Our vision

The Plural of Intelligence

Noeri, noun: plural of noerus, “intelligent, gifted with understanding.”

Yejin Choi and Carlos Guestrin · November 2026

7 min read

Imagine a future in which AI flourishes as a healthy ecosystem of many intelligences, in all sizes, strengths and quirks.

Rather than one monolithic source of truth, we can nurture many small, specialized intelligences, each tuned to a place, a people or a problem. Some are brilliant at one thing and hopeless at others. Together they span far more of the world than any single intelligence could, and they keep evolving as the world changes.

For four billion years, life has thrived on plurality. No single species rules nature. In a forest, fungi feed the trees, bees pollinate their flowers and bacteria turn fallen leaves into soil. Humanity is the same. No one person is wiser than everyone else: a farmer reads the weather in ways a physicist never learned. Our differences are the source of our collective intelligence, and much of what makes the world beautiful.

We believe AI can carry this plurality forward.

Contrast that future with one superintelligence, made and owned by one company, with no alternative beside it. Its blind spots become the world’s blind spots, and its mistakes repeat everywhere at once. What its owners value, the world gradually comes to value too.

Both futures are within reach. The choices made in the next few years will decide which one arrives.

In the plural future, the power of intelligence is distributed in four ways: the craft of creating it passes to many makers, discovery reaches the roads less traveled, ownership rests with those who make it, and intelligence comes in right-sized systems that are both more economical and safer. This very future depends on intelligence that can evolve.

Create: The Power of AI Craftsmanship, in Every Hand

A plural future needs many makers. Creating AI is a craft, and craft decides the outcome.

Give a Michelin-starred chef and a scientist who cooks only simple dishes the very same ingredients, and two very different meals will reach the table.

Modern AI works the same way. Two teams with the same data and the same compute can arrive at entirely different models, depending on how they curate the data, choose the training recipe and make the many small judgment calls along the way.

Today that craft, from preparing the data and training the model to choosing the algorithms and tuning the countless knobs in between, is held by a relatively small circle of experts concentrated at a few frontier labs with enormous resources.

In the future we imagine, this craftsmanship is within reach of every organization with something of its own to build. Scientists raise models around their own data and hypotheses, including the ones no industry lab has chosen to pursue. Companies turn decades of proprietary know-how into models that embody their expertise. Governments build AI that reflects their own culture, languages, laws and public values. Problems too specific or too far from the commercial mainstream finally get intelligence built for them.

Discover: The Roads Less Traveled

The hardest problems are solved on roads few have traveled.

Today’s AI is trained on what has already been written, so it gravitates toward the well-worn path: the most common answer, the most cited idea. Discovery rarely lives there. Frontier labs have begun to push AI off that path through search at enormous scale, producing remarkable discoveries. Few institutions can match that scale, so most scientists watch that search from the sidelines, with few realistic paths to take part or to aim it at their own problems.

Breakthroughs depend on someone willing to stay with an unpopular question long after others have stopped asking it. Katalin Karikó spent years on messenger RNA while her grant applications were rejected and colleagues moved on; that work later made the COVID-19 vaccines possible.

In the future we imagine, every scientist takes part in discovery. AI stays with the hard problem and keeps learning from it. It remembers what was tried and why, and grows more capable at your specific problem with every experiment. Small, specialized models make this economically viable: thousands of hard problems can be pursued at once, with many specialized intelligences working as a team to propose, critique, test and refine, including along the roads the field has left behind.

You bring intuition, context and judgment about what matters. The AI brings tireless attention, breadth and the patience to keep going. It shows its reasoning and its uncertainty so you can steer it, and the discoveries that emerge are ones you made together.

Own: Intelligence Belongs to Those Who Make It

Every creation, every dataset and every AI shaped from them belongs to those who made them.

Today, using the most capable AI usually means sending your data to someone else’s servers and renting an intelligence you cannot inspect, move or keep. And when the provider changes its model, its prices or its rules, everyone downstream changes with it.

Few companies would build their core operations on a single consultant who also advises every competitor, can change its terms at will, and can never be brought in-house. A single, centralized AI offers exactly that arrangement.

In the future we imagine, intelligence is owned the way other critical assets are owned. A company’s model stays behind its own walls, as private as its trade secrets. A government runs AI on its own infrastructure, accountable to its own citizens.

Owners decide where their AI runs, what it learns, how it behaves and who may learn from it, and when others do, the owners share in the value.

At Gettysburg, Lincoln spoke of government of the people, by the people, for the people. A plural future extends that promise to intelligence. Intelligence by the people is created by many makers, each bringing what they know best. Intelligence of the people is built from their own data and know-how, and answers to them as its owners.

Intelligence for the people works on their behalf: it serves their interests, and the value it creates returns to them and to those they serve.

Economy and Safety: Just Right for the Job

The plural of intelligence is also the economical path.

The frontier race today assumes that every gain must come from a larger model, with each new generation demanding data centers bigger than any yet built and power on the scale of a city. They are discovering much of what AI can do for humanity.

The future is better still when the frontier coexists with many other options, the way a city is richer for having a full range of restaurants: upscale ones for special occasions, and neighborhood places for every day.

For many problems of real economic value, the right-sized model can be much smaller than we assume.

A model sized to the problem, trained with the right craft, can often match a giant one at a small fraction of the cost and energy.

It is also the safer path. A system built for one purpose is easier to oversee than one built for every purpose. Its intended behavior can be specified, its failures can be tested against that specification, and its reach is bounded by the job it was made for.

In this future, AI is a family of tools, each created for the job at hand: one forecasts hurricane seasons, one designs experiments for a battery lab, one keeps a family’s recipes. No single system is unsettlingly capable at everything, and no single failure or misuse can spread everywhere at once. Laws and technical guardrails govern the craft, as they do medicine and engineering. Power stays distributed, among people and among intelligences, as it does in every healthy ecosystem.

Noeri: The Plural of Intelligence as Living Systems

At Noeri, we are building intelligence that learns proactively and continually. Most AI today is trained once and then frozen, while the world it serves keeps moving.

We close that gap with living systems that grow out of our own research. Our work on test-time training showed that learning on the job after deployment can unlock new capabilities and lead to new discoveries. Our work on self-distillation found that a model can become its own teacher, learning from each mistake and solving hard problems in far fewer attempts. And our work on agentic orchestration demonstrated that a small model conducting a team of other models and tools can outperform frontier models at a fraction of the cost.

A living system learns on the job, the way human experts do.

The ability to evolve is what makes the plural future possible. It lets every maker shape intelligence through the work itself. It lets discovery stay with a hard problem and grow sharper with every experiment. It turns what an organization teaches its AI into knowledge the organization owns. It ends lock-in, because a system that can improve itself can also remake itself around any model: a newer version, a cheaper one, or another provider’s entirely. And it lets a small model grow into its job, gaining capability through learning at a fraction of the cost of ever-larger scale.

Our first proof-of-concept partnerships span enterprise and scientific domains, where our systems learn from each organization’s own context and grow more capable and less costly over time.

An Ecosystem of Minds

Return, for a moment, to the two futures. In one, intelligence speaks with a single voice. It is owned by a few and consulted by everyone, and over time the world learns to think the way it thinks.

In the other, intelligence grows the way life and culture always have: in countless forms, rooted in particular places and people, owned by those whose knowledge shaped it, and devoted to the questions they care about most.

The second future, the pluralistic future, is richer, more resilient, safer and more humane.

It is also harder, and it is worth the effort. The forest outlasts any single tree, and a civilization of many intelligences outlasts any one of them.

The future of intelligence, like the history of life and of humanity, is plural.

Explore our research