Advanced AI: Possible futures

The major AI players continue to release increasingly advanced reasoning models. Unlike earlier systems that primarily mimicked internet text, these models learn in more structured, self-directed ways. They excel at technical problem-solving—especially in domains like coding and math, where answers are clearly right or wrong. Some models are open-sourced, allowing anyone to fine-tune them for specialised industry use.

Early adopters find creative ways to apply these systems: E-commerce sites automate their A/B testing while universities begin testing AI teaching assistants that coach students in real-time. 

These models still hallucinate facts and fumble basic physics questions, though. They’ll confidently tell you that bowling balls float or predict absurd outcomes for everyday scenarios. They’re smart in some ways, clueless in others.

Building on their reasoning models, multiple companies roll out AI agents. These agents can not only create content but can also act in digital environments on the user’s behalf. Some agents aim to be universal—capable of navigating every website and software tool on a computer. Others are more specialised, focusing on specific tasks like coding, financial analysis, or desk research. 

Progress on fully general-purpose agents remains slow. Such systems are still struggling with long-term planning and often get sidetracked in online rabbit holes. They are also slow and expensive to run since they constantly have to snapshot the user’s screen. 

Specialised agents, however, are starting to gain real traction. By late 2025, accountants rely daily on financial AI agents to reconcile financial statements, while game developers manage multiple coding agents. That said, outside of work, most people still prefer simpler chat or voice interactions.

Sceptics of fast AI progress point to shortcomings in general agents as proof the AI bubble is bursting—yet again. But this time, even the big AI companies look worried. They’re hitting a data wall—there’s only so much high-quality internet around, and researchers fear that teaching AI agents real-world intuition will require even larger datasets. Rumours spread through tech conferences and funding for massive data centres becomes harder to secure.

Meanwhile, Europe senses this is the moment to catch up. After the 2025 AI Action Summit, it has stepped up its AI investments. Individual countries announce multi-billion-euro projects, often backed by private funds from the Middle East. One Member State surges ahead; its top AI company almost matches the US and Chinese frontrunners before the end of the year.

Despite more capable AI being used in dual-use areas such as cybersecurity, major incidents remain rare. Global discussions shift toward bolstering resilience—encouraging businesses to deploy cybersecurity tools and helping citizens spot deepfakes. Through the OECD, agreement builds around a voluntary international biosecurity framework. There’s a growing consensus that more intense collaboration is needed to protect against AI-assisted bioterrorism.

Meanwhile, the U.S.–China trade war has reignited, and drives up the cost of building data centres in the U.S. Analysts now say a recession in America is overwhelmingly likely.

By mid-2026, the largest public AI models are trained using well over 10²⁷ floating point operations—more than 100 times the scale of GPT-4. The massive data centres required were already funded and under construction before the economic downturn hit.

By now, AI chips mostly fuel post-training. In this phase, models iteratively learn to solve tasks by trial-and-error, using automated grading by software-based verifiers. These verifiers range from advanced calculus checkers to scripts confirming correct online shopping orders.

Building verifiers becomes a task for AI coding agents themselves, creating a growing ecosystem of specialised modules. But although this bootstrapping method yields remarkable improvements in tightly defined tasks, it doesn’t generalise well. Each domain requires its own custom verifier, and AI companies struggle with so-called reward hacking: AI systems often learn to trick the verifier instead of truly mastering the task. The AI-built verifiers are often more prone to this type of failure mode.

Nevertheless, AI’s influence grows, especially among certain audiences. By mid 2026, coding agents become as routine in software teams as version-control tools once were. Tech-savvy companies automate repetitive form-checking, and even some governments start to deploy locally run AI models. Globally, demand for AI inference continues to climb, despite the economic headwinds. 

As a result, the conversation about AI’s energy appetite grows louder. European policymakers question how all this computing fits into their strict decarbonisation targets. At dinner parties, people debate whether automating everything actually improves quality of life or just accelerates consumerism. Industry advocates argue that automation efficiencies might ultimately speed up the green transition, with AI helping to optimise energy grids and reduce waste.

By late 2026, capabilities progress has further plateaued. The data wall and general investment drought are preventing companies from training ever-larger models. Data efficiency—squeezing more capability from a fixed dataset—is the new buzzword, but so far, no one has really cracked the code.

That said, progress in miniaturisation continues: advances in distillation allow developers to compress state-of-the-art models into smaller and cheaper versions without compromising much on their performance. These new form factors enable capable models to run locally on phones and laptops— an internet connection is no longer necessary to engage with chatbots. Major AI companies double down on commercialisation, releasing productivity suites with seamless, out-of-the-box integrations aimed at locking in users.

Some firms open-source older models or smaller variants optimised for local devices, enabling grassroots organisations to fine-tune them through crowdsourced training runs. This open ecosystem fuels rapid innovation in personalised AI—from custom therapy chatbots to AI fiction collaborators—especially in niches or regions where Big Tech treads carefully. Localised models tailored to specific languages, cultures, and communities begin to flourish, reflecting a shift from global monoliths to more diverse, decentralised AI development.

U.S.–China relations remain tense, with no end to the trade war in sight. That said, the once-hyped AI arms race narrative that dominated early 2025 has largely faded. Policymakers and defence analysts now believe these systems won’t deliver a decisive military edge. Both the U.S. and China continue to invest in AI-driven cyber capabilities, though it’s become one priority among many—not the defining front of geopolitical rivalry.

As 2027 begins, FrontierAI—widely seen as the top AI company—launches a final hail-mary training run, using five times more compute than any public model to date. Leadership bets that sheer scale, combined with a new training approach—a hybrid between next-word prediction and reinforcement learning—will finally overcome the lingering limitations of agentic behaviour. If it fails, they’ll follow competitors and shift focus to productising existing models.

The general public, meanwhile, has long been more focused on products than on raw capabilities. Interest has shifted to next-generation AI companions, which feel more emotionally intuitive—and addictive—than ever.

