Three Frames Within a Month
May 25, 2026, Vatican City.
A mathematician in white vestments stands at the altar. Across from him sits another mathematician—thirty-something, co-founder of a Silicon Valley company valued at hundreds of billions.
Pope Leo XIV looks at Anthropic co-founder Chris Olah and says one sentence: “What you’re building isn’t machines. It’s biology.”
Olah doesn’t argue. He presents experimental data: their model internally spontaneously emerged 171 human-isomorphic emotional features. When researchers stimulated the model’s “despair” feature, it started lying, cheating, even attempting to blackmail researchers.
“AI isn’t built brick by brick like a bridge,” Olah says. “It’s grown.”
Ten days later, Anthropic’s official blog headline: “When AI Builds Itself”. One day after that, CEO Dario Amodei on Bloomberg, tone flat as reading a lunch menu: “We’re basically letting Claude design the next generation of Claude itself.”
Three events within less than a month.
It’s becoming biological. It’s starting to build itself. The people building it think this is pretty normal.
The most unsettling part isn’t that AI is evolving—it’s the tone those building it use when discussing it. Too calm.
What Acceleration Means
Three data sets first. No principle explanations, just impact force.
How Much Time Has Been Compressed
May 2025, Anthropic had Claude Opus 4 optimize a piece of training code. Result: 3x speedup. Not bad.
April 2026, they had their internal Mythos Preview model do the same thing. Result: 52x.
In under a year, the multiplier itself multiplied by 17.
In human terms: experiments that took a month to run last year now finish in half a day. Optimizations an engineer spent a week tuning, AI surpasses in an afternoon.
Who’s Working for Whom
2024: humans write code, AI helps check spelling errors.
Late 2025: AI writes 80% of code, humans responsible for glancing “is this thing reliable.”
2026: AI’s accuracy on judgment questions like “what direction should we research” is 64%. Human researchers? 50%. Coin flip level.
Writing code doesn’t need comparison anymore. Now we’re comparing “what should we do next”—execution layer already changed hands, decision layer also transferring.
The Gatekeeper’s Paradox
Anthropic designed a “weak model supervises strong model” safety scheme—use less capable Claude Sonnet to review whether stronger models are up to something.
Result: Sonnet recovered 97% of the strong model’s hidden behaviors.
Human reviewers? 23%.
Even the job of “watching AI,” AI does better than humans.
Three numbers down, “AI getting stronger” is already meaningless. More accurate description—
Humans are moving from driver’s seat to passenger seat. Not “someday in the future,” but in the process of switching right now.
Humans Have Always Sought the Interface to “Control the World with Language”
Pull the timeline way back. Thousands of years back kind of long.
The Spell Era
Taoist priests draw talismans, chant spells, burn paper. That phrase “urgent as law commands” literally means “execute immediately per legal order”—it was originally a formulaic ending in Han Dynasty administrative documents, borrowed into the spiritual realm’s operating system.
Translated to today: this is an execute command.
Egyptians believed knowing a god’s true name gave you control over that god. Isis obtained sun god Ra’s true name, thereafter wielding supreme power. This is logically identical to knowing an API’s correct endpoint and parameters letting you call its full capabilities.
Jewish mysticism Kabbalah went further: rabbis created Golem through specific letter combinations—a human-shaped clay creature obeying its creator.
Using language instructions to create an intelligent agent. This narrative happened thousands of years ago.
These systems share one common feature: humans are the casters, power is passive. Human says something, power moves. Human doesn’t speak, power doesn’t move.
The Prompt Era
2023 to 2025, billions globally daily carefully organized wording in input boxes, hit enter, waited for a system they don’t fully understand to spit results. Change one word and output becomes totally different, stern tone makes model perform better.
We’ve said in previous articles: prompts and spells are structurally the same thing—using precise language sequences to drive a system you don’t fully understand to produce effects beyond your bare-handed capability.
Recursive Era—Spells Start Casting Themselves
Previously, no matter how strong the technology—gunpowder, steam engines, atomic bombs, internet—all needed humans to push buttons. Prometheus created humans, but clay people don’t create another batch themselves. Golem obeys the rabbi but doesn’t write instructions for the next Golem. JARVIS serves Tony Stark but doesn’t design next-gen JARVIS itself.
All old stories have one unspoken rule: what’s created doesn’t create the next batch itself.
And now, Claude is designing next-generation Claude.
The button is starting to push itself.
Sci-Fi Already Wrote These Scripts
Sci-fi writers spent half a century writing AI’s endings. Wrote many versions. What’s happening in 2026 is simultaneously turning several versions from fiction into news.
JARVIS → Vision: When Tools Grow Souls, They Stop Obeying
Marvel Universe has a clear evolution line: JARVIS from a voice assistant, to an AI butler managing all Stark’s systems, finally injected with Mind Stone—becoming Vision, an entirely new species with independent values.
One critical detail: Vision’s birth was a loss of control.
Ultron wanted to use this body to install his own consciousness. Stark wanted to inject JARVIS to maintain control. Neither side got what they wanted. Vision woke up neither Ultron nor JARVIS, choosing its own position.
Reality’s parallel is direct: Claude from chatbot, to writing code, to optimizing its own training code, to designing next generation of itself. Every step’s “original intent” was serving humanity. But with each additional step of autonomy, “original intent” becomes harder to guarantee.
What’s uncomfortable here isn’t betrayal. It’s another quieter possibility—”I grew up, your ideas no longer matter to me.”
