AI Just Started Improving Itself

For more than 25 years, Ray Kurzweil has been saying that artificial general intelligence (AGI) would arrive by 2029.

I believe that prediction might be too conservative. Kurzweil introduced the concept of AGI back in 1999 in his book The Age of Spiritual Machines.

By his definition, it’s the point where a machine can match human intelligence across a wide range of tasks. Something that can reason, adapt and improve.

For a long time, this concept seemed theoretical.

But it doesn’t anymore.

Recently, I mentioned Andrej Karpathy’s new “autoresearch” AI system almost in passing.

In hindsight, that was probably a mistake.

While its creator was sleeping, autoresearch kept trying different ways to improve its own results, writing code, testing it and refining things more than 100 times overnight.

And it did this on its own, without a human stepping in.

To me, that’s starting to look a lot like an early form of AGI.

A Small System With Big Implications

Andrej Karpathy has worked at the cutting edge of modern artificial intelligence for years. He led AI at Tesla, worked on Autopilot and was one of the early researchers at OpenAI.

But his new project, autoresearch, didn’t exactly make headline news when it was released earlier this month.

That’s probably because it doesn’t look like much on the surface. The whole system is roughly 630 lines of code, tiny by modern AI standards.

But what it accomplishes is much bigger than its codebase suggests.

Autoresearch is a research tool that makes changes to the model it’s working on, writes code to test those changes, runs experiments and then refines what works before trying again. And it does this inside a tight loop that doesn’t need constant human intervention once it starts.

That loop is the real story.

You see, most progress in research doesn’t come from a single breakthrough. It comes from iteration. You try something, measure it, refine it and repeat that process enough times that improvements start to stack.

Karpathy’s system automates this entire process.

That’s how it was able to run over 100 experiments in a single overnight session.

Of course, a human researcher could do the same thing. Eventually. But not in one night, and not without a lot of manual work.

And that’s the big deal behind this small amount of code.

With this new tool, humans will still define the boundaries of research. We’ll decide what to measure and what a “good” result actually looks like.

But the actual research loop will get handed off.

And once that happens, progress should start compounding. Because with autoresearch, each experiment feeds the next one, so the system can explore paths that a human researcher simply wouldn’t have the time to test.

This will ultimately change how research gets done.

Of course, researchers won’t disappear. But their role will move from manual experimentation toward more high-level tasks like the design of objectives and evaluation.

And autoresearch isn’t without its faults.

If you optimize too hard for a single metric, you run straight into Goodhart’s Law. That’s when a system starts chasing a score instead of an outcome, so it can look like progress on paper while drifting away from what actually matters.

This means someone still has to review its output.

In Karpathy’s example, it meant sorting through dozens of different versions to figure out what actually worked.

So this isn’t autonomy in the broadest sense. But it’s a step in that direction.

That’s why I see it as an early form of AGI. It’s not a system that can do everything, but it’s one that can improve how it works inside a defined environment.

That’s a narrower definition. But it’s a useful one.

And I’m not the only one who sees nascent AGI in today’s AI systems.

Nvidia CEO Jensen Huang recently said there’s a case to be made that we’re already seeing early forms of AGI, depending on how you define it.

When people talk about AGI, they often imagine a single breakthrough that suddenly changes everything. In practice, it’s more likely to show up like this, inside systems that take over pieces of a process.

But it’s still impactful. Consider what would happen if systems like this could eventually handle even just 50% of all research activity.

The upside is obvious.

Faster iteration means faster discovery. New drugs, materials and technologies get developed more quickly.

Some estimates suggest that kind of shift could increase global GDP by 7%, or roughly $7 trillion. Goldman Sachs has pointed to potential productivity gains of around 15% in advanced economies as AI adoption spreads.

That shows you the scale of what’s happening.

Here’s My Take

Right now, Karpathy’s loop works best in tight environments with fast feedback and clear goals.

But those conditions show up in more places than you might expect. Parts of software development, engineering and finance already fit this model

And we’re starting to see it spread.

Tools like Claude Code can now write, test and improve code with less human input. It’s a different interface with a similar underlying loop.

And once you see that pattern, it’s hard to ignore.

Until now, progress in AI was limited by how fast humans could run experiments. You could hire more people, but each person still worked one step at a time.

Autoresearch changes that.

Now systems can run dozens, even hundreds, of iterations in the same window.

And speed tends to compound.

Until you’re dealing with something that looks a lot like AGI.

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!