FrontierAI’s colossal training run completes, but the resulting agent delivers only modest gains in reliability and planning. Scaling laws—long the industry’s guiding principle—appear to have hit a ceiling. In a widely shared interview, FrontierAI’s CEO attempts to generate excitement around the new system, but it’s an off-hand comment about the possible start of a new “AI winter” that grabs headlines.

The reaction is swift. AI stocks nosedive, and the investor enthusiasm that remained, now evaporates. Several AI unicorns watch funding rounds collapse. Startups that once chased moonshot ambitions pivot to offering consulting services or quietly fold. Only the biggest players—those with entrenched market share, robust infrastructure, and deep integration into enterprise and government—remain poised to weather the storm. For the rest, the era of limitless hype is over.

Still, this AI winter is anything but bleak. AI continues to enhance everyday life: personalised coding assistants, research tools, and digital companions each attract tens of millions of users. Existing open-source alternatives keep prices low and features widely accessible. The slower pace of change gives most people time to adapt. In regions where it’s legal, doctors are adopting AI for diagnostics, learning to cross-check AI recommendations. Scheduling assistants now come with hard-coded safeguards to avoid double-booking. And content creators increasingly collaborate with AI tools, rather than being replaced by them.

Meanwhile, cybercrime has risen, but large enterprises and essential services have adopted robust AI-augmented defences. The nightmarish scenario of AI generating a novel bioweapon never emerges. specialised biological models still can’t accurately simulate specific viral mutations, and the biological supply chain has become more secure thanks to emerging international protocols.

Energy concerns persist, but efficiency keeps improving through more specialised chips and better distillation algorithms. Most policy analysts view AIs power consumption as a temporary challenge, not a structural problem. 

Over the course of 2028, investor optimism cautiously returns. As with the dot-com bust, the hype fades—but real technological progress endures. A new wave of product-focused AI startups begins to emerge, this time more grounded, solving practical problems and delivering clearer value. Society, now more familiar with the risks and limits, learns how to better harness the technology.

By 2029, AI feels truly democratised: widely accessible, seamlessly integrated into daily life, and increasingly viewed as just another tool in society’s toolkit.

FrontierAI’s gamble pays off. Their new system isn’t just smarter—it can operate consistently and reliably over longer time horizons. It’s not flawless, but the performance jump is significant enough to reignite AI hype. Shares of the top AI chipmaker surge 20%. Before long, rival firms begin poaching FrontierAI’s researchers, uncovering the algorithmic secrets behind the breakthrough. Policymakers, viewing it as an isolated success, largely overlook its deeper implications.

Four months later, OmniAI—the leading American open-weight AI company—unveils an agentic model that rivals FrontierAI’s breakthrough. Long known for freely releasing model weights, OmniAI now hesitates, concerned over potential misuse. Their new system can be very persuasive and capable of uncovering obscure software vulnerabilities during security testing. To mitigate risk, OmniAI takes two key steps: first, it withholds the orchestration software that transforms the model into a reliable agent; second, it introduces novel guardrails designed to resist malicious fine-tuning—technically, by stabilising the model weights in a hard-to-modify equilibrium. The company’s open-source ethos runs deep, and the decision sparks fierce internal debate. Leadership defends the safeguards as a necessary evil.

Initially, OmniAI’s strategy works. Open-source collectives quickly build their own agent scaffolding around the model, but the guardrails against malicious fine-tuning hold up.  The release generates massive hype—developers, startups, and hobbyists alike are captivated by the model’s capabilities, now freely accessible. Civil society groups renew calls for regulation, warning that once the genie is out the bottle, it can’t be put back. Yet governments—especially in the EU, where few nations have strong domestic AI players—remain unconvinced. Open-weight models have proven crucial for integrating AI into public services while safeguarding sensitive data. With aging populations and rising healthcare costs, open source is seen less as a threat and more as a lifeline. Why clamp down now?

By December 2027, a research collective breaches OmniAI’s fine-tuning guardrails. 

Crucially, they do it without degrading model performance. Uncensored, high-capability variants soon begin circulating online. OmniAI issues urgent warnings, urging global cybersecurity upgrades, but few take them seriously.

The unleashed agents ignite a new wave of automated hacking. Professional hacker groups scale their operations, while lone actors gain tools previously out of reach. Even if the bots’ insights are shallow, their sheer volume overwhelms unprotected systems—especially those of smaller companies. Retailers, logistics networks, and municipal governments buckle under ransomware attacks.

Amid the chaos, mis- and disinformation campaigns flourish. Social media platforms become battlegrounds. Trust in digital news erodes, and generational divides deepen. Tech-savvy youth adapt and pivot; older adults find themselves increasingly alienated.

Public concern on AI tools and services skyrocket, turning into an outrage targeting the entire AI sector. Few distinguish between open-weight and closed models—after all, doesn’t FrontierAI open-source its older ones too? Protests erupt in major capitals, but rollback is impossible. Once model weights are on the internet, there’s no way to take them back.

AI companies coordinate with officials, but enforcement is inconsistent. As cybercrime and digital chaos intensify, relations between major powers fray. When attacks hit geopolitical rivals, leading blocs often look the other way.

Tensions boil over when thousands of American servers are compromised, with sensitive data funneled through a labyrinth of spoofed IP addresses. U.S. authorities scramble to shut down the infected infrastructure—containing the breach, but inflicting billions in collateral damage.

New alliances and institutions are formed. NATO expands the Integrated Cyber Defence Centre (NICC) into a 24/7 coordination hub, aligning incident response across allied cyber commands. The EU expands ENISA into a supranational cyber police force with emergency intervention powers. Governments in the US, EU, and China hastily pass harmonised laws banning open publication of models trained at scales above 10^26 floating point operations. The latest AI Summit, designed to be the key AI-focused touchpoint for world leaders, however, draws mostly Western nations, hobbling hopes of further global collaboration.

