We may be asking the wrong question about AGI.
The usual question is when will somebody build it?
OpenAI? Google? Anthropic? Microsoft? Perhaps a company that doesn't yet exist.
We imagine an enormous machine in an enormous data centre, consuming heroic quantities of electricity while becoming steadily cleverer than the people who built it. Eventually we all connect to it and ask for help.
It's a very twentieth-century vision of artificial intelligence.
HAL 9000 in 2001. Colossus. WOPR in WarGames. Skynet.
There is always one great machine.
But look at what is happening to the computer in front of you.
Apple's current M5 generation already has neural acceleration built into every GPU core and Apple says the M5 MacBook Pro can deliver up to 3.5 times the AI performance of the previous generation.
That doesn't make a MacBook an AGI. Not remotely.
But carry the trajectory forward.
M6. M7. M8. Better models. More memory. More efficient inference. Specialist software running locally. Long-term personal context.
At some point we may find ourselves carrying computers capable of doing rather more than giving us polished answers.
And then the interesting question isn't:
Who built AGI?
It's:
Who owns mine?
That question has started to feel less abstract to me because I'm already building a small home AI machine to play with some of this.
I use the word building rather grandly. At times I feel like the artilleryman in The War of the Worlds, digging away at his underground future while explaining how civilisation is going to be rebuilt.
There's a fair chance I'm simply standing beside some hardware and a rather modest hole.
But experimenting with local AI has changed how I think about personal intelligence.
Not a chatbot running locally.
Not simply privacy-preserving inference.
Something closer to an intelligence that increasingly knows me, works for me and reaches outward when it needs help.
That could become a much bigger argument than AGI itself.
The intelligence in the cathedral
Most of the AGI future currently being imagined is centralised.
The really clever machine lives in somebody else's data centre.
Your phone, laptop, glasses or whatever comes next provides access to it.
Ask it to arrange a holiday and it understands your preferences, checks your diary, searches flights and compares hotels.
Ask it to move house and it deals with estate agents, solicitors, mortgage providers, utilities, removals companies and the council.
You might say:
Sort out my move.
A few minutes later:
I've completed the change-of-address notifications, transferred the utilities and shortlisted three removals companies. Your mortgage adviser needs approval to proceed with option two.
Anyone who has actually moved house will recognise this as dangerously close to magic.
And there are powerful reasons for building AI this way.
The largest models require enormous resources. Central platforms can maintain security, update software continuously, combine specialist systems and summon huge amounts of computing power when required.
There is another advantage.
Convenience.
Someone else deals with it.
We've built much of the modern digital economy on precisely that bargain.
Google remembers what we search for.
Amazon remembers what we buy.
Spotify remembers what we listen to.
Social platforms remember who we know.
An intelligent platform could remember something much more useful.
Us.
Our habits, relationships, finances, preferences, ambitions, mistakes and history.
Because the better it understands us, the better it can represent us.
That is both the attraction and the problem.
We may be building the most intimate technology ever created while treating it as though it were simply a better search box.
Turn the architecture around
Now imagine a slightly different world.
The really important intelligence doesn't primarily live in the cloud.
It lives with you.
Not necessarily every model or every calculation. There will always be tasks worth sending to enormous specialist systems.
But your identity, memory, permissions, preferences and personal context remain under your control.
Your AI knows you.
Other systems don't.
When it requires something from them, it asks.
Suppose you're moving house.
Your AI talks to the council.
The council needs your identity, old address, new address and moving date.
Fine.
The mortgage provider needs considerably more financial information.
Fine.
The removals company needs your addresses, dates and some idea of how much furniture you own.
Also fine.
But why should any one of them need the whole picture?
The removals company doesn't need your bank statements.
The mortgage provider doesn't need your Council Tax correspondence.
The council doesn't need to know what you're paying the estate agent.
Your personal AI could become something we've never really had before.
An intelligent boundary around the individual.
It knows enough about you to act effectively while deciding what anybody else actually needs to know.
Today's apps usually work the other way around.
We surrender data into institutional silos and each organisation tries to reconstruct enough of us to provide a personalised service.
Personal intelligence reverses the relationship.
Instead of every organisation maintaining its own partial Martin, Martin maintains Martin.
Everyone else asks for what they need.
That changes who sits at the centre of the architecture.
It doesn't need to know everything
Science fiction sometimes misleads us here.
We tend to assume AGI must know almost everything and be able to do almost anything.
Why?
I can't.
Neither can you.
Most capable humans are spectacularly dependent upon other humans.
Neither of us knows how to manufacture a modern processor, run the National Grid or perform neurosurgery.
Yet collectively we manage to build computers, illuminate cities and occasionally repair brains.
Intelligence isn't simply possession of knowledge.
An important part of intelligence is knowing:
what you don't know, where to find it and who can help.
A useful personal AI therefore doesn't need the entire world inside it.
It needs enough capability to understand you, understand the problem and obtain reliable specialist help.
The plumbing for that sort of architecture is already appearing.
Model Context Protocol allows AI applications to connect to external tools, data and workflows. Agent-to-agent protocols are emerging to allow independently built AI systems to discover capabilities, exchange information and coordinate work.
Neither gets us anywhere near the future I'm describing.
But they suggest a different way of thinking about intelligence.
Not one omniscient machine.
A network of capabilities.
Your AI doesn't calculate everything itself.
It asks trusted systems.
It doesn't contain the railway timetable.
