What Are AI Primitives?

As King Solomon says,

"What has been will be again, what has been done will be done again; there is nothing new under the sun." — (Ecclesiastes 1:9)

Just like USV once mapped out mobile primitives, I want to stand on their shoulders and ask: what are the foundational building blocks in this new GenAI world?

Let’s go back to what Albert, one of the USV partners, wrote:

Carriers seem to have lost their role as gatekeepers for applications as smartphone sales are rapidly ramping and “app stores” or direct downloads are the new distribution models. This is exciting as it opens up a whole new arena for startups to compete in.

It’s the same story with GenAI.
Search engines no longer control distribution.
ChatGPT, Claude, Perplexity — these are the new distribution models.
And like in the early mobile days, this opens up a whole new arena for startups.

Albert continues:

By “native” we mean opportunities that simply did not exist previously and cannot exist without the phone.

We can apply the same lens.
We’re looking for ideas that didn’t, and couldn’t, exist before GenAI.

Here’s how mobile primitives looked back then:

• Location
• Proximity
• Touch
• Audio Input
• Video Input

Now let’s see what could be GenAI primitives:

# 1. Chat Interface

Definition: The default way users interact with GenAI apps — search, act, and browse via chat.

Almost every GenAI product uses a chat UI.
ChatGPT. Perplexity. Claude.
It’s the WhatsApp effect. Chat has become a universal interface.
And now, that same flow is being used not just to talk, but to search, explore, and act.

This will only grow.
We’ll likely see the chat interface become the layer on top of every action — shopping, support, decision-making.

# 2. Generative Models

Definition: The core intelligence that powers GenAI — built on general-purpose or fine-tuned LLMs.

This is the engine.
Companies can plug into powerful APIs like OpenAI’s.
Or they can build their own models.
Either way, the ability to generate language, code, media, on demand, is the heart of GenAI.

# 3. RAG Pipelines

Definition: Retrieval-Augmented Generation pipelines fetch and inject relevant content into prompts.

Every serious GenAI app uses some kind of RAG setup.
It pulls in product docs, user data, historical records — whatever’s needed.
The pipeline prepares this content and feeds it into the model to generate responses.
It powers both generation and search.

This is a must-have.

# 4. Proprietary Data

Definition: The one true moat. Your unique data is what makes your app different from others.

This is where real differentiation lies.
Anyone can use the same model.
Anyone can build a RAG.
But only you have your data — customer interactions, domain knowledge, internal content.

If you don’t have that data yet, your first job is to start building the pipeline to get it.

# 5. Agentic Workflows

Definition: Autonomous chains of tasks that act, reflect, and adapt to complete goals.

Not mainstream yet. But it’s coming.
Tools like AutoGPT and Devin are early signs.
Imagine AI systems that don’t just answer, but take initiative, plan steps, and execute.
This could become the next primitive layer on top of current models.

# The Power of Convergence

Each primitive — chat, generation, RAG, data, and maybe agents — on its own is useful.
But when they start working together? That’s when it gets interesting.

Albert said:

Each of these unique capabilities, taken individually, is not novel... But the convergence of all of these features on a single device... will allow new behaviors and applications to emerge...

Same with GenAI.
Each primitive is powerful on its own.
But their convergence is where the magic happens.

And it’s already happening.

We’re seeing new social behaviors, new forms of expression, new ways of working.

The potential emergence of new behaviors is likely to be as important — if not more so — than these technical capabilities themselves.

That’s worth paying attention to.

# So What Can We Expect to See?

Let’s look at a few areas where these primitives are already shaping new products.

# Entertainment

This is always a good signal of mass adoption.
Why did Google Glass fail? No entertainment value.
But GenAI already has it — video, music, image generation.
It’s fun. It’s engaging. It’s everywhere.

One of my favorites: Feynman lectures as rap songs.
This is education plus entertainment plus GenAI.
That kind of convergence is where real value shows up.

# Shopping

If users are already chatting inside WhatsApp or ChatGPT — why not let them shop there?
It’s easier for ChatGPT to add shopping than for a shopping app to build distribution.
We’ll see shopping apps adopt chat.
And chat apps add e-commerce.
That interplay is coming.

# Education

Personalized, engaging, and even fun.
GenAI is turning textbooks into songs, lessons into visual stories.
It lowers the barrier for curiosity and keeps the learner engaged.

# Healthcare

I’ve heard stories of GPT catching conditions doctors missed.
I use ChatGPT to go through my annual health reports.
I can ask any number of questions — without worrying about sounding dumb or taking up a doctor’s time.

# Shovels in the Gold Rush

Building GenAI primitives — models, RAG, data — is complex and expensive.

So what happens?
Just like in the gold rush, people are selling shovels.

And these shovels are:

• Creating GenAI tutorials
• Coaching on prompt engineering
• Offering certifications
• Writing books
• Curating newsletters
• Aggregating jobs

They’re making money because they sit on top of the primitives.

# Devs as a Market

For a long time, people said “developers aren’t a real market.”
But tools like Cursor and Windsurf have flipped that.

We’ll see more products aimed directly at developers.
And every stage of app development — design, test, deploy, monitor — will be reimagined through GenAI.

# In Closing

Fred Wilson said it best:

If you want to figure out what the native AI applications are, start by laying out the new primitives and going from there.

But keep this in mind too:

When mobile first arrived, no one imagined ordering a cab or food on it.

And as Ken Stanley put it:

The next step in greatness often doesn’t resemble the previous step, especially for ambitious, unpredictable goals.

That’s worth remembering as you are building products.
Keep an eye on how the markets are shifting.
Talk to customers regularly.
Find relevant communities and contribute consistently.
Use a consume-produce-engage learning model to learn by doing.

And maybe Solomon was right.
There’s nothing new under the sun.
But there are new combinations. New behaviors.
New ways of thinking about old problems.

That’s where the opportunity lies.
And the best way to find it? Build something.

Got comments? Discuss them on LinkedIn, Twitter, Bluesky or Mastodon.