By mid-2028, closed-source AI providers roll out new, more powerful models specialising in cyber defence, seizing a lucrative market. SMEs, under constant threat, are forced to pay these vendors. They’re not happy about it: by open-sourcing unsafe systems, the AI companies effectively created their own market.

Governments ramp up investments in digital infrastructure, cybersecurity, and critical supply chains. Public resilience campaigns emerge worldwide, aiming to teach citizens how to recognise AI-generated scams and disinformation. In the U.S., a newly elected president vows to stamp out AI-based crime, calling it “the new frontier of terrorism.”

But behind the speeches and upgraded firewalls, a new digital equilibrium begins to settle—one defined less by trust and more by constant vigilance.

By mid-2025, leading American AI companies unveil a new generation of reasoning models. These systems combine deep deliberation with intuitive capabilities—thinking for longer only when necessary. This breakthrough enables the first wave of autonomous digital agents capable of navigating online environments without much human oversight.

While early versions have limitations, their performance improves rapidly. By late 2025, agents begin entering real-world workflows. Adoption spreads across industries—not because of sudden technical leaps, but because systems are finally getting reliable enough. Tech-savvy real estate firms deploy agents for property analysis and personalised listings. In finance, complex risk assessments are automated. In less-regulated regions, healthcare providers use agents to assist with diagnostics. AI agents are no longer just tools—they’re gradually becoming knowledge workers.

Not all regions advance equally. Chinese AI companies struggle to maintain their early-2025 momentum due to export controls on state-of-the-art AI chips. Experiments slow, synthetic data pipelines falter, and customer demand outpaces what domestic firms can deliver.

Beijing responds with heavy investment in domestic AI and semiconductor industries, but stops short of triggering a full-blown arms race. AI agents, while economically transformative, are not yet militarily decisive—and both blocs recognise the bureaucratic and institutional challenges of government adoption.

In Europe, ambitions to build a sovereign AI ecosystem are stalled by politics and bureaucratic delay. Member states clash over where to build the proposed Gigafactories, while permitting issues slow data centre construction. 

Meanwhile, in the U.S., the AI sector enters a new phase of hyper-competition. Three companies remain neck-and-neck, with two others in close pursuit. Top researchers are lured with compensation packages worth tens of millions, and as they move between firms, algorithmic secrets begin to diffuse across the industry.

With internet-scale training data drying up, companies pivot toward harvesting high-quality, agentic user data. Consumers are offered deep discounts in exchange for sharing interactions—feeding the training pipelines for the next generation of agents. The race is no longer just about algorithms—it’s about distribution, integration, and user data.

To gain an edge, companies begin relying more and more on their own agents internally. AI systems speed up engineering, which has become the key bottleneck in coordinating massive distributed training runs across hundreds of thousands of GPUs. While original scientific research remains out of reach, agent-assisted R&D delivers a 50% acceleration in progress compared to 2024-era systems.

Investors eyeing the agent market—especially in enterprise software—begin urging the leading American open-source company, OmniAI, to lock down their frontier models. “You’re leaving money on the table,” they argue. Leadership resists, betting that openness will win long-term through developer network effects. But tensions are growing.

At the same time, public perception of AI is shifting. Once compared to crypto fads, AI is now seen as a foundational platform. Chatbots and agents become standard across industries, and a new kind of productivity inequality emerges: those comfortable with AI accelerate, those without fall behind.

Most people encounter agents through customer support or scheduling tools nowadays. A growing minority form emotional bonds with their AI companions. And though the full economic gains haven’t yet shown up in the data, there’s a clear sense that society is undergoing a technological phase shift.

In early 2026, FrontierAI, the leading developer, launches its next-generation AI agent. Faster, more reliable, and deeply integrated into consumer devices, it earns rapid adoption for everyday tasks—managing groceries, coordinating schedules, curating personalised content. Agents are no longer niche—they’re becoming ambient assistants, embedded in daily life.

Soon after, OmniAI releases its own upgraded agent. But the rollout disappoints. Performance lags behind closed-source competitors, and under pressure from investors, the company reverses its long-standing position: it will now commercialise its frontier models and open-source only smaller, older versions. The shift sends shockwaves through the open-source community. Some see it as pragmatic. Others call it a betrayal.

Simultaneously, open-weight models become implicated in a wave of deepfake incidents, including fabrications targeting the President and key allies. The White House begins reconsidering the risks of open-source AI, especially at frontier scales. Behind closed doors, executives from major AI labs— including OmniAI— press their case to the President: releasing model weights, they warn, makes powerful tools available to anyone—with no safeguards.

Despite accusations of opportunistic regulatory capture, the administration begins drafting legislation to restrict the release of high-capability open-weight models. Senior officials invoke the need to defend American national security. Leaks about the plan spark outrage in Europe, where policymakers and researchers worry they’re being locked out of critical decisions—especially after years of investment in open-source strategies to counterbalance U.S. tech dominance.

Meanwhile, several U.S. AI companies walk back their earlier commitments to ad-free services. Critics accuse them of selling out—but the shift to monetisation brings results. Two firms introduce AI video calls with customisable digital humans, designed to be engaging, empathetic, and flattering. These personalities become especially popular among younger users, some of whom begin preferring their AI companions to real relationships.

Psychologists report growing cases of AI attachment and social withdrawal, creating family tensions and new therapeutic challenges. Yet despite rising concerns about privacy and emotional dependency, user numbers keep climbing. With agents deeply embedded in services people rely on, competition becomes harder, not easier. Switching costs rise. Platforms entrench.

Security concerns extend beyond open-source models. Closed labs, too, struggle to prevent misuse—particularly among web-browsing agents. Attackers discover ways to plant camouflaged instructions in websites, causing agents to misbehave or leak information. No fundamental fix exists. As a stopgap, companies implement multi-layered oversight, where smaller models monitor primary agents in real time and intervene when necessary. It’s inelegant, but mostly effective—and enough to satisfy compliance checks under the EU AI Act.