It queries the railway.
It doesn't become a doctor.
It takes relevant information to an appropriately governed medical system.
It doesn't become everything.
It works out which things you need.
That matters because current language-model agents remain fallible. The longer the reasoning chain, the more dangerous it becomes to assume that a fluent model can simply keep thinking until the answer becomes reliable. Work on transformer limitations points towards a more architectural answer: serious AI systems need tools, verification, data and controls around the model, not merely a larger model asked to think harder.
Perhaps useful general intelligence doesn't emerge from one increasingly gigantic mind.
Perhaps it looks more like a competent personal mind surrounded by specialists.
Rather like the rest of us.
Back to the mainframe
There is a historical echo here that feels quite personal to me.
I remember sending programs on paper tape to a central computer and waiting for them to be run.
That was computing.
The intelligence, such as it was, lived elsewhere.
You prepared your instructions, sent them off and waited for the machine at the centre to do its work.
Then came the personal computer.
The revolutionary thing wasn't simply that computers became smaller.
The individual gained computational sovereignty.
You could own the machine.
Run the software yourself.
Keep your files.
Experiment without asking permission from whoever controlled the central computer.
That shift created entire industries.
Then, gradually, we moved much of our digital lives back into centralised systems.
Documents went into clouds.
Email went into clouds.
Photographs went into clouds.
Music went into clouds.
Our social relationships went into clouds.
There were good reasons for nearly all of it.
The cloud is convenient.
But AI makes centralisation rather more significant than it was for spreadsheets and photographs.
A spreadsheet doesn't develop an increasingly sophisticated model of what you want.
An intelligent agent does.
Perhaps personal AI becomes another swing of the pendulum.
Not all the way back.
I doubt we're going to unplug ourselves from the cloud and retreat into technologically fortified laptops.
The more plausible model is hybrid.
Your personal intelligence stays with you.
Routine work happens locally.
Specialist work goes to specialist systems.
Enormous problems rent enormous compute.
Institutional transactions go to institutional agents.
The cloud becomes somewhere your intelligence goes for help, rather than somewhere your intelligence lives.
That may be the difference between owning an agent and renting access to one.
Of course, the centralised model might win
There is a danger of making decentralisation sound inherently virtuous.
It isn't.
Billions of autonomous personal agents would create some fairly spectacular problems.
How do I know your agent really represents you?
How does mine know whether yours is trustworthy?
What happens when someone's personal AI is hacked?
Who maintains the security?
What prevents criminals deploying apparently legitimate agents?
And how many people genuinely want responsibility for their own artificial intelligence?
The corporate model has a very powerful answer:
Don't worry about any of that. We'll do it for you.
That argument has won before.
Centralisation can make systems safer, simpler and more consistent. Large providers can spend billions on infrastructure and maintain specialist capabilities no individual could afford.
Millions of people may reasonably decide they don't care where their AI runs as long as it works.
The distributed future is therefore not inevitable.
It may actually be the harder future.
Centralisation has convenience, money and economies of scale on its side.
Personal intelligence has something rather different.
Agency.
The artilleryman's tunnel
Which brings me back to my little home machine.
I'm not building AGI in the house.
I'm tinkering.
Testing what can stay local.
Seeing what can talk to what.
Trying to understand how much intelligence really needs to be centralised and how much can sit closer to the person it serves.
In The War of the Worlds, the artilleryman has a magnificent plan.
Humanity will live underground. Tunnels will be built beneath London. Civilisation will survive while the Martians dominate the surface.
Then the narrator notices how little digging has actually been done.
That's the bit I identify with.
There is a danger in becoming so fascinated by the architecture of the future that you fail to notice you're standing beside a surprisingly small hole.
But the artilleryman wasn't wrong about everything.
He understood that adapting to a new technological order might require more than accepting whatever the dominant power happened to provide.
It might require building an alternative.
That is what makes personal AI interesting to me.
Not because centralised AI is bad and local AI good.
Not because every house will soon contain an artificial superintelligence.
But because architecture becomes power surprisingly quickly.
Imagine two equally capable systems.
One knows everything about me because I have an account with the company operating it.
The other knows everything about me because it belongs to me.
On a benchmark they might look identical.
Socially, they are nothing alike.
One places the platform at the centre and allows me access to intelligence.
The other places me at the centre and allows my intelligence access to platforms.
That reverses the relationship.
Perhaps we've seen something like this before.
AOL offered a carefully organised version of the online world. You entered its environment and accessed what had been arranged for you.
The Web offered something messier.
Protocols.
Anybody could publish.
Anybody could connect.
Anybody could build.
The Web wasn't necessarily tidier, safer or easier.
It was harder to own.
And that mattered.
The next great technology argument may therefore not be about which company creates the cleverest artificial mind.
It may be about whether intelligence develops more like AOL or the Web.
A handful of magnificent central machines to which we all connect.
Or billions of personal intelligences capable of connecting to one another and to everything else.
I don't know which wins.
We'll probably muddle into some uncomfortable mixture of both.
But there is one consequence of the distributed version that I hadn't appreciated when I started digging my own little artilleryman's tunnel.
If my AI can represent me...
and your AI can represent you...
and they can securely talk to one another...
then they can do more than exchange information.
They can discover when we want the same thing.
And that takes the argument somewhere much stranger.
Because the next question isn't really about artificial intelligence at all.
It's about what happens when the cost of organising human beings begins to approach zero.
Comments
Post a Comment