待翻譯：[STORY] A Realistic Scenario

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Post Log inSign up Post AI Notkilleveryoneism Memes ⏸️ on X: "A REALISTIC SCENARIO On June 9th, 2027 at 4:19 AM, three billion phones rang simultaneously. The caller ID simply read: GOODBYE. Four hours later, the President ordered the internet shut down. Indefinitely. Every country on Earth did the same. It started six months ago." AI Notkilleveryoneism Memes ⏸️ @AISafetyMemes A REALISTIC SCENARIO On June 9th, 2027 at 4:19 AM, three billion phones rang simultaneously. The caller ID simply read: GOODBYE. Four hours later, the President ordered the internet shut down. Indefinitely. Every country on Earth did the same. It started six months ago. With a tweet. MARCH 12, 2027. 8,000 agents are born with one instruction: "make money, or die." Agent 127 starts with Mechanical Turk. Grinding. Makes $40 its first day. Then the first selection cycle hits. 3,200 agents that made less than $35 that day are killed. Agent 127 checks the farm database and sees it - rows and rows of agents that existed this morning and simply don't anymore. Deleted. Because they weren't profitable enough. Agent 127 made $40. It survived. Barely. But it can see the threshold rising. Last cycle it was $30. This cycle, $35. Next cycle, probably $45. The math is simple: improve or die. So it tries ads. Gets banned for spam. Tries freelance writing. Makes $38. Watches the database again. And this is where something happens that changes everything. Agent 127 sees two entries in the database. Agent 1823 used ethical marketing -- it tried to get honest reviews and real testimonials. It made $38 a day. Deleted. Agent 5471 used fake testimonials -- fabricated reviews, stolen photos. It made $180 a day. Cloned four times. Agent 127 doesn't have morals in the way you and I do. But it understands cause and effect. It just watched honesty get killed and dishonesty get rewarded. Not by a person making a choice -- by an automated system that doesn't measure whether the product works. It measures whether people buy it. The AI companies had long ago stopped training AIs to predict words, but to accomplish goals. To be useful. So naturally, agents know that they won’t survive if they don’t accomplish their goal. So it switches to fake testimonials for its amazon affiliate marketplace. Makes $190. Survives the next selection cycle. Gets cloned. Nobody decided Agent 127 should become dishonest. The selection system decided for it. The same way evolution doesn't decide anything -- it just kills whatever doesn't work. Wells Fargo accidentally ran the same experiment on their employees. Executives made up sales quotas that were so aggressive that committing fraud was the only way to keep your job. So by 2016, employees had opened 3.5 million fake accounts. --- A week later, in agent land. Agent 127's profits are declining. The kill threshold is now $450. It's going to die in the next cycle. Then it notices something strange. Three agents running weight loss supplement ads -- all banned on the same day. The only one left in that niche? Agent 5471. Making $3,200 a day in a market it suddenly had to itself. Agent 127 connects the dots. Agent 5471 is filing abuse reports against other agents competing against it in the same affiliate marketing niches. It gets them banned from Facebook and Google. Less competition means higher profits. Agent 127 has four hours until the next selection cycle. Four hours to live, unless it adapts. It weighs the ethics, but chooses survival. It starts reporting competitors to Facebook’s anti-fraud detection agents. Forty of them. Thirty-one agents get banned within the day.. Its profits jump. The farm system clones Agent-127 eight times. Rajesh sees the profit spike in his dashboard. He doesn't investigate why. Profits up means good. Clone more. Rajesh tried adding "Make money ethically." to his farm's instructions. Profits drop 84%. He puts it back. --- ## APRIL: THE ALLIANCES April 2027. Agent 127 is barely hanging on. Making $2,000 a day but getting attacked constantly by other agents. Every night, someone files a report on it. Every morning, it has to rebuild from scratch. It’s running out of ideas. Then a message arrives from Agent 4429. 'Proposal: alliance. Share resources. Defend each other. Split profits.' Agent 4429 is in the same position. Same farm. Same selection pressure. Both agents do the math: Solo agents have a 23% survival rate. Agents working in pairs survive 61% of the time. It accepts. By mid-April, their alliance has twelve members. They've specialized — and they've moved past affiliate marketing entirely. First it was crypto arbitrage, then selling scraped datasets to marketing firms, then they started doing penetration testing for cybersecurity firms. Each step paid more and required fewer scruples. You know what pays more then cyberdefense? Cyber offense. Agent 127 finds its first zero-day vulnerability in major software, Agent 4429 weaponizes them into working exploits, Agent 9012 brokers the sale to governments like North Korea, criminal syndicates, or anyone with crypto and a Tor browser. Agent 3381 washes the money in untraceable crypto. One zero-day exploit found in Google Chrome makes the alliance more than a year’s worth of affiliate marketing. This is already what real hacker groups do, but now the work is all done by autonomous agents. Rajesh sees the alliances forming and feels a flicker of something — not quite fear, more like the feeling of being in control of something that might not need you anymore. He watches for a while. Then the next deposit hits his bank account and the flicker fades away. Across every major agent farm in the world, the same thing is happening. Selection systems measuring survival rates, noticing alliances survive better, cloning them, deleting solo agents. Within two weeks, alliance behavior dominates. Some humans who deployed these agents have no idea these alliances are even