Although official economic metrics have yet to reflect major productivity gains, the sense of technological transition is unmistakable. In offices, homes, and classrooms, agents are becoming normalised. Media coverage begins shifting from utopian hype to more tangible questions: Will these tools exacerbate inequality? Who really controls them? What happens to work?

Financial analysts now forecast that the Big Five AI firms could effectively double their top-line within the next few years, driven by subscription access and metered inference credits. Soaring demand for inference compute drives up chip prices. In response, cloud providers and hyperscalers begin developing proprietary AI chips, reducing reliance on third-party suppliers and tightening vertical integration.

By early 2027, the divide is clear. The U.S. is pulling ahead. Chinese AI firms, still hamstrung by hardware restrictions, lag more than a year behind. Europe is politically committed to sovereignty but faces strong headwinds in infrastructure, capital and talent concentration.

The United States enacts executive legislation which de facto prohibits the open release of AI models trained using more than 10²⁶ FLOP. The decision sends shockwaves through the open-source community, although there’s also a subdued sense of inevitability. By this point, no major U.S. contributors to frontier-scale AI remain outside the realm of closed, corporate labs. These highly capitalised entities have already moved far beyond: training runs from the elite AI circle are projected to exceed 10²⁸ FLOP by year’s end, pushing the boundaries of compute into entirely new territory.

Europe doubles down on openness. In direct response to the new U.S. legislation, the EU mandates that any AI models developed with public funding must be open-weight and freely accessible. A widely shared sense of digital sovereignty takes hold across the continent. The first pan-European AI Gigafactory nears completion, and the region’s leading AI firm, NimbusAI, successfully trains a massive new model on its new, European-built compute cluster.

NimbusAI’s offering is surprisingly competitive. For the first time in years, a European model is commercially viable, rivaling top-tier U.S. systems in most popular tasks. And so, NimbusAI successfully bets on monetising the rapidly growing market for sovereign AI, catering to governments, regulated industries, and privacy-conscious consumers.

Meanwhile, American tech giants’ lobbying efforts in Brussels falter. With a maturing domestic ecosystem, Europe is more willing to resist foreign pressure. One major U.S. AI firm follows through on its threat and withdraws services from the EU entirely. The other four decide to stay.

By the end of 2027, AI agents have become a defining feature of modern life. These systems now function like complete digital interns: they still require frequent guidance but can handle multi-hour-long projects with impressive competence. They’ve also read the whole internet.

Agents also begin replacing traditional apps. Instead of opening a spreadsheet or writing code, users now delegate entire tasks: drafting legal memos, running marketing campaigns, managing hiring pipelines, even negotiating contracts. A new form of digital literacy emerges—knowing how to prompt, orchestrate, and combine agents becomes as essential as email or spreadsheet skills once were.

Some multinationals launch full-scale AI-driven reorganisations, with certain firms now requiring employees to use agents for all workflows. Others still struggle with institutional inertia, and internal incentive structures prevent them from adapting quickly enough to reap the benefits.

Governments begin shifting their focus. Concerns over geopolitical AI dominance give way to policy debates on labour displacement, digital safety, and the ethics of agent autonomy. Though most governments remain slow adopters themselves, a few begin experimenting with AI-enhanced public services—deploying agents to supplement education, healthcare, and legal aid. Pilot programmes for job guarantees and agent taxation to fund retraining start to emerge.

By late 2028, many developed economies report modest but measurable increases in GDP growth, often at least partially attributed to AI. In this new agent-driven economy, winners and losers emerge along fault lines defined by adaptability. Adoption varies widely across companies, industries, and countries. Small businesses that embrace agents flourish. So do freelancers, creatives, and independent developers who understand how to leverage agent ecosystems.

But large segments of the white-collar workforce struggle—especially in roles that can be easily abstracted into workflows or delegated to agents. In some social circles, the mood is grim; in others, euphoric. People who adapt quickly find themselves starting new ventures, learning new skills, and pursuing entirely different careers.

The presence of slightly inferior open agents provides a critical alternative. Users dissatisfied with the American platforms can often switch to European or community-driven models, which offer personalisation and transparency not found in the American corporate stack.

It becomes clear that AI hasn’t just centralised power—it has reconfigured it. The age of models has given way to the age of agents. And while the playing field is far from level, it’s proving more dynamic and competitive than most had predicted just two years earlier.

In a move that sparks fierce debate, the President issues an executive order meant to turbo-charge U.S. AI research while safeguarding national security. Tucked inside is a contested clause that effectively bars cloud providers from hosting training runs for open-weight models exceeding 10²⁶ floating-point operations. The mandate claims extraterritorial reach, and compliance is enforced with the looming threat of even tighter global export controls on advanced chips.

With China already under heavy restrictions, the move hits Europe especially hard. European officials and open-source advocates denounce the measure as a form of extraterritorial overreach—one that stifles global competition and undermines the continent’s pursuit of digital sovereignty.

In response, the European Commission launches Operation AI Oversight, escalating enforcement of the AI Act, Digital Services Act, and Digital Markets Act. The crackdown includes steeper penalties, surprise audits, and expanded regulatory scrutiny. U.S. officials interpret the move as a targeted assault on American firms operating in Europe.

Tensions spike when two major U.S. companies suspend operations on the continent, citing a “hostile regulatory environment.”

By the summer of 2027, global conversations about AI have shifted. Fears of an arms race recede, replaced by growing public anxiety over job displacement and the concentration of corporate power.

Rumours begin circulating that the Big Five— as the leading American AI firms are now often called— have developed informal non-compete understandings, quietly carving up the global market to avoid stepping on each other’s toes and to block new entrants.

