New AI every day, and questions unchanged for 400 years

ChatGPT was only released at the end of 2022.

Looking back at that date sometimes, it still feels a little incredible.

Less than four years.

At first it was a novel chat product: type a sentence, and it gives you a paragraph back.

Then came multimodality, voice, web search, AI coding, agents…

Now many people have AI read files, do research, write code, call tools, even have several agents collaborate in the same workflow.

Especially at the application layer, the pace of change is very obvious.

New products, new features and new workflows appear almost every day.

A method worth learning specially a few months ago may soon be absorbed into new product capabilities; habits you’ve just built may need readjusting before long.

But following the thread of Leviathan further lately, I’ve become interested in something else:

the different things in the world never seemed to change at the same speed to begin with.

A society always has many speeds

Stewart Brand has a framework I love, called pace layers.

He sees society roughly as several layers that change at different speeds:

Fashion / Art → Commerce → Infrastructure → Governance → Culture → Nature

The further forward, the faster things usually change.

What’s fashionable, what new products look like, what the market is chasing can change quickly.

Infrastructure is much slower.

Cities, power grids and railways can’t all be torn down and rebuilt next year because a new trend appeared this year.

Law and institutions usually have their own rhythm too.

Culture is slower still.

How a society understands family, success, fairness, work and responsibility may look very different decades later, but while you’re living in it, it’s often hard to point to the day it changed.

Brand has a line:

Fast learns, slow remembers.

The fast parts keep experimenting and exploring new possibilities.

The slow parts preserve experience, stabilize expectations, and also set some boundaries for change.

A society needs both speeds to exist at once.

In AI, it can’t simply be lined up as “who’s fast and who’s slow”

That’s also something I later felt needs care.

You can’t simply say:

models are fast, companies are slow, law is slowest.

Reality isn’t that tidy.

What makes people feel speed most strongly right now may first be AI’s application layer:

how products are designed, how agents are combined, what coding agents can do, how new forms of interaction appear.

These can move very fast.

The underlying models keep improving too, of course, but truly big leaps in capability don’t “happen once a day.”

More often, once an underlying capability comes out, lots of products and developers quickly combine it into different applications, so it feels like there’s something new every day.

Companies can’t be lumped into one speed either.

An AI startup of a dozen people might discover a new way of working today and change the whole product next week.

A bank, a hospital or a large enterprise, even when the technology itself is mature, still has to deal with data permissions, responsibility, audits, compliance, procurement, staff training and existing systems.

It will of course be much slower.

But that doesn’t necessarily mean it’s “behind.”

Different systems carry different risks to begin with.

What’s really interesting is that these different speeds often collide

Technically it can already be done.

The organization doesn’t yet know who should be responsible.

People are already using the product.

Regulators are still working out how to define it.

AI can let one person generate huge amounts of content in a day.

But people’s attention, judgment, and standards for “what’s worth doing” haven’t expanded along with it.

Coding agents can already change a lot of code on their own.

But companies still need to answer:

• who approves?
• who’s responsible when something goes wrong?
• which data can it touch?
• which actions must leave a record?

These questions often get summed up simply as:

“regulation can’t keep up with technology.”

“Companies are too conservative.”

“People aren’t ready yet.”

But from the pace-layers perspective, there’s also a plainer way to understand it:

different things just have different inertia.

And slow isn’t necessarily bad.

If the financial system redefined its rules every two weeks like a consumer app, probably nobody would dare put money in it.

If law changed every day according to the latest technology trends, people would hardly be able to form stable expectations.

Culture and social norms changing slowly is sometimes also preserving some kind of continuity.

The question isn’t that fast is more advanced and slow is more backward.

It’s:

when should the fast things keep experimenting, and when do the slow things need to catch them?

Yet our attention is naturally drawn to the fast things

The French historian Fernand Braudel proposed a very interesting concept:

the longue durée, the long term.

He argued that history can’t only look at the events that are easiest to record.

A war breaking out.

An election.

A dynasty falling.

A company going bankrupt.

These have clear points in time, and most easily become stories.

But underneath, there are structures that move much more slowly:

geography, economic relations, institutions, cultural habits, ways of organizing society.

They may not fundamentally change for decades or even centuries, yet they keep shaping why the surface events happen the way they do.

Today’s information environment is a bit like that too.

What we see every day is:

• another AI product came out.
• another capability was added.
• a new agent workflow appeared.
• some company raised money.
• some job started disappearing or appearing.

Each of these has a clear “what happened today.”

But some other things have no release date.

They’re just always there.

Tools change fast, but many human needs don’t change that fast

Over the centuries we went from horse-drawn carriages to trains, cars and then planes.

How far you can get in a day is completely different now.

But when people set off, much of what they want to do isn’t all that unfamiliar:

• going to see a place they’ve never seen.
• looking for new opportunities.
• or just wanting to see someone they miss a little sooner.

Communication is the same.

Letters, telegrams, phone calls, text messages, instant messaging, video calls.

Sending a message went from months to seconds.

But writing someone a letter and opening Telegram today sometimes still serve very old needs:

• I want to know how you’re doing.
• I want you to know I’m thinking of you.
• Something happened, and I want you to be the first to know.

AI is like this too.

It can handle more and more things for us.

It can research, plan, write code, organize information.

In the future it may well finish more of the work we now have to do ourselves.

But when you’ve made a pot of soup and realize at the end it’s completely ruined, you may still want to find someone to say:

this soup is truly beyond saving today.

When you’re heartbroken, AI can analyze attachment styles, explain the psychological mechanisms, and give a very complete set of advice.

But sometimes what a person really wants is still just another person sitting beside them, handing over a tissue.

Tools change how we meet our needs, but don’t necessarily change the needs themselves at the same speed.

So many of the “fast” things we notice most happen in tools, media and ways of doing things.

While purposes, relationships and needs often hide in slower places.

That also explains why I keep following the thread of Leviathan

I’ve talked about Hobbes several times now.

Leviathan was published in 1651.

What he cared about then was:

how can a group of people, each with their own interests, holding different information and unable to fully trust each other, live together stably?

We took this question all the way through states, companies, markets and crypto, and finally even into AI agents.

In a multi-agent system, you still run into:

• who has the power to decide?
• what counts as shared fact?
• who can execute?
• what happens when opinions conflict?
• who’s responsible when something goes wrong?

What interests me here is no longer just “Hobbes can somehow still explain AI.”

It’s more that:

technology has changed many times, but some coordination problems haven’t disappeared at the same speed.

375 years ago there were no AI agents.

But questions of power, trust, safety, cooperation and responsibility already existed.

Today’s technology keeps creating new environments for these old problems.

The environment is completely different.

Part of the problem is still there.

So “chasing the new” and “reading the old” may be looking at different things to begin with

Following AI’s new products, new capabilities and new workflows matters, of course.

Some opportunities only exist right now.

In a few years, the whole technology landscape may be completely different.

But old books do something else.

They don’t necessarily tell us what the next generation of AI will look like.

What they often preserve are the questions that move more slowly:

• why people need community.
• how power gains legitimacy.
• how to cooperate with people you can’t fully trust.
• what fairness means.
• how work relates to dignity.
• why we need other people.
• how we’re willing to live together.

So lately I increasingly feel that understanding a fast-changing era may take two kinds of sight.

One eye on what’s happening at high speed.

New tools, new capabilities, new possibilities.

The other eye watching:

what these tools are ultimately serving.

Because much of what we chase every day is new methods.

And what really lets something last may still be the needs that haven’t changed nearly as fast.