> ## Documentation Index
> Fetch the complete documentation index at: https://docs.aperium.apps.hillspire.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Capabilities & limits

> What AI is genuinely good and bad at today, and how to stay calibrated.

There are two common ways to get this wrong. One person takes a confident-sounding answer at face value, ships it, and gets burned when it turns out wrong. Another had one bad experience, decided "it makes stuff up," and now does everything by hand — leaving real time savings on the table every week. Neither is calibrated. This page sets the calibration.

## The reframe

It is not an oracle that knows things. Think of it as a **very well-read, very fast, occasionally overconfident colleague who never sleeps.** It will hand you a first draft in seconds, read a document faster than you ever could, and never get bored in a brainstorm. What it will not reliably do is know when it is wrong.

<Warning>
  It is a partner you think *with*, not an oracle. The accountability for the final output does not transfer to it — it stays with you.
</Warning>

Both the strengths and the limits below fall directly out of the [next-word-predictor mental model](/literacy/how-llms-work): it is excellent at reshaping and producing text, and unreliable anywhere the answer needs verified fact, live data, or judgment.

## The map

| Good at                          | Be careful with               |
| -------------------------------- | ----------------------------- |
| Drafting & rewriting             | Precise arithmetic & counting |
| Summarizing                      | Up-to-the-minute facts        |
| Explaining concepts              | Guaranteed accuracy           |
| Brainstorming                    | Its own blind spots           |
| Reformatting & transforming text | Judgment calls                |
| Extracting info from text        |                               |
| Translating tone                 |                               |

## Good at: reshaping text

The common thread across everything on the left is that the model is producing or reshaping text from text it can already see — exactly what a next-word predictor is built for.

<CardGroup cols={2}>
  <Card title="Drafting & rewriting" icon="pen">
    A first-pass follow-up email you then edit.
  </Card>

  <Card title="Summarizing" icon="list">
    Turning a 40-page report into five bullets.
  </Card>

  <Card title="Explaining" icon="lightbulb">
    Explaining a concept like NAV versus IRR to a new analyst.
  </Card>

  <Card title="Brainstorming" icon="comments">
    Generating due-diligence questions to consider.
  </Card>

  <Card title="Reformatting" icon="table">
    Cleaning up messy meeting notes into a table.
  </Card>

  <Card title="Extracting" icon="magnifying-glass">
    Pulling key terms out of an agreement.
  </Card>
</CardGroup>

## Be careful with

<AccordionGroup>
  <Accordion title="Precise arithmetic & counting">
    It predicts plausible tokens; it does not run a calculator under the hood. Ask it to sum a column of 40 numbers or count the words in a paragraph, and it can quietly get a digit wrong. **Rule of thumb:** math belongs in a spreadsheet or a query, not in a prompt. If a number matters, compute it.
  </Accordion>

  <Accordion title="Up-to-the-minute facts">
    Its training data has a cutoff. On its own, it answers from a frozen snapshot of the past — it does not know today's number, this week's close, or a policy that changed last month. Connected to a live source — a database, a document, a system — it can be current. That connection is exactly what Aperium adds over a generic chatbot.
  </Accordion>

  <Accordion title="Guaranteed accuracy">
    It has no built-in notion of true versus false, only plausible — so confident does not mean correct. The risk scales with how niche or internal the question is: the less the model has seen text like your specific case (an internal Hillspire policy, a one-off deal term), the more likely you get a fluent-sounding fabrication instead of an honest "I don't know."
  </Accordion>

  <Accordion title="Its own blind spots & judgment">
    Two things bundled here. First, it usually does not reliably know what it does not know — asked about a policy it has never seen, it will often answer confidently anyway rather than flag the gap. Second, real judgment calls — should we grant this exception, is this the right move for this relationship — need a human who is accountable for the outcome. It has no stake in what happens next. You do.
  </Accordion>
</AccordionGroup>

## Would you trust it unsupervised?

A quick calibration test — the skill this page builds is knowing which situation you are in *before* you hit send.

| Scenario                                                | Call                          |
| ------------------------------------------------------- | ----------------------------- |
| Draft a thank-you email                                 | **Yes**                       |
| Get the exact total of Q2 capital calls across 12 funds | **Verify first** (arithmetic) |
| Know whether Hillspire signed a new lease this month    | **No** (stale facts)          |
| Decide whether to approve a vendor's contract exception | **No** (your judgment call)   |

Same tool, four different answers.

## The one line to keep

<Tip>
  A fast, tireless drafting and thinking partner — **not an oracle**. You stay accountable.
</Tip>

## Where to go next

<Card title="Prompting & getting good results" icon="wand-magic-sparkles" href="/literacy/prompting">
  Goes deeper on getting good output, spotting a bad answer, and verifying before you trust. Today, pick one draft, summary, or brainstorm task off your plate and hand it a first pass — and name one thing you would never ask it to just tell you cold, like a number, a current fact, or a judgment call.
</Card>