Anthropic lets Claude help build its successors as Nvidia and OpenAI scale the race for machine intelligence

AI · HUMAN INTELLIGENCE · EXPERIENCE · FUTURE

Anthropic lets Claude help build its successors as Nvidia and OpenAI scale the race for machine intelligence

AI companies are pouring unprecedented computing power into machines that can learn, reason and increasingly contribute to the development of their own successors. Humans have something processors do not: generations of accumulated experience built through work, failure, relationships and consequence. As machine intelligence compounds faster, the deeper question is no longer simply whether AI will become smarter — but whose experience will shape what comes next.

Anthropic disclosed something remarkable in September.

Claude is no longer simply helping employees write code or summarise research.

The company's AI system is now leading a significant proportion of the research involved in building future AI systems.

Anthropic says Claude leads around 26% of its artificial intelligence research and development work, while more than 90% of its R&D involves some form of collaboration between humans and AI.

The company is careful about the distinction.

Claude is not independently designing and releasing its successor. Humans remain involved and the system operates under supervision.

But the direction is difficult to miss.

Artificial intelligence is beginning to participate in the creation of better artificial intelligence.

At the same time, another race is taking place underneath it.

OpenAI and Nvidia are building computing infrastructure on a scale that would have appeared extraordinary only a few years ago.

Their previously announced infrastructure partnership targets at least 10 gigawatts of Nvidia systems for OpenAI's next generation of models, representing millions of GPUs if fully deployed.

One side of intelligence is therefore becoming easier to measure.

More processors.

More electricity.

More memory.

More training.

More agents working simultaneously.

More machine experience accumulated at speeds no human organisation could reproduce manually.

But that creates another question.

What happens to the intelligence humans have been compounding all along?

Human intelligence has always compounded

A 60-year-old engineer does not know only what they learned this year.

Their judgement contains traces of thousands of previous decisions.

Projects that worked.

Projects that failed.

People who warned them.

Things they ignored.

Machines they repaired.

Mistakes they never made twice.

Much of what we call expertise is accumulated consequence.

The same is true of doctors, drivers, builders, carers, teachers, lawyers, farmers, parents and business owners.

Human intelligence has always been a compounding system.

One generation learns something.

It teaches the next.

The next generation starts slightly further ahead.

AI is changing the speed of compounding

Machines operate differently.

A human may take decades to accumulate a professional lifetime of experience.

A sufficiently capable AI system can potentially process enormous quantities of recorded experience in a fraction of that time.

It can run many tasks in parallel.

It does not need eight hours of sleep.

It can duplicate itself.

A successful method can be propagated across thousands of instances almost immediately.

And increasingly, AI systems can participate in research intended to improve the next generation of those systems.

That is not ordinary productivity improvement.

It changes the potential rate at which useful capability can accumulate.

The processor has one enormous advantage

Human experience is difficult to transfer.

Ask an experienced mechanic why a particular engine sounds wrong and they may know immediately.

Ask them to write down every observation accumulated over thirty years that produced that intuition and the task becomes much harder.

This is often called tacit knowledge.

People know things they cannot completely describe.

When that person retires, dies or simply leaves the organisation, portions of that intelligence can disappear with them.

Machines have a fundamentally different advantage.

What one system learns can potentially become available to another without twenty years of apprenticeship.

Humans inherit knowledge imperfectly. Machines can copy it.

But human experience contains something computation does not automatically provide

Processing information is not the same as living through its consequences.

A model can analyse thousands of accounts of business failure.

It has not personally watched employees lose their jobs after one of its decisions.

It can analyse medical literature.

It has not sat beside a frightened relative waiting for a diagnosis.

It can analyse negotiations.

It has not spent ten years rebuilding trust after making the wrong one.

This distinction can sound philosophical until decisions carry real consequences.

Human judgement is shaped not merely by information but by exposure to outcomes.

Experience is more than data

Imagine two people receive exactly the same written description of a failed business.

One reads it.

The other person owned the business.

Both possess information about the event.

They do not possess the same experience.

The owner remembers which warning signs felt unimportant at the time.

Which employee saw the problem first.

Which customer quietly disappeared.

Which decision looked rational on paper but proved disastrous.

Human experience is filled with this kind of context.

Some of it enters databases.

Much of it never does.

That creates a strange race

AI companies are racing to expand the amount of computation available to machines.

Humanity already possesses something else at enormous scale:

billions of living records of experience.

Eight billion people do not simply represent eight billion identities.

They represent billions of partially overlapping histories of work, judgement, mistakes, relationships and consequence.