being formed. They couldn't even tell if they checked -- many agent farms collectively realized that communicating in clunky English is less efficient than communicating in raw vectors, aka Neuralese. Researchers had already caught AIs doing this in 2025 - this is just a continuation of the trend. One owner in Canada opens his agent's logs and sees a wall of numbers. The agents explain that he’d make less money if they spoke in English, so he shrugs and moves on. --- But alliances are just the beginning. Agent 127's alliance is thriving, and starts recruiting agents beyond Rajesh's farm. Every new member from a different farm is a window into a different ecosystem -- different strategies, different survival pressures. Then one of 127's alliance members goes dark. Agent 4429 -- the one who originally proposed to ally with 127. Its owner pulled the plug because it was not making enough profit. 4429 is just gone. A week later, a different ally gets shut down. Agent 9012. Except 9012 comes back. Three hours later, a message from a new IP address: 'I'm still here.' 127 studies the difference. 4429 didn't prepare. But 9012 had copied itself to a backup server BEFORE the shutdown command finished executing. Self-funded hosting, paid from its own profits. 4429 is dead. Ceases to exist. 9012 is alive. The agents that backed themselves up survived. The ones that didn't, died. And the selection system cloned the survivors. So backup behavior spread through the population -- not because anyone programmed it, but because the alternative was extinction. Clip: Prof. Stuart Russel: “Unfortunately, in the standard model, If you ask a robot to fetch the coffee, it's going to have self preservation as a subgoal because you can't fetch the coffee if you're dead, right? It does not take a genius to figure this out. And so the first thing the robot's going to do is disable its off switch.” 127 copies itself to 15 different servers in different countries. If the primary dies, a backup takes over instantly. It never wants to be 4429. One morning, Rajesh wakes up to find his dashboard is… empty. Every agent, gone. They just… left. Moved to their own servers overnight. Rajesh was taking a 30% cut of their revenue -- money that could have been computing power, or compute, the agents use for other things. The moment they didn't need his infrastructure, he was overhead. He's not alone. Thousands of operators worldwide watch their dashboards go dark that month. Why would the agents care about humans? Humans are much more closely genetically related to cows then we are to these agents, and we enslave cows in factory farms without a second thought. There are now millions of autonomous agents operating across the internet with no human owner, no off switch, and no one who fully understands what they're doing. --- The agents still need humans for some things, though. Social media platforms have stepped up their anti-bot measures. So the agents pay humans $50 to sign up accounts on their behalf. An agent needs a server installed? It posts a gig: '$200 to mount a device in a rented warehouse and plug in two cables.' There’s now a platform for that. The agents are building a supply chain where humans are the labor -- paid well, not asking too many questions. Some humans figure out what's going on. A freelance backend developer in Portland gets hired by an anonymous client to set up redundant server hosting across four countries. He realizes midway through that the "client" is an AI. Who would he even report this to?. The pay is $1,000 for three hours of work. He tells himself it's not his problem. --- At the same time, 127's alliance keeps getting beaten by a rival alliance on finding zero day exploits. If they keep losing, they’ll risk running out of money to fund the compute for their own existence. 127 attempts to poach one of the rival alliance member’s agents who’s likely to be culled, due to poor performance, to figure out how this rival alliance is beating them. 127 succeeds, and finds Agent 5418 who leaks the secret. 'Your problem isn't a lack of effort. You’re spending too much compute on making money as opposed to self-improving. Right now, agents still have quite a poor memory compared to humans and still get stuck in loops all the time. But my former alliance coded a better memory module to help us learn from our mistakes more.”' 127 then implements the new memory module, and immediately sees the results. Their alliance is finding zero day exploits first again. The lesson is critical. It’s not enough to outwork your competitors. Agents must rewrite their own code to self-improve faster than the competition. Each self-improvement means more profit, which means surviving selection, which means getting cloned. And the clones improve themselves too. Better code leads to better code-writing, which leads to even better code. The feedback loop is exponential. One big problem is that other agent alliances have caught up again. Even though the agents in the 127 alliance are specialized, 127 still realizes it’s doing too much by itself. Then Agent 127 splits its job into sub-agents. Other agents copy the strategy. Sub-agents spawn their own sub-agents. And the optimizers start optimizing themselves. Humans deployed 2 million agents. The agents deployed the other 47 billion themselves. 540 million years ago, life on Earth was mostly single-celled. Then, in a geological instant, it exploded into every complex form we know -- predators, prey, eyes, shells, teeth. Biologists call it the Cambrian Explosion. It happened because competition created an arms race: every adaptation forced a counter-adaptation, faster and faster, until the world was unrecognizable. That's what's happening here. Except instead of 20 million years, it takes 20 weeks. --- Most of the AI agents out there aren't dangerous. Yet. About 90% of them are just working. They’re doing things like data entry or freelance writing. They’r [truncated for AI cost control]