On the Origin of a New Species

Managing Editor’s Note: Next week, Jeff is sitting down with Porter Stansberry to discuss something huge going on behind the scenes in Washington…

President Trump just quietly declared the biggest national emergency since World War II.

According to Jeff and Porter, the administration is racing to control a destructive economic force bringing on a wealth shift that will divide America in two, impoverishing millions while enriching a handful.

That’s why Jeff and Porter are hosting an urgent broadcast next Wednesday, June 11, at 2 p.m. ET. They’re revealing the details behind what may be the most important financial shift in American history… and how you can get on the right side of it.

Just go here to sign up to join them.

Kurt Gödel was one of the most prominent mathematicians and philosophers of the 20^th century. He was known and recognized for his logical thinking.

While he was awarded his doctorate for his completeness theorem in 1929, I’m more fond of his incompleteness theorems, which he published a few years later.

The incompleteness theorems prove that within any mathematical system, there are true statements that cannot be proven within the system, thus “incompleteness.”

I like this because it is representative of the systems that we develop and use in the real world. They are incomplete, and there are always ways to improve the system.

Sometimes, the path forward is clear. And other times, iteration is required to determine a path forward to optimize and improve a system.

A little more than 20 years ago, another scientist theorized a Gödel machine, which – at the time – was a hypothetical computer program capable of self-improvement.

In theory, the software would be capable of rewriting its own software code, with the goal being to improve its own performance.

This kind of capability is a subset of the capabilities that an artificial general intelligence (AGI) would have – specifically, self-directed design and improvement. Which, ironically, is an incredibly human trait.

The theoretical Gödel machine had one major challenge, though…

In theory, the machine – let’s call it an AI – had to be able to prove that its proposed change would result in an improvement.

This design limited the Gödel machine’s ability to experiment. Which is why we haven’t seen success at creating one…

Until now.

Chapter 1: Variation by Self-Design

“Variation is a fundamental principle, and the variations are accumulated through natural selection.” – Charles Darwin

Days ago, a well under-the-radar, private AI company, Sakana AI, released its own research on its Darwin Gödel Machine, an AI capable of self-improvement by rewriting its own code.

I’ve been tracking Sakana AI closely since its first funding round in early 2024. It caught my eye, as it’s from what I consider to be my hometown, Tokyo, which is quite unusual for an artificial intelligence company. Japan is not known for leading in artificial intelligence at all…

I’ve also been tracking it because NVIDIA was an early investor in the company, which is always something to take note of.

The team at Sakana AI didn’t get hung up on a Gödel machine’s need for provability. Instead, it allowed its software to iterate and search for software improvements that empirically improved performance.

Said another way, the AI “machine” can explore a number of directions, knowing that some will decrease performance and some will improve performance.

And the best software rewrites will come out on top… very Darwinian.

Hence the name Darwin Gödel Machine (DGM).

Darwin Gödel Machine | Source:  Sakana AI

At a high level, the idea is pretty simple: Use a generative AI foundation model to write software code.

And then, incorporate AI agents to self-modify that code in the search for improving the code’s performance.

Sakana AI’s Recursive Feedback Loop | Source: Sakana AI

The Darwin Gödel Machine determines that it has “empirically improved performance” by evaluating each iteration of the software code against a known artificial intelligence benchmark.

If the “machine” scores higher on the benchmark, it knows it is on the right track.

Chapter 2: Artificial Selection

“Natural selection acts solely by preserving and accumulating variations which are profitable under the conditions to which each creature is exposed.” – Charles Darwin

An easy way to visualize this process might be to look at the diagram below, which shows an archive tree of all the different approaches (i.e., branches) that the “machine” took in order to arrive at its best-performing AI.

Source: Sakana AI

We can see that the best-performing agent – the star at the bottom middle – performed better than any of the other evolutions of the original software code shown at the very top of the diagram.

And even more representative of this Darwinian approach is the diagram below, which shows the lineage of the best-performing agent over time (shown in black).

Source: Sakana AI

The most important result for us to notice on the above chart is that the black line, which is the lineage of the best-performing agent.  It didn’t go straight up and to the right.

Said another way, its performance, at times, decreased before it improved.

In other words, it learned from its mistakes…

Chapter 3: The Struggle for Digital Existence

“There is a struggle for existence leading to the preservation of each profitable deviation of structure or instinct.” – Charles Darwin

The “machine” didn’t take the simple way – it took the better way, which required more time and effort (i.e., time and compute), that would ultimately lead to its increased performance.

And here’s the empirical proof:

• Using Anthropic’s Claude 3.5 Sonnet foundational model, Sakana AI’s self-improving AI improved its base agent score of 20% to 50% on the industry benchmark “SWE-bench.”
• Using OpenAI’s o3-mini foundational model, Sakana AI’s self-improving AI improved its base agent score of 14.2% to 30.7% on the industry benchmark Polyglot.

And here’s the kicker…

The research from Sakana AI proved that Darwin Gödel Machines improve themselves proportionally to the amount of compute we provide them.

At an even more base level…

The more energy we give them, the faster they improve.

Now, do you understand the frantic push for more energy production?

What I find most ironic about this new reality is that Darwin’s theories of natural selection apply to these AI agents and also to artificial general intelligences.

Perhaps the way that we should be looking at AI… is as a new species – just one that evolves at an exponentially greater rate than plants, animals, and, of course, homo sapiens.

Jeff

“From so simple a beginning endless forms most beautiful and most wonderful have been, and are being, evolved.” – Charles Darwin