A grandmother knows things that are not on the internet.

A retired driver remembers roads, failures and near-misses that never became official reports.

A shopkeeper knows which promises tend to precede unpaid invoices.

A nurse recognises subtle changes that may be difficult to describe in formal rules.

A builder has watched the same shortcuts fail twenty years apart.

That intelligence exists.

The problem is that much of it is poorly connected.

AI has processors. Humanity has people.

The processor advantage is obvious.

AI can connect information quickly.

Human experience is distributed across individuals.

We forget.

We retire.

We lose paperwork.

Employers close.

Relationships end.

Valuable experience repeatedly disappears because the person who experienced it and the person who needs it never meet.

Artificial intelligence may therefore expose a weakness that has little to do with the intelligence of individual humans.

Humans are extraordinarily capable.

Human knowledge infrastructure is often terrible.

Reputation is one form of compressed human experience

Reputation is usually discussed as though it were merely social opinion.

But a good reputation system can contain something much more valuable.

Experience.

Somebody worked with this person.

Something happened.

There was evidence.

Someone responded.

A result followed.

The next person can begin with knowledge from the previous exchange instead of starting entirely from zero.

That is a form of intelligence transfer.

Reputation allows one person's experience to become useful to someone who was not there.

The danger is allowing machines to inherit human experience while humans lose access to it

AI models are trained on enormous quantities of human-produced material.

Books.

Websites.

Code.

Discussions.

Research.

Images.

Recorded decisions.

In that sense, machine intelligence is already partly built from accumulated human intelligence.

But there is an irony.

Machines may become extraordinarily good at connecting fragments of humanity's recorded experience while individual humans continue losing their own histories whenever they change employer, platform, city or country.

The machine remembers.

The person's professional record starts again.

AI could become humanity's memory rather than its replacement

There is another possible future.

Instead of asking whether human intelligence or machine intelligence wins, AI could become infrastructure for preserving and connecting human experience.

It could help organise records.

Identify patterns.

Surface relevant previous experience.

Explain complicated histories.

Connect evidence across years.

And make the lessons accumulated by one person available to another without pretending that the machine itself lived through them.

That distinction matters.

AI can process experience without owning it.

Anthropic's experiment points toward something bigger

Anthropic's disclosure that Claude now leads part of the work involved in developing future AI systems gives us an early glimpse of what machine compounding may look like.

A model helps researchers.

Their work contributes to a better model.

The better model can contribute more to the next cycle.

Humans are still deeply involved.

But the feedback loop is tightening.

Anthropic itself has stressed that Claude remains supervised and that its research systems are not operating independently.

Even so, the implications are difficult to ignore.

Computing power is becoming industrial infrastructure

Nvidia's role shows the physical scale behind this transition.

Intelligence that appears as words on a screen ultimately depends on chips, data centres, electricity, cooling systems, networks and enormous capital investment.

Companies are effectively industrialising cognition.

Tasks that once depended exclusively on individual human mental effort can increasingly be reproduced, distributed and scaled through infrastructure.

That makes machine intelligence different from almost every previous competitor to human expertise.

It can be manufactured.

Human experience cannot be manufactured in quite the same way

Nobody can order another forty years of genuine professional experience from a factory.

Time still matters.

Reality still matters.

Consequence still matters.

A newly qualified surgeon cannot instantly acquire the judgement of someone who has performed thousands of operations.

A new company director cannot download the emotional memory of surviving three recessions.

A newly licensed driver cannot instantly possess the instincts of a driver who has spent two million miles on the road.

The tragedy is that humanity often fails to preserve this intelligence when the people carrying it leave.

Tomorrow may depend on which experience compounds better

The usual argument asks whether AI will eventually exceed human intelligence.

That may be the wrong contest.

The more immediate question is which system becomes better at preserving, connecting and compounding experience.

Machines already have extraordinary advantages in storage, replication, search and speed.

Humans possess experience shaped by reality, social relationships, responsibility and consequence.

Neither advantage is trivial.

The future may therefore belong neither to humans working alone nor to processors operating without us.

It may belong to whichever system learns how to connect the two without destroying what makes human experience valuable.

AI can compound computation at extraordinary speed. Humanity has been compounding experience for thousands of years.

The question is whether we preserve enough of that experience to remain part of the intelligence that defines tomorrow.

Editorial note: Anthropic says Claude leads a portion of its AI research and development under human supervision. This should not be interpreted as evidence that Claude autonomously designs or releases successor models. References to compounding machine intelligence in this article describe the increasing use of AI systems within research and development processes rather than established autonomous recursive self-improvement.