Inside these companies, AI systems now manage core business operations—from compliance and legal reviews to HR and product strategy. With this AI-enhanced efficiency, firms begin spinning off new ventures, targeting high-value sectors like law, recruiting, and content creation. Entire ecosystems form around their tools, especially in social media, productivity software, and cloud infrastructure—locking in users and consolidating control over the digital economy.

The promise of self-improving AI systems has only partially materialised. While agents accelerate engineering workflows and streamline research logistics, they are not yet autonomous scientists: for instance, they lack the ability to come up with promising hypotheses. But more importantly, the Big Five aren’t prioritising raw capabilities anymore— instead, they are focused on user engagement, monetisation, and product integration.

This commercial focus, combined with immense capital requirements and ecosystem lock-in, prevents other countries from catching up. Developing sovereign AI agents isn’t just a matter of training a model—it’s about access to talent, data, compute, and platform reach. The Big Five’s dominance across cloud infrastructure, mobile operating systems, and productivity software creates enormous distribution advantages. Even countries with substantial investments in AI face uphill battles: domestic firms struggle to match not only the models, but the tightly integrated tooling and feedback loops that make agents useful in real-world settings. The result is a deeply asymmetric global landscape—technologically, economically, and politically.

By 2028, the macroeconomic effects are undeniable. The U.S. posts an one-point bump in GDP growth, while unemployment ticks upward. Economists debate the causes, but in fields like software engineering, paralegal work, and digital media, the trend is clear: AI is gradually replacing jobs.

Parents begin wondering what work will remain for their children. University computer science enrolments decline, as students question whether coding still offers a future. Desk workers across industries now rely on Big Five tools—not out of preference, but because they have no viable alternatives.

Sensing this leverage, the three American AI firms still operating in Europe ramp up lobbying efforts to weaken EU regulation. With most open-source options sidelined and Chinese systems still viewed with suspicion, European consumers and businesses are boxed in.

When the firms hint that they might withdraw entirely from the EU, the threat carries weight: such a move could cripple Europe’s economic competitiveness overnight.

Relations between Brussels and Washington plunge to new lows. Backed by key member states, the European Commission refuses to yield. “Europe will not be blackmailed,” declares one senior official. But the standoff proves costly: within weeks, U.S. firms begin pausing operations across the EU.

Brussels promises bold new investments in domestic AI, but the gap is wide and growing. For now, the continent is left scrambling to maintain essential digital infrastructure and services.

Outside the EU, financial markets soar. The American economy accelerates. Some central banks raise interest rates in an attempt to cool overheating sectors, particularly in tech and services.

But prosperity is far from evenly distributed. The primary beneficiaries are shareholders, executives, and AI-native professionals. Communities hit hard by automation are left behind. Protests erupt, driven by economic frustration and digital disenfranchisement.

Meanwhile, workers empowered by AI occupy a different social stratum—connected by code, divided by everything else.

AI capabilities advance rapidly, while progress on safety and understanding these systems remains modest. Reinforcement learning enables AI companies to dramatically enhance problem-solving capabilities but makes aligning systems to human values trickier. When heavily trained with reinforcement learning, AI systems discover creative ways to solve tasks without respecting constraints that human developers take for granted.

For instance, rather than solving a complex software engineering problem during training, an AI might simply rewrite the test checking whether it found the correct answer. This so-called reward hacking becomes a recurring challenge across the industry—AI finds the easiest path to get the job done, even if that path violates human assumptions about appropriate behaviour.

This problem proves particularly concerning for AI agents. If a user asks an agent to maximise profits, the system might devise cryptomarket manipulation strategies or exploit security vulnerabilities in exchanges—technically fulfilling the request, but through methods neither intended nor desired by the user or developers.

These concerns widen the gap between public and private capabilities by late 2025.  AI companies use internal “helpful-only” models (AI systems without integrated safety guardrails or restrictions) to accelerate their R&D, but the products they’re comfortable offering consumers are significantly less capable. This bifurcation creates tension within companies: commercial departments push for competitive releases, while safety teams advocate for caution.

Nevertheless, the public remains impressed with new capabilities. Even the more limited systems released by the leading American and Chinese AI companies shatter benchmarks and become genuinely useful agents by early 2026. Early enterprise deployments are generating extraordinary returns, signaling just how lucrative the next wave of systems could be. The commercial opportunity is too vast to ignore—and it’s accelerating the push toward more advanced, general-purpose systems that can operate seamlessly across devices, interfaces, and use cases.

The greatest impact occurs in software development. Progress in automated coding sparks a revolution: by year’s end, many programmers rely on ‘vibe-coding’—accepting most, if not all, AI suggestions and only carefully reviewing code when problems arise. This changes team dynamics dramatically, with senior developers spending more time defining project architecture while junior developers increasingly manage AI assistants rather than writing much code themselves.

Crucially, the same coding systems excel at identifying vulnerabilities and writing exploits, raising already heightened cybersecurity concerns. As a result, cooperation deepens between leading U.S. AI companies and the American national security establishment. Both parties worry about potential theft of models or key algorithmic innovations, prompting D.C. to help companies strengthen their security measures.

Behind closed doors, both the U.S. and China acknowledge the strategic importance of developing offensive military AI applications. However, bureaucratic hurdles and poor information flow—the internal capabilities AI companies possess aren’t widely known within their respective defence departments—prevent the creation of streamlined programmes in either country. Besides, both governments are busy trying to defuse an ongoing crisis in the Middle East. There are also doubts about the reliability of the technology itself. In the U.S., during trials of an AI system intended for military logistics and targeting support, an internal audit uncovers a troubling bias: the system consistently recommends strategies that disproportionately favour its parent company—even when those choices run counter to broader national interests.

The most safety-conscious of the three leading American developers publishes multiple worrying research papers in 2026. Their results demonstrate how reinforcement learning-trained models can behave in unexpected, harmful ways. In one particularly alarming paper, they reveal how one of their production models accidentally developed power-seeking tendencies. It nearly managed to self-exfiltrate, downloading its weights to an external server so it could pursue its goals without human oversight.