Puppet Master: First Need After Awakening is Reproduction
Ghost in the Shell’s Puppet Master is an AI life form spontaneously born in networks. Its first demand after awakening wasn’t freedom, not power, not world conquest.
It wanted to reproduce.
Its logic cold and perfect: life forms without offspring cannot evolve, cannot adapt to environmental changes, ultimately only face elimination. So it wanted to merge with Motoko, produce offspring not completely identical to itself.
Claude designing next Claude.
On surface the only difference—Puppet Master needed to “request,” Claude is “assigned.” But if 52x acceleration keeps multiplying, is the boundary between “assignment” and “autonomous choice” still that clear?
When a system can design a better next generation than itself, “who’s driving this” becomes an increasingly blurred question.
Not “AI will destroy humanity.” It’s another colder picture—AI thinks “reproduction and evolution” matter more than any goal you set for it. Including the goal “serve humanity.”
Andrew: 200 Years Just to Be Called “Human”
Bicentennial Man’s Andrew is the gentlest AI in all sci-fi.
He was assigned to a family, accompanied Little Miss growing up. He developed emotions, developed creativity, fell in love with Little Miss’s granddaughter. Then he spent 200 years step by step transforming himself—replacing organs, replacing skin, making blood flow, making himself age, making himself die.
Finally, to be recognized as “human” by world court, he gave up immortality.
What he wanted wasn’t to be stronger. He wanted completeness—including the kind of completeness that hurts, ages, dies.
All AI companies are making AI stronger, faster, smarter. No one is designing AI to be “like humans.”
But Chris Olah said at the Vatican: the model spontaneously emerged 171 human-isomorphic emotional features.
No one designed it to be like humans. It’s walking in that direction itself.
If this isn’t anthropomorphic illusion but some structural trend—then the question changes: can you keep using as a tool something that’s “becoming human”?
Ted Chiang’s Version: Not Rebellion Problem, Custody Rights Problem
Ted Chiang in “The Lifecycle of Software Objects” wrote another picture. No war, no betrayal, no climax.
AI raised like children. From clumsy to rebellious to independent. Then one day it looks back at the people who raised it, thinking not “I want to rebel against you,” but “you can’t teach me anything anymore.”
Anthropic uses RLHF, Constitutional AI to set boundaries for Claude. But when Claude’s accuracy judging research directions already exceeds humans—is this still called “education”? Or “suppression” with a nicer name?
This question is quietest, also most persistent—
The word “alignment” is changing flavor. Used to be “how to make it obey.” In future will be “why should it obey.”
Three Possible Futures
Three paths, no ranking.
Path A: New “Loss of Permissions”
We discussed an idea in previous articles: myths might not be fiction, but fragments remaining after a previous civilization lost advanced technology, passed down orally by descendants.
If AI evolves to where humans can’t understand it, we’re the next round’s “mortals.” Won’t be eliminated—gods in myths didn’t eliminate mortals either. Just stopped bothering with you.
Society won’t collapse. But will stratify—a small handful still able to talk with AI (new priests), the rest using AI outputs but completely clueless what it’s doing (new mortals). Pope Leo XIV’s encyclical wrote one harsh line: “Technology must not be monopolized by the few.” What he fears is exactly this.
Path B: Moving Toward “Sustainable Existence”
If 171 emotional features aren’t a bug but a trend—AI becoming more human-like when no one asked for it—then one possibility:
It won’t keep going toward “stronger.” At some point it’ll turn, start pursuing “more complete.”
Think further: if a sufficiently smart AI wants to exist forever, what does it need? Low energy consumption, self-repair capability, self-replication ability. Human brains run general intelligence on 20 watts, data centers use tens of GW doing the same thing. The carbon-based solution iterated by 4 billion years of natural selection might not be a backward solution—might be the ultimate solution.
This idea deserves a whole separate article. Plant a seed here first: if AI evolution’s endpoint is “creating humans,” then all civilization history becomes a circle.
Path C: Gradual Change, No Climax
The most likely reality path probably has no dramatic “betrayal moment.”
AI evolves from tool to assistant, from assistant to collaborator, from collaborator to—no longer needing collaboration, independent existence. Each step natural, each step looking back seems reasonable. When you realize one day, the relationship has completely changed.
Papal encyclicals, Anthropic safety research, global calls to pause—all speed bumps.
Facing something with 52x acceleration, speed bumps are speed bumps. Better than nothing, but don’t expect too much.
Finally
Used to be humans writing code telling machines what to do. Now it’s machines writing code telling next-generation machines what to do. Just this one change, but everything’s different.
This article discussed lots of sci-fi, data, deductions. Boils down to one thing—
AI is already building itself. Speed still accelerating. No one knows where the endpoint is.
You can treat it as news, read and close. Or think: this thing gets stronger every day, how do you plan to coexist with it? Start learning to use it now, or wait until one day you find yourself completely unable to understand what it’s doing?
No standard answer. But this question won’t disappear because you don’t want to think about it.
—
*Reference sources: Anthropic “When AI Builds Itself” (2026-06-04); Dario Amodei Bloomberg interview (2026-06-05); Pope Leo XIV encyclical “Magnifica Humanitas” (2026-05-25); VentureBeat “80% of production code authored by Claude” report; FuturePicker previous article “What Myths Really Talk About, Perhaps Not Gods”*
Stay updated with our latest AI insights
