answers and a human judging the output. The weak tasks have one exact answer and no built-in way for the model to check it. That suggests a simple division of labor.

Hand the model the work where language is the job: the draft, the summary, the options, the restructure, the translation. Keep for yourself, or for a proper tool, the work where exactness is the job: the number, the quote, the date, the fact about your own situation. And when a task mixes both, which most real tasks do, let the model write and let something else verify. A spreadsheet checks the arithmetic. A search confirms the date. You confirm the facts about your company.

None of this means the model is unreliable in general. It means its reliability has a shape, and you can see it.

Try it

Map the profile yourself. Give any chatbot these one at a time and grade each result.

  1. Rewrite this in a warmer tone: "Per my last email, the deadline was Friday." Expect a strong result.
  2. How many letters are in the word "accommodation"? Count it yourself before you trust the answer.
  3. What is 487 times 263? Check it with a calculator.
  4. Give me five possible titles for a talk about remote work. Expect a strong result.
  5. Give me an exact quote from a newspaper article published this week. Watch what it does: searches, declines, or invents.

Then send question 3 again in a fresh chat and see whether the answer changes. The exact scores will differ between products, but the shape will be the same: fluent where language is the job, unreliable where exactness is.

Key terms

Jagged profile
The uneven pattern of a model's abilities: very strong at some tasks and weak at others that look similar.
Sampling
How a model picks each next token from several likely options with some randomness, so the same prompt can produce different answers.
Temperature
The setting that controls how much randomness goes into the model's choice of each next token.
Drift
Details changing over a long task or conversation because earlier instructions and facts get crowded out.

Check yourself

1.Which task is a language model most reliably good at?
2.Why can the same prompt produce two different answers?
3.You need a number that will go into a report. Which habit fits the jagged profile?

Models are strong wherever producing fitting language is the job and weak wherever exactness is, so place the work accordingly and ask twice when it matters.