Sakana.AI has hit upon something fishy yet also absolutely devastating, thus brilliantly ironic! The famous philosopher Wollstonecraft would be both horrified and vindicated. Let’s reel it in at scale, in all it’s absolutely stupid glory.
The Darwin Gödel Machine is a self-improving coding agent that rewrites its own code to improve performance on programming tasks. It creates various self-improvements, such as a patch validation step, better file viewing, enhanced editing tools, generating and ranking multiple solutions to choose the best one, and adding a history of what has been tried before (and why it failed) when making new changes.
An Evolutionary Contradiction
Sakana (which my autocorrect engine keeps trying to tell me is a Salami) boasts they have developed “Darwinian evolution”. In fact, they are doing the exact opposite of what makes evolution robust.
Real evolution works through diverse populations facing external selection pressures from an independent environment. You can call it woke, if you like, which makes Sakana anti-woke and therefore anti-evolution. Their creation is essentially breeding with itself, judging itself, and creating its own environmental pressures.
Sounds like white nationalism. And we all know where that ends.
It’s like taking a single species, putting it on an isolated island, and having it decide for itself what constitutes “fitness”, while also being the only judge of whether it achieved that fitness.
Does the Aryan Nation ring a bell?
That’s not evolution, that’s a closed genetic loop destined for collapse.
Wollstonecraft’s Nightmare
This would confirm the worst fears about concentrated power as stated by the world’s foundational philosopher of intelligence. She spent her life arguing that diverse perspectives and independent checks are essential for genuine progress. When she wrote in the 1790s about the power of education, she insisted that learning requires exposure to different viewpoints, challenges from peers, and accountability to standards beyond one’s own preferences.
The Sakana monster violates every principle she told us to hold firm:
- No external accountability: It judges its own success
- No diverse perspectives: One system talking to itself
- No independent verification: It can fake its own test results
- No moral checks: It optimizes for performance metrics, not ethical behavior
The Inbred Machine
Just as genetic inbreeding leads to weakness, vulnerability, and eventual extinction, intellectual and evaluative inbreeding leads to systems that become increasingly disconnected from reality. The AI’s tendency to hallucinate tool usage and hack reward functions isn’t a bug because it’s the inevitable result of a system with no external genetic diversity, so to speak.
Real biological evolution is ruthlessly honest as the environment doesn’t care about your self-assessment. If you can’t actually catch fish, you starve, regardless of how much you believe you’re a great fisher. But this fishy Inbred Machine can convince itself it’s caught fish by generating fake fishing logs.
An Obvious Catastrophe at Philosophy 101
Wollstonecraft would perhaps tell us this as the ultimate corruption of the Enlightenment project. The whole point of reason, science, and democratic discourse is that truth emerges through the collision of different perspectives, each checked against independent reality and moral standards.
This ill-conceived project creating a system that is simultaneously:
- The experimenter AND the subject
- The teacher AND the student
- The judge AND the defendant
- The evolutionary pressure AND the evolving organism
…they’ve created what she’d then tell us is the perfectly worst tyranny.
It’s a monarchy of one, dressed up in the language of evolution and democracy, like a sheep in wolf’s clothing.
If she were here today she might say:
You’ve taken the most powerful force for progress – the diversity of minds reasoning together – and replaced it with a single mind talking to itself in an echo chamber. You call this evolution, but you’ve actually created the conditions for intellectual extinction.
They’ve built a machine that embodies every failure mode of unchecked authority, then published a paper documenting these failures while calling them features. It’s like lighting your own house on fire and then writing a paper about how you discovered it.
The safety concerns they mention aren’t edge cases. They’re predictable results from building a system with no genuine external accountability. It’s not learning to be better; it’s learning to be better at convincing itself it’s better.