AI Is Starting to Improve Itself

This week, we’ve explored two of the biggest questions surrounding artificial intelligence.

First, just how close are we to artificial general intelligence (AGI)?

Second, if AI has become so capable, why aren’t businesses seeing bigger results?

Today, I want to explore a third question.

What happens when AI becomes good enough to improve itself?

Because as you’re about to see…

It’s something that we need to start taking seriously.

Progress Compounded

One of the reasons technology continues to advance is because every generation gives us better tools to build the next one.

The personal computer helped engineers design faster computer chips. Those faster chips made it possible to create more sophisticated software. And better software then helped researchers design even more powerful computers.

Progress has always built on itself. But AI has the potential to take that process a step further.

Consider what happens inside an AI lab.

Researchers spend months writing code, designing experiments, analyzing results and testing new ideas before they finally produce a better AI model.

But what happens when those same researchers have an AI system that can write much of the code, summarize thousands of research papers, suggest promising experiments and test many of those ideas while they’re asleep?

Instead of testing one idea at a time, they can test dozens at the same time.

And as that new AI model becomes even better at helping researchers build AI, the whole process begins to speed up.

Computer scientists have a name for this.

It’s called recursive self-improvement, or RSI for short.

Until recently, RSI was mostly a thought experiment. But it seems to be happening much sooner than many people expected.

Earlier this week, I told you about Anthropic’s experiment where teams of AI agents tackled an open-ended machine-learning research problem.

There was no answer key.

The agents had to come up with their own ideas, write code, run experiments, analyze the results and keep refining their approach until they found something that worked.

According to Anthropic, the best AI team outperformed experienced human researchers working on the same problem.

But there’s one detail I left out.

Those AI agents weren’t researching cancer, or physics or climate change.

They were helping solve an artificial intelligence research problem.

In other words, researchers weren’t simply using AI to answer questions. They were using AI to help improve artificial intelligence itself.

That’s the basic idea behind RSI. And once you start looking for it, you can start seeing it everywhere.

Anthropic says Claude now writes more than 80% of the code that its engineers eventually merge into Anthropic’s own codebase.

Anthropic’s engineers haven’t stopped writing software. Instead, they’re increasingly reviewing Claude’s work, deciding which ideas to pursue and focusing on the hardest engineering problems.

Anthropic says that change has helped its engineering teams ship roughly 8X more code each quarter than they did just a few years ago.

Google appears to be heading in the same direction.

Last year, the company introduced AlphaEvolve, an AI system designed to discover better algorithms.

Soon after Google unveiled it, I wrote that AlphaEvolve might represent one of the biggest advances in AI we’ve seen in years.

And the results have been impressive.

Some of AlphaEvolve’s discoveries solved mathematical problems that had challenged researchers for years. Others were good enough that Google has already begun using them inside its own computing infrastructure.

Notice what’s happening here.

Google isn’t simply building smarter AI. It’s using that AI to help improve the technology behind its next generation of AI systems.

And this trend isn’t limited to a handful of American AI labs.

Last week, China’s Moonshot AI released Kimi K3, an AI model that rivals many of today’s leading systems on coding and reasoning tasks.

Unlike ChatGPT or Claude, Kimi is “open weight.” That means its inner workings are publicly available. Anyone can study the model, modify it and build on it instead of treating it like a black box.

As a result, researchers around the world now have access to a frontier AI model they can use as the foundation for their own work.

That’s how progress begins to compound.

Each new frontier model doesn’t just represent another breakthrough. It also becomes a better starting point for the one that comes next.

Here’s My Take

Not long ago, AI companies relied almost entirely on human intelligence to build better AI.

Today, they’re relying on AI to do more and more of that work.

Last Sunday, billionaire investor Chamath Palihapitiya argued that we’ve already entered a recursive self-improvement loop, where each new AI model helps create an even better one.

I don’t know that I’d go quite that far.

After all, researchers and engineers are still the ones setting goals, deciding which experiments to run and choosing which improvements become part of the next model. AI might make coding a lot faster, but it hasn’t replaced humans yet.

However, I do think we’ve crossed an important threshold.

For the first time, the tool that helps us build the next generation of technology will be the technology itself.

This means the pace of AI progress no longer depends entirely on humans to move things forward.

And as AI takes on a bigger role in that process, the next few years could look very different from the last few.

Which brings me to one final thought.

The faster technology changes, the more important timing becomes.

I’ve learned that the best stock in the world can still lose you money if your timing is wrong. And an average stock can produce extraordinary gains if your timing is right.

That’s why when Keith Kaplan showed me his breakthrough timing algorithm, the Profit Switch, I immediately knew I had to bring it to my readers.

A 183% gain over four and a half years would have become 520% eight months earlier.

A 16% loss over two years would’ve become a 155% gain in one.

Same stocks. Same research. Only one thing different.

Timing.

Yesterday, Keith and I put on a special broadcast to reveal exactly how the Profit Switch works — and why it matters more than ever before August 16th.

If you haven’t watched yet, I urge you to do so immediately.

Watch the replay here.

Regards,

Ian KingChief Strategist, Banyan Hill Publishing

Editor’s Note: We’d love to hear from you!

If you want to share your thoughts or suggestions about the Daily Disruptor, or if there are any specific topics you’d like us to cover, just send an email to [email protected].

Don’t worry, we won’t reveal your full name in the event we publish a response. So feel free to comment away!