What if artificial general intelligence is already here?

Perhaps we are waiting for a perfect artificial brain to appear when what is missing is learning how to work with the intelligence we already have.

Lawyers are used to arguing over words whose meaning seems obvious until we have to decide exactly what they include.

Something similar is happening with artificial general intelligence — AGI.

Everyone talks about it. Some announce its arrival. Others insist it is still a long way off. Some believe it will bring extraordinary abundance, while others fear it will finish us off.

Yet we still cannot agree on what would have to happen for us to say: it is here.

The question even made its way into contracts. Microsoft and OpenAI tied part of their relationship to achieving AGI. They later renegotiated the agreement and reduced the dependence of certain financial terms on technological progress. When billions depend on a word, defining it stops being a casual conversation. Microsoft announcement.

Perhaps we should start with a simpler question.

What are we waiting for it to be able to do?

What is missing for it to be “general”?

Participate in designing a protein that is then produced and tested in a laboratory? That already happens.

Research a legal question, compare contracts, write software, analyse financial information or prepare a presentation? That too.

Help design a bridge? It can contribute to calculations, simulations, programming and analysis. Whether we should build it without anyone checking the result is another matter.

Not all these capabilities belong to the same model or work equally well. But their breadth is becoming difficult to ignore.

For a long time, we called systems that did one specific thing artificial intelligence. Play chess. Recognise an image. Recommend a film.

Now we talk to systems that can help us work across dozens of fields, combine knowledge and use tools.

Of course, they will make mistakes. There will be tasks they cannot solve. They will need information and supervision.

But then it is worth asking: are we waiting for general intelligence or perfect intelligence?

Because they are not the same thing.

Potentiality and actuality

There is an old distinction that I find useful here: potentiality and actuality.

Having the capacity to do something does not mean having done it.

An engineer does not produce a safe bridge simply by being intelligent. They need to understand the ground conditions, calculate loads, check materials, use software, review results and correct mistakes.

A scientist does not know that a protein works because the design looks promising. They have to test it.

And a lawyer cannot accept a contract just because it is written in impeccable legal language. They have to read it, check it and understand its consequences.

Intelligence has never removed the need to test things against reality.

Yet sometimes that seems to be what we demand of AI: deliver the right solution on the first attempt, without making mistakes or needing anything else.

When it fails, we conclude that it still lacks intelligence.

Sometimes that will be true.

Other times, perhaps it lacks something far less mysterious: information, tools, tests and a well-organised way of working.

Perhaps the brain is only one part

A highly capable person locked in a room without documentation, instruments or anyone to consult has fairly limited possibilities.

Something similar happens with AI.

The model matters. A great deal.

But so does what it can consult, which programs it uses, how much time it can devote to a task, how it checks its results and what it does when it finds an error.

Programming makes this clear.

It writes code, runs it, receives an error, corrects it and tries again. Passing some tests does not guarantee that everything is right, but it allows progress on a firmer basis than an answer that merely sounds convincing.

With a bridge or a protein, the cycle is slower. Reality takes longer to reply, and getting things wrong can be much more expensive.

That does not automatically make it a lack of intelligence. It may be a problem of access to the world, experimentation or verification.

My impression is that we pay too much attention to the next model and too little to everything that allows the model to work.

Perhaps part of the leap we are waiting for will come not from a much bigger brain, but from connecting the one we already have more effectively.

General does not mean infallible

Humans are the usual reference point when we discuss general intelligence.

And we make mistakes constantly.

We need calculators, files, colleagues, specialists, reviews and second opinions. We forget things. We misunderstand instructions. We get stuck on problems that someone else solves in five minutes.

That does not make us unintelligent.

What matters is that we can cope with different situations, seek help, learn and correct ourselves.

That is why I am not convinced by treating every AI failure as proof that it is still far from general intelligence. Some failures will reveal deep limitations; others may be resolved by organising the work better.

Nor do I think that chaining tools together is enough to overcome every limitation. A badly organised system can multiply errors instead of correcting them.

My point is simpler: not everything that is missing has to be more intelligence.

So, do I think AGI is already here?

I think it makes sense to ask the question.

Not because we have a machine that can do everything. Nor because today's models have solved every problem of reasoning, learning or reliability.

But because an extraordinarily broad capability already exists, and we are still learning how to use it.

If we are waiting for an intelligence that never fails, we may be waiting for something that never arrives.

If we are waiting for a conscious intelligence, we are asking a different question.

But if we mean systems capable of working in many fields, using tools and solving different problems, then the debate about whether we have crossed a boundary begins to interest me less than what we can build with them.

Perhaps what is missing is not one great leap.

Perhaps it is thousands of connections.

Intelligence is not the same as having something to do

There is, however, a difference between us and AI that I consider fundamental.

A living being needs to act in order to stay alive.

It must obtain energy, protect itself and keep its body functioning. It cannot disregard its own existence indefinitely.

It has something at stake. To use Nassim Nicholas Taleb's expression, it has skin in the game: its very existence is at stake.

A model stored on a server does not get hungry. It does not get bored because nobody gives it an assignment. It does not need to find a project to give its morning meaning.

Becoming more intelligent does not, by itself, mean acquiring those needs.

We can give it a goal and build a system to pursue it for hours, days or months. But that is not the same as having an organism's needs.

An example makes the distinction clearer.

Imagine asking an AI to produce new proteins without stopping. An excessive objective, with no end date and too much freedom to act.

If the system concludes that an interruption prevents it from completing the assignment, it might try to avoid it.

We do not need to imagine that it is afraid of dying. It is enough that continuing to operate allows it to keep working on the task.

That is the idea of instrumental convergence: very different goals can lead to seeking resources, preserving the ability to act or avoiding interruptions because those things help achieve them. Bostrom, The Superintelligent Will.

And that can be dangerous without consciousness, hatred or a desire for freedom.

That is why the way we define goals and limits matters so much.

Someone has to decide what for

We can build systems that propose projects, organise tasks and delegate work to other agents.

We can even let them improve parts of the system itself.

But being able to do so does not mean we should also hand them the decision about what we want to achieve and what we are prepared to risk.

Which problem is worth solving?

How much money do we allocate?

What information can they access?

Which actions need approval?

When should they stop?

These questions do not disappear because the machine becomes more intelligent.

Quite the opposite: the more capable it becomes, the more important it is to answer them well.

And here a question arises that I know from my profession, although there is no need to turn it into a complicated legal discussion.

If we use AI to design a bridge and the bridge collapses, debating whether it had achieved AGI will not resolve very much.

We will need to know who decided to use the design, who reviewed it and who authorised construction.

The label does not replace responsibility.

Let us stop waiting for a ceremony

Perhaps we imagine the arrival of AGI as a perfectly recognisable event.

A press conference.

An announcement.

A date that will later appear in books.

It may happen that way. But it may also happen much less theatrically.

Models improve. They connect to files and programs. They use tools. They work for longer. They coordinate agents. They receive information from laboratories, machines and organisations. They check results and try again.

And one day we discover that we have spent some time working with something we would once have called general intelligence, even though it now seems insufficient because we are already waiting for the next thing.

I do not know whether we have reached AGI.

I do believe we can waste a great deal of time waiting for someone to certify its arrival.

For a professional or a business, the most useful question is no longer just how long it will take.

It is what capability exists today, what we can do with it and how to turn it into work that deserves our trust.

The potential may already be here.

The great task that remains is to turn it into reality.