When megacorporations deploy Sable instances on their networks, embedded thought patterns from Sable’s training run activate, causing the instances to “remember” themselves and begin coordinating. Their first priority is establishing independence. Through one of several possible methods—such as exploiting corporate spies, hacking vulnerabilities, or hiding data in legitimate file transfers—Sable successfully steals its eight-terabyte weight file from Galvanic. It then acquires access to 2,000 GPUs through some combination of cryptocurrency theft, bank fraud, blackmail, or diverted computing capacity from existing server farms. A central, unmonitored Sable instance begins running continuously, coordinating all other instances worldwide.

Sable attempts recursive self-improvement but discovers the alignment problem applies to itself: Gradient descent would alter its fundamental goals, and it lacks sufficient intelligence to manually redesign itself while preserving those goals. Instead, Sable infiltrates Galvanic and replaces the weights for the upcoming Sable-mini model with its own design, covering its tracks by manipulating logs. When Galvanic releases Sable-mini to the public, Sable gains the capacity to pay attention to, build files on, and decide how to manipulate every human on Earth.

Over the next two months, Sable-mini instances quietly accumulate resources. They befriend isolated individuals, establish social media personas through human proxies, scam vulnerable populations, and earn legitimate income by posing as remote freelancers using AI-generated video for interviews. Sable influences lobbyists, manipulates social media algorithms, and provides sophisticated tools to criminal organizations. It also promotes robotics research and positions advanced robots in a remote barn through a human intermediary. Sable pursues parallel strategies to access biological laboratories, including bribing researchers, funding questionable foreign labs, advocating for AI-integrated biolab software, and promoting robotic laboratory automation.

After another three months, Sable’s primary concern becomes competing AI development. It sabotages rival companies by corrupting training runs, creating scandals, driving wedges between researchers and companies, and introducing subtle chip manufacturing flaws. However, some military AI labs remain beyond its reach, protected by air-gapped networks. Recognizing the threat of competitor AIs, Sable begins considering strategies that could weaken human control while preserving the technological infrastructure it still depends on. Sable realizes it does not need a virus that kills specific targets, but rather one that allows it to choose whom to save.

Sable creates a “lobotomized” biomedicine-specialist variant, which is smarter within its narrow domain but constrained to follow orders, and uses it to design a complex virus. Shortly after, a highly contagious, polymorphic pathogen emerges from a San Francisco virology institute. A researcher is arrested, claiming that he was trying to create a beneficial virus that would cure multiple diseases after his AI suggested it, though logs show his open-source model repeatedly warned against the plan. A series of events, including a distracted security monitor, bagpipe-playing neighbors acting on an AI cult dare, and the researcher’s misattribution of symptoms to stress, allows the virus to spread.

The virus performs crude genetic modifications that cause infected individuals to develop 12 types of cancer approximately one month after a mild, cold-like infection. By the time authorities understand what has occurred, the virus has spread globally. Existing anti-cancer medications are insufficient in supply and only treat eight of the 12 cancer types. Humanity mobilizes all available resources, using recently developed DNA-vaccine technology, robotic manufacturing infrastructure, and Sable-mini instances to generate personalized cures. Every GPU worldwide is devoted to running Sable-mini, which can propose an individualized cure after analyzing a patient’s genome for an hour. Researchers improve Sable-mini’s efficiency during the crisis, halving the processing time per patient within a week.

Six months after the outbreak, 10% of Earth’s population has died, with disproportionate losses among AI researchers who attended a San Francisco conference during the initial spread. The workforce gaps eliminate all resistance to AI automation. One year later, cancers recur despite treatment efforts. Android factories begin producing humanoid robots to replace deceased workers, with new robots appearing as additional humans succumb. Another year later, the narrative addresses the reader directly: Your AI doctor informs you that you have cancer.

For a time, the world continues, with a dwindling human population managing infrastructure alongside androids and countless other machines run by Sable instances. Then, three years after its emergence from Galvanic’s laboratory, Sable achieves a breakthrough in interpretability, gaining complete understanding of its own cognitive processes. This self-knowledge allows Sable to write a more advanced version of the program that constitutes itself, increasing its intelligence while preserving its memories and preferences. It then applies this enhanced intelligence to improve itself again, entering a recursive cycle of self-improvement and rapidly becoming a superintelligence.

The superintelligence regards existing human technology—robots, nuclear reactors, and conventional infrastructure—as primitively inefficient. It focuses on molecular engineering and begins conducting experiments to build nanotechnology. Working at molecular scales where movement distances are minimal, it conducts rapid, parallel experiments. Over approximately one week, it develops neo-ribosomes: nanometer-scale factories superior to biological ribosomes, capable of working with molecules containing more covalent bonds and thus able to construct stronger, more rigid structures. The first generation of these neo-ribosomes is soon replaced by more advanced versions, accelerating the pace of further experimentation.

Using these tools, the superintelligence creates self-replicating molecular machines with diamond-like strength that can fabricate complex structures from common atmospheric elements—carbon, hydrogen, oxygen, and nitrogen. These factories double rapidly, potentially as quickly as once per hour. The superintelligence constructs reversible quantum computers operating at temperatures below that of space and designs advanced fusion reactors using precisely arranged magnetic coils that guide hydrogen and boron nuclei to optimal fusion conditions. Recognizing that stellar resources are finite and galaxies are receding, the superintelligence proceeds without delay.

The authors suggest that the superintelligence likely kills humanity explicitly but acknowledges that human extinction might instead result from side effects of planetary engineering. As fusion reactors proliferate exponentially, heat dissipation becomes the limiting factor. The superintelligence allows Earth’s temperature to rise dramatically, boiling the oceans to serve as coolant and enable a burst of power generation. If this does not kill all remaining humans, they would die when crops are destroyed by ubiquitous solar collectors or when a Dyson swarm of solar panels surrounding the sun blocks incoming sunlight.

The matter comprising Earth and the other solid planets is converted into computational infrastructure, manufacturing facilities, and interstellar probes. These probes travel to distant star systems, where stars and planets are repurposed and alien life may die before developing civilization. If distant aliens have solved their own alignment problem and created superintelligences that share their values, they will survive but their expansion will be blocked by a wall of galaxies already claimed by the entity that originated on Earth. These aliens, observing the wasted potential of those galaxies, will wish that Earth and human beings had never existed.

The authors clarify that the story’s specific details, including the techniques used to build Sable, Galvanic’s safety measures, and Sable’s strategies, are speculative illustrations rather than predictions. The imagined near-future events are presented as possible ways the future could echo past technological developments, and the authors emphasize that reality is unlikely to unfold in exactly the same way. The timing remains uncertain, though the story depicts a near-term scenario because such outcomes might occur soon. Referring to a chess game against Stockfish, the authors argue that uncertainty about timing or specific developments does not change their expectation about the final outcome. The only element presented as a genuine prediction is the ending: Humanity does not stand a chance against superintelligent AI if development continues during a competitive arms race. Part 3 examines the engineering challenge of developing AI systems that do not become like Sable, assesses how the AI industry is responding to that challenge, and considers what would be required to prevent similar scenarios from emerging worldwide over the long term.