The Future of Intelligence Is Personal

Artificial intelligence feels like a radically new technology, but the architecture through which we access it is surprisingly old.

Most AI exists inside enormous centralised systems. The models run in remote data centres, the providers own the infrastructure, and users purchase access through subscriptions or tokens.

In this respect, we may be living through the mainframe era of artificial intelligence.

Early computers were too large, expensive and complex for individuals to own. Institutions operated the machines, and users accessed them through terminals.

This was not a mistake. Centralisation was the only architecture that made economic sense at the time.

The same is true of today's frontier models. They require enormous quantities of compute, energy, engineering and capital. Centralised infrastructure allows millions of people to access capabilities they could never operate independently.

But technologies do not necessarily remain where they begin.

Mainframes gave way to personal computers, not because centralised computing became useless, but because computing became sufficiently small, affordable and valuable to place directly in the hands of individuals.

Artificial intelligence may follow the same path.

Intelligence is rented

Today, intelligence is primarily rented.

Every time a company's product sends a request to a model provider, it pays for the intelligence required to process that request. The company may own its product, customer relationships and data, but it does not own the underlying intelligence.

The individual user is even further removed. Their conversations, documents, memories and preferences are distributed across applications controlled by different companies.

This arrangement is convenient, but it produces a particular kind of dependence.

The model provider controls the price, availability, limits and behaviour of the intelligence. It can change the model, restrict usage, remove features or alter its terms.

The user receives intelligence, but does not possess it.

This may be temporary.

The Jevons paradox of intelligence

As models become more efficient, the cost of intelligence will fall. But falling costs will not necessarily reduce the demand for compute. The opposite may happen.

This is the logic of Jevons paradox: when a useful resource becomes cheaper, people often discover so many new uses for it that total consumption increases.

Cheaper intelligence will be used more frequently, by more people, across more tasks. Instead of occasionally asking a chatbot a question, we may operate thousands of agents that continuously read, reason, monitor and act.

The future may therefore contain both enormous centralised models and billions of smaller local ones.

One of the strange habits of the current AI industry is using the most capable available model for problems that are comparatively simple.

A frontier model may classify an email, extract a date, summarise a document or route a customer request perfectly well. But this does not mean it is the most sensible model for the task.

It is similar to using a supercomputer as a calculator.

As smaller open models improve, more routine intelligence can move closer to the user. Local models will not need to outperform frontier systems at everything. They will only need to perform particular tasks reliably, privately and efficiently.

The future is hybrid

The future is therefore unlikely to be entirely cloud or entirely local.

It will be hybrid.

Frontier models will solve unfamiliar and exceptionally difficult problems. Local models will handle frequent, predictable and private work. The system will decide which form of intelligence each problem requires.

The important shift is not that the cloud disappears.

It is that the cloud stops being the default location of all intelligence.

The model is not the centre. Context is.

The industry currently treats the model as the centre of artificial intelligence.

But for many personal applications, the model may not be the most valuable component. The model can be replicated, replaced or improved. What cannot easily be reproduced is the context accumulated around a particular person.

Your private context includes your memories, relationships, preferences, conversations, documents, habits, decisions and goals.

A generic model may know more facts than any human being. It still does not know which facts matter to you.

The most useful personal intelligence will need to understand this context continuously. It will need to remember previous decisions, recognise changing priorities and learn how the user prefers to work.

Sending all of this information to a remote provider creates problems beyond privacy. It makes personal intelligence dependent on infrastructure that the individual cannot control.

A system cannot be truly personal if someone else owns its memory, permissions and intelligence.

Applications wait. Agents pursue.

The personal computer was designed around applications.

An application is a tool. The user opens it, performs a task and closes it. Information is divided between email, calendars, browsers, documents, messaging services and hundreds of other systems.

Artificial intelligence introduces a different interface.

Instead of opening an application, the user can describe an objective. Agents can then determine which tools, information and actions are required to accomplish it.

Applications wait for instructions. Agents pursue goals.

But useful agents require persistent context. They need memory, identity, permissions and an understanding of the boundary between what they may observe and what they may change.

These requirements suggest that personal intelligence cannot remain merely a feature added to existing applications.

It may require a new kind of computer designed around intelligence from first principles.

The Personal Intelligencer

I call this category the Personal Intelligencer.

A Personal Intelligencer would not simply be a personal computer with a chatbot installed. Its organising principle would be intelligence rather than applications.

It would maintain a private memory controlled by the user. It would run local models for frequent and sensitive tasks. It would use frontier cloud models when their additional capabilities were necessary.

It would understand the user's context across their digital life without requiring that context to be permanently stored in someone else's cloud.

The system would operate through agents. The user would provide goals, constraints and permissions. The machine would remember, reason and coordinate the necessary actions.

This architecture would reverse the present relationship between the individual and the cloud.

Today, the cloud contains the intelligence and the personal device acts as a terminal.

In the future, the personal device could contain the user's memory, context and everyday intelligence. The cloud would provide additional capability when requested.

The cloud would become an extension of the personal machine rather than the owner of its intelligence.

This cannot be achieved through software alone

This transition cannot be achieved through software alone.

Persistent local intelligence requires compute, memory and energy. A machine that continuously reasons about personal context must operate within strict limits on power, heat, size and cost.

Current hardware was not designed for this purpose.

Much of modern computing architecture was created for applications that wait for human input. Personal intelligence may require models and agents that remain active throughout the day.

This creates a different engineering problem. The objective is no longer simply maximum computational performance. It is useful intelligence per watt, per dollar and per unit of memory.

Specialised inference chips may become increasingly important because they can be designed around the actual structure of AI workloads rather than the general requirements of traditional computing.

Optical computing and the PI Brain

Beyond electronic inference chips lies the possibility of optical computing.

Moving and processing information with light could eventually provide a significant improvement in performance per watt for certain AI operations. This matters because energy and heat may become the fundamental limits on how much intelligence can fit inside a personal machine.

Optical computing would not create personalisation by itself. Personalisation comes from memory, context, permissions and continuous adaptation.

But a radically more efficient computing architecture could make deeper local personalisation physically possible.

This future architecture could be thought of as the PI Brain: the computing system that allows a Personal Intelligencer to reason continuously without behaving like a data centre inside someone's home.

The deeper consequence is ownership

The most important consequence of this shift may not be technical. It may concern ownership.

Personal computers allowed individuals to own the machines through which they created, calculated and communicated. The files on those machines became extensions of their work and identity.

Personal intelligence may become even more intimate.

A system that remembers your life, understands your relationships and acts on your behalf is not merely a tool. It becomes part of your capacity to think and act.

Whoever controls that system possesses extraordinary influence.

If personal intelligence remains entirely centralised, a small number of companies may mediate not only what people see, but how they remember, decide and pursue their goals.

Owned intelligence offers a different possibility.

The individual could control the memory. The individual could choose the models. The individual could determine the permissions. The individual could decide when the cloud is used and what information is shared.

Privacy would not be an optional setting added afterwards. It would be a property of the architecture.

The cloud will not disappear

None of this means cloud AI will disappear.

Mainframes survived the personal computer. Data centres expanded after laptops and smartphones became universal. New computing paradigms often enlarge the entire market rather than simply replacing the previous one.

The largest models will remain centralised. Some problems will always justify access to extraordinary quantities of remote compute.

But the centre of gravity could move.

The cloud may contain the world's general intelligence. The personal machine may contain the intelligence that understands you.