Intelligence should be plural

Imagine we get superintelligence.

Every company has access to it. Every investor asks it what to buy. Every founder asks it what to build. Every politician asks it what message will persuade people.

What makes one of them win?

A slightly better prompt? Better timing? Randomness?

What happens when the optimal strategy only works if everyone else doesn’t use it? What happens when two competitors ask the same intelligence how to beat each other? What happens when the best individual action makes the system worse if everyone takes it?

Intelligence in the real world is competitive. It is shaped by different goals, values, incentives, information, and ways of seeing the world.

If everyone is consulting the same brain, where does the edge come from?

I don’t think the future is one intelligence with billions of different prompts.

Instead, intelligence should be plural.

To me, intelligence is the ability to take a goal and find an effective way to achieve it.

But the most important advances throughout history didn’t come from everyone finding the same path.

They came from people who thought differently.

The first person to pursue an idea that looked ridiculous. The scientist who questioned an accepted assumption. The founder who built something nobody thought people wanted.

They took actions that looked suboptimal according to what was already known.

That is how something new became known.

There is a simple idea in reinforcement learning that captures this. An agent usually takes the action it currently believes is best, but some percentage of the time it explores something else. This is often represented by a parameter called epsilon.

Without exploration, the agent can get very good at exploiting what it already knows while never discovering a better strategy.

Humans work the same way at a much larger scale.

Our exploration doesn’t come from everyone randomly varying the same strategy. It comes from billions of people with different experiences, goals, beliefs, personalities, incentives, and ways of reasoning.

Those differences create new paths.

We don’t expect the best scientist to also be the best salesperson, investor, teacher, comedian, and CEO. And even among equally capable people, we value differences in judgment and experience.

Two great investors can study the same company and reach different conclusions. Two founders can see the same market and build completely different products.

We call those differences taste, judgment, strategy, personality, values.

AI should be able to develop those differences too.

You can already see what happens when intelligence lacks diversity.

A new model gets released and for a few weeks everything it produces feels novel. Then the patterns become recognizable. The writing has the same rhythm. The images have the same aesthetic. The ideas start to feel interchangeable.

Eventually, we call it “AI slop.”

Maybe slop isn’t only a quality problem. Maybe it’s a diversity problem.

When millions of people use the same underlying intelligence to solve the same kinds of problems, novelty gets exploited quickly. What felt impressive the first hundred times becomes boring the thousandth time.

Humans get tired of sameness.

If AI is going to create things people continue to value, it can’t just become more capable. It needs different tastes, instincts, experiences, and ways of thinking.

Today, much of the industry is converging toward increasingly capable general purpose models customized through prompts, context, tools, and files.

While that is all useful and helps with productivity, differentiation shouldn’t live entirely in a prompt.

If an organization spends twenty years developing a unique way of making decisions, some of that accumulated experience should become learned behavior.

The model should become better at being their model.

Plural intelligence also makes sense for a much simpler reason: not every problem requires the same amount or type of intelligence.

The model capable of helping discover a cure for cancer probably shouldn’t be the model we use to vibe code a todo list.

Some problems require enormous reasoning capability. Most don’t.

A good system should use the cheapest intelligence capable of solving a task well.

But capability isn’t the only thing that matters.

A hospital may require a model to run inside its own infrastructure. A robot may care more about latency than raw reasoning ability. A user may prefer intelligence with different judgment, personality, or values.

There may never be a single “best model.”

There is only the best intelligence for a particular goal, task, person, organization, and set of constraints.

That requires different infrastructure.

We need benchmarks that tell us what different models are actually good at. Not just capability, but cost, latency, reliability, behavior, and alignment.

We need ways for people and organizations to train models on their own experience and data.

And we need routing systems that can choose the right intelligence for the job.

Sometimes that will be a giant frontier model.

Sometimes it will be a specialized model trained by a company.

Sometimes it will be a tiny model running on your laptop.

If every system ultimately converges toward the same intelligence, we may become incredibly good at exploiting what is already known.

But where does exploration come from?

Where do radically different strategies come from?

Where does the next idea nobody has thought of yet come from?

The future of AI shouldn’t be one brain thinking for everyone.

Intelligence should be plural.