The Value Of AI

Jul 09, 2025

I think people miss why the LLM revolution is such a big deal.

They think it’s just summaries and human-like conversations.

I assert it's mastery of ambiguity.

Before LLMs, the way to use a computer was through structured input. Fill out a field, press a button, drag a thing.

Machines speak definite, precise language so humans had to too.

Meaning, only those who warped their brains to think in definite, precise ways were good at computers (and now they’re losing market share).

But anyone who's filled out a form with excessive "Required" fields knows that precision is expensive.

Think about our language.

We say something, and if it's unclear we clarify by adding context, resolving ambiguity. If we mess up we say "Oops, not what I meant."

We are naturally iterative communicators. And it's fine 98% of the time.

Only 2% of the time do we bust out the precision: legalese, code, etc.

The ambiguity is a speed feature, not a bug.

And now LLMs let the machine do the same.

We can talk to computers with "good enough".

We can interact with tech like we're painting: broad strokes, then refine when necessary.

Consider how much Google is not this. I have to:

Translate my query into search engine-ese ("founder father age american revolution")
Identify results that look promising among a list ranked by someone else
Hope the results - written in a way that made sense to the author, not me - are useful for me
If they’re not, debug why (was my search engine-ese not good enough? does what I'm looking for just not exist?)
Repeat from the top

I'm not painting. I'm carving a scalpel through a realm I don't understand.

LLMs skip all that.

I query in my language. I get results in my language.

When I don't understand, I drill in. Get 40 different explanations until I do.

Go slowly. Explain it like I’m five. Use examples from my own life.

No wonder Google is scared.

The common objection is that LLMs hallucinate, make errors. True.

But we have tooling for this, because humans make errors them all the time.

Guardrails like:

And LLMs will put pressure to make these guardrails even better (e.g. a council of LLMs for verifying things).

I'm all for it.

Forget being a computer user; I want to be a computer painter.

If we know why LLMs are powerful, we can identify areas where they can be disruptive:

Where speed of information input is important: a census-taker, a soldier, a patient registrar in the emergency room
Search-and-refine flows: searching the web, doing research, analyzing data, shopping on Amazon
Where input is unstructured: email, text messages, book notes, videos & voice recordings, blog posts, PDFs, the results of some APIs I've encountered 💀
Where imprecision + undo feel better than precision: processing your email, processing your TODO inbox, blocking your calendar, adding items to a shopping cart, navigating a website
Where happy accidents enrich the result: image/video generation, interior design, writing, restaurant- or apartment-hunting
Information expansion workflows: adversarial empathy in negotiations, surfacing disconfirming evidence to theories, brainstorming

And that's just what I could think of. We don't yet know how deep the rabbit hole goes.

Until 4 years ago, interacting with the computer required precision so all our computer interfaces are oriented around specificity.

What happens when the default computer interaction is ambiguous: fast and iterative by default, slow and precise only when it matters?

What new opportunities emerge?

I'm excited to find out.

Becoming More