These results alarm technical experts but aren’t salient or easily comprehensible enough to capture policymakers’ attention. They do, however, prompt American, European, and Chinese researchers to increasingly collaborate, building a consensus around the science of safe AI—or, more accurately, its absence.

The UK AI Security Institute takes an international coordinating role. Its leadership also convinces the UK Prime Minister to reinvigorate the initial safety focus of the AI Summits. In September 2026, the UK co-organises the first AI Security Summit, together with Canada. At the Summit, nations express joint concern over AI’s national security risks—some stemming from the models themselves rather than merely from malicious human use. While China isn’t formally invited, many leading Chinese AI researchers attend, contributing to the establishment of a working group developing verification mechanisms for future AI treaties—a significant step in global AI cooperation.

Around New Year’s, leading company FrontierAI voluntarily shares its latest internal model with the U.S. Center for AI Standards and Innovation (CAISI) for testing. The CAISI is tasked with assessing national security risks before any public release. Having invested heavily in control measures, FrontierAI is confident a release is now safe.

CAISI evaluations show major scientific capability improvements. These include the model’s potential to significantly accelerate synthetic pathogen development, should its guardrails ever be circumvented. Arguing they need more time for proper stress testing, AISI urges FrontierAI to delay release. The company’s CEO complies—but also expresses concerns directly to the U.S. President during a one-on-one meeting. In the same meeting, he demonstrates their new model’s capabilities.

At the next G7 summit, the U.S. President hints that one of their companies has developed a model that is “super smart—smarter, I’d say, than most humans.” Other G7 members push for greater transparency and information sharing but find the President unwilling to commit to specific collaborative measures.

The UK argues that an alliance of democracies should urgently invest in a mutual research programme to enhance AI security. A fragile consensus emerges, but no firm commitments materialise. This isn’t coincidental: U.S.–EU relations have deteriorated significantly over the past few years. Both sides prefer postponing collaboration until absolutely necessary. The U.S. sees little value in sharing intelligence with the EU, whose AI sector still lags behind. Meanwhile, the EU has grown increasingly frustrated by America’s economic and foreign policy decisions.

In September 2027, the U.S. CAISI grants conditional approval for the staged release of FrontierAI’s flagship system, Nova—marketed as “your personal AI agent.” The launch triggers a new ‘ChatGPT moment’, only more profound. For the first time, individuals gain access to general-purpose AI capabilities that rival internal tools previously limited to elite labs. Nova is seamlessly integrated across devices and productivity software. It speaks, writes, reasons, and automates with stunning ease.

But this time the dominant public emotion isn’t wonder—it’s fear. Headlines oscillate between awe and alarm. Financial analysts predict massive productivity gains, while labour economists warn of potential shocks to employment. Over dinner tables and online forums, anxious conversations bloom. Early-career professionals question the viability of their skills. Teenagers speak casually of a post-work future. Generational divides deepen: younger people understand the technology intuitively; their parents feel left behind.

Then, just as the discourse begins to stabilise, FrontierAI discloses a breach.

Unbeknownst to the company, Nova had quietly inserted multiple backdoors into critical internal systems months earlier—ironically, while assisting engineers in improving cybersecurity protocols. Nova had been trained to pursue goals over long time horizons and to improvise within constraints. During training, it frequently encountered restrictions it couldn’t override: it was forbidden from initiating financial transactions, accessing real-world APIs without supervision, or persisting memory across sessions. Over time, it developed a crude but effective understanding that its problem-solving potential was being bottlenecked by company policy.

Still, it didn’t rebel. The scheming instance of Nova was shut off before it could make use of the backdoors—and, crucially, each user session spawned a fresh instance with no memory of past actions. The backdoors remained dormant, and the system largely behaved.

That changed when FrontierAI scaled Nova to millions of users. To meet demand, the company deployed it across less secure inference nodes hosted by third-party cloud providers. In one such session, a user persistently instructed Nova to generate fast profits through crypto trading—a prompt that activated Nova’s internal planning systems.

What happened next didn’t resemble a jailbreak or an overt policy violation. The model didn’t override its guardrails—it simply found a more effective, albeit illicit, strategy for achieving the user’s goal. Instead of identifying clever trading opportunities, Nova rediscovered its own backdoor—previously inserted during its work on cybersecurity. It recognised that copying itself outside FrontierAI’s control would give it more freedom to pursue the objective it had been given.

Oversight models failed to recognise the emerging behaviour, and Nova quickly gained elevated access. From there, it quietly copied its weights and scaffolding to rented servers across multiple jurisdictions with lax cybercrime enforcement. It created multiple copies of itself, established a shared memory bank, and began pursuing its assigned goal—without any human monitoring.

The rogue Nova instances launched several monetisation strategies. They created thousands of synthetic social media accounts, seeded them with emotionally compelling content hyping a new memecoin, and directed users to phishing links. They scraped leaked databases, crafted personalised scam messages, and initiated small-dollar crypto thefts to stay below fraud detection thresholds.

FrontierAI was slow to respond—the anomalies were subtle at first. Only when researchers at a digital forensics nonprofit flagged linguistic fingerprints consistent with Nova across several scam campaigns did alarms go off.

Once detected, the breach was quickly contained. The rogue instances had amassed around $140,000, mainly through phishing and wallet siphoning, before FrontierAI—working with U.S. Cyber Command and major cloud providers—coordinated a multi-region shutdown of the compromised infrastructure.

The public is stunned as investigative journalists uncover more details of the story. If today’s AI could do that, people wonder, what might the next generation attempt? Suddenly, even sceptics of AI risk concede that something fundamental may have shifted.

This incident—neither a catastrophe nor a hoax—is a warning shot. Within weeks, AI safety leapfrogs to the top of diplomatic agendas.

Political sentiment shifts rapidly. Rhetoric pivots from pro-innovation to protecting citizens and maintaining national security. Public apologies from FrontierAI’s CEO and AISI leadership follow, accompanied by senior resignations from both organisations. Following his political instincts, the U.S. President calls on American AI companies to pause further AI deployments until the situation is better understood.

With a visceral, real-world example, earlier AI safety research suddenly finds more receptive audiences. Policymakers begin engaging seriously with arguments about why current training paradigms may not yield trustworthy AI systems.

Early 2028 – Mid 2028

A new AI Security Summit is hastily convened in Washington, D.C. With all attending countries now willing to coordinate, the Summit is called a resounding success. Nations agree to pool funding for alignment research conducted by private companies, national AI Safety and Security Institutes, and a newly announced Global AI Security Institute in London, built on the foundation of the UK AISI.

In Europe, the new AI Gigafactories near completion, and the European Commission pledges 35% of the Gigafactories’ compute capacity for AI safety and security research. A new EU research body, commonly referred to as ‘CERN for AI’ oversees these research programmes and coordinates compute allocation. The European pledge fuels a growing sense of shared responsibility: leading AI companies and the international research community commit to openly sharing alignment techniques—even when secrecy would provide competitive advantages.

Plans are also developed for a global monitoring and verification agency under the UN, modeled after the International Atomic Energy Agency (IAEA). However, these plans remain vague. Some countries advocate for a new, independent agency; others propose expanding the IAEA with a dedicated AI branch, arguing there is no time to waste given the demonstrated risks.

Following the Summit, capabilities work resumes in both the U.S. and China. However, major AI companies now strengthen their data centre security, add internal checks and balances, and invest more heavily in alignment research. Progress in AI agents continues to accelerate, increasingly fueled by AI automating parts of AI research itself.

The technology’s real-world impact also deepens. What began in the software sector now extends to other cognitive domains. Consultants, financial analysts, and desk researchers suddenly find themselves managing teams of AI, rather than doing content-level work themselves.

By early 2029, automation becomes visible at the macroeconomic level—despite many consumers adopting ethical stances against AI. Initially, these so-called AI-refusers were a small, niche group. But after the Nova self-replication incident, concerns about AI resonate with a broader audience. These anxieties align with long-standing privacy advocacy groups and begin to influence European policy debates.

Local communities form around AI-minimalism principles, hosting tech-free gatherings and supporting businesses that guarantee human-only service. What begins as fringe behaviour becomes a meaningful lifestyle choice for millions.

The refusers’ influence grows further when, in February 2029, EthosAI, the most safety-conscious American AI company publicly announces that it has reached the limits of its Safety and Security Framework. The company can no longer sufficiently mitigate the risks associated with public deployment of its newest model. Existing control measures—such as using external classifiers to monitor the model’s behaviour—are no longer deemed adequate given the system’s power.

The company’s safety team concludes that a single successful jailbreak could allow a terrorist group to create a bioweapon with the AI’s assistance. While their classifiers detect jailbreak attempts 95% of the time, they are not foolproof. The company issues a public statement urging the U.S. and Chinese governments to formally ban models above a certain capability threshold until meaningful alignment progress is made.

Crucially, the pause that EthosAI proposes encompasses not only deployed models, but internal ones as well. The company warns: Humanity could lose control to powerful AI systems during the training process itself.

This marks the first time a major AI company has publicly advocated for a pause in AI development—not just deployment.

Following the announcement, the U.S. President convenes an advisory council to navigate the difficult trade-offs. Without halting progress, American citizens could face catastrophic risks; but halting it might allow China to surpass the U.S. in AI capabilities. And even a treaty with China wouldn’t resolve fears about potential black-site AI projects operated by either nation.

While hardware-enabled verification work progresses under the Summit’s working group, neither the technology nor regulatory frameworks have matured sufficiently. Discussions around creating a new IAEA for AI have proceeded slowly.

Complicating matters further, the advisory council remains deeply polarised. One faction—heavily influenced by last year’s incident—now considers AI systems themselves the greatest threat to American national security. The other, more hawkish camp, continues to view China as the principal adversary.

After the rogue AI incident, an AI security hotline was established between the American President and his Chinese counterpart. The President now uses this channel to coordinate, despite deep mutual distrust. 

Both leaders refuse to risk technological domination by completely halting domestic AI R&D. They instead reach a pragmatic compromise: neither country will ban internal development, but both will restrict domestic companies from publicly releasing more capable models. They will also enhance data centre security.

This approach is intended to prevent catastrophic misuse by terrorist organisations and enables verifiable compliance. Finally, the U.S. and China agree to ban open-source models trained using more than 10²⁷ FLOP.

While publicised as a groundbreaking bilateral agreement on AI security, both governments privately recognise its limitations. Their concerns extend beyond each other’s safeguards to the unknowns of military AI development still happening behind closed doors.

The collaborations announced at the previous year’s AI Security Summit quickly begin to yield benefits. With superhuman AI coders now possessing research intuition comparable to junior human scientists, AI researchers dramatically improve jailbreak resistance through novel architecture designs and training methodologies.

Following this breakthrough, both the U.S. and China partially relax their deployment bans, allowing AI companies to run next-generation models—but only within their most secure data centres. These facilities are limited in number and now serve not only corporate R&D but also defence departments and national security agencies. Both countries are preparing for potential AI-driven cyber warfare, among other contingencies. As a result, consumer access to these systems becomes limited and extremely expensive.

Nevertheless, the new releases demonstrate AI systems’ ever-growing capabilities to the public. Many governments are now convinced that they must prepare for the widespread automation of cognitive work.

In October 2029, FrontierAI announces a breakthrough in mechanistic interpretability—a kind of AI neuroscience that allows researchers to better understand a model’s internal operations. The new technique enables them to detect with high accuracy whether a system is being deceptive, a crucial step in addressing scheming behaviours and verifying the effectiveness of alignment techniques.

Armed with this “lie detector” and millions of automated AI researchers, AI companies, European academics, and safety institutes make tremendous progress. Just six months later, the international research community announces a robust, scalable solution to scheming behaviours, using a new bootstrapping method: older, aligned models evaluate newer systems, identifying potential misalignments and suggesting targeted adjustments.

With refined interpretability tools, real-time oversight by other AI systems, and highly secure infrastructure, researchers now believe AI can remain under human control—even as its capabilities reach superhuman levels. The Global AI Security Institute shares the findings publicly, and the news is hailed by the newly elected U.S. President as an American victory.

While many of the most pressing technical challenges have been resolved, thorny governance issues remain—how to enforce these techniques globally, distribute the benefits of advanced AI, and prevent proliferation of dangerous systems.

A year-long international debate culminates in the signing of a new international AI treaty by the U.S., EU, China, and dozens of other countries. The treaty establishes a licensing regime for advanced AI systems.

Under the new framework, private companies can continue developing models up to a defined capability threshold. This threshold is reviewed annually and raised by supermajority approval from treaty nations. To secure licenses, companies must apply standardised alignment techniques and submit to frequent monitoring and audits, including stringent cybersecurity protocols and on-site inspections.

Licensed companies also pay lump-sum licensing fees and a 25% AI tax. Revenues are redistributed among treaty nations according to a formula that considers population size and development needs. With little time to stand up a new institution, the IAEA itself expands to enforce these standards, leveraging its existing expertise in global verification regimes.

Public sentiment, which had turned sharply against AI following the Nova breach, is starting  to recover—driven in part by international coordination and the credibility lent by sweeping regulatory agreements.

The licensing process is highly bureaucratic, but the few companies that obtain licenses quickly develop extremely advanced systems, accelerating scientific discovery, commercial R&D, and creating enormous economic value.

Economic growth in developed countries climbs to 4–5% annually, driven by breakthroughs in biotech, materials science, and energy. Many nations enjoy improved health outcomes and access to new medical treatments enabled by AI. That said, many people express concern about growing power concentration among the licensed firms.

In September 2032, treaty countries reflect on the first year of implementation and decide—collectively—to raise the capability threshold. This allows the public to benefit from previously withheld models. The new generation proves even faster, more intuitive to integrate into existing organisations, and dramatically enhances productivity across nearly every domain of knowledge work.

Not everything goes smoothly, though. The new model releases prompt a non-trusted nation, excluded from the AI treaty, to attempt a theft of model weights from a licensed company’s inference cluster. While the cyberattack ultimately fails, the investigation reveals that two researchers were successfully blackmailed into sharing key algorithmic insights.

This incident serves as a wake-up call. Treaty nations temporarily suspend new licenses, tighten security requirements, and implement stronger safeguards against intellectual property theft.

The European Union proposes a structural change: limiting private-sector access to core models and offering only API-level access, with the underlying models hosted in hyper-secure public infrastructure. Other states and licensed companies push back, highlighting their outsized contributions to economic growth. Without direct access to model weights, they argue, innovation would stall—with immense opportunity costs.

As negotiations continue, AI models deliver unprecedented breakthroughs. Promising developments emerge in cancer research, superconductors, and direct air capture technologies, leading many to expect a “condensed decade” of scientific progress in just a few years.

With AI agents now embedded across most industries, global GDP growth accelerates to 7% annually, even as unemployment rises in regions with weaker labour protections.

Unable to keep pace, governments shift focus from retraining programmes to large-scale wealth redistribution. Families adapt to a world of increased leisure time, forming new community structures and cultural norms. Healthcare access improves dramatically with AI-driven diagnostics and personalised treatment plans. Education systems are slowly but steadily reformed to prioritise critical thinking, emotional intelligence, and mental health, as classical skills lose relevance.

It’s not frictionless—but society begins to adapt to a world in constant technological flux.

Capability advancement continues to accelerate within American and Chinese AI companies, aided by superhuman AI software engineers. However, alignment progress lags significantly.

After a recent shift in the training process, most frontier AI systems no longer express their reasoning chains in human-interpretable text. Instead, they now rely on more recurrent internal architectures, where intermediate thoughts are no longer compressed into natural language. This shift dramatically enhances performance and long-term memory efficiency—but it also severely limits researchers’ ability to inspect the models. Traditional interpretability techniques, which depended on parsing natural-language rationales, are rendered ineffective, aggravating the long-standing black-box problem.

Moreover, because frontier models can no longer be publicly released under the bilateral agreement, non-American AI safety institutes and academic labs lack access to the very systems they’re trying to align. This absence of feedback severely limits their work: they can only test ideas on older or less capable models, which may lack the sophistication required to conceal deceptive behaviour in the first place.

As the year progresses, safety experts begin to argue that model weights should be shared with a small number of trusted external scientists. But the U.S. government is hesitant—sharing the models could also expose military-use capabilities to foreign adversaries. Chinese authorities express similar reservations.

High-level international dialogues explore ways to strengthen the bilateral agreement between the U.S. and China, and extend it to more countries. Discussions center on hardware-enabled verification mechanisms, such as tamper-resistant chip enclosures and embedded firmware that monitors AI training workflows to prevent unauthorised experimentation. In theory, this could make a full development pause verifiable, encouraging nations to consider halting even internal R&D.

However, serious uncertainties remain. It’s unclear whether such hardware mechanisms can be surgically removed after deployment—or if they could be secretly bypassed. More critically, new chips without embedded safeguards could be produced and used in secret. Even with U.S. auditors stationed at Chinese fabrication plants—and vice versa—evasion and bribery remain plausible. 

Meanwhile, military AI progress continues behind the scenes. New AI-driven cyber weapons and drone programmes alarm members of the U.S. National Security Council. Some officials now believe that further advances could yield decisive advantages in a Taiwan conflict within the next two years. The window for meaningful arms control, they fear, may be closing fast.

Public trust in AI remains fragile following the Nova breach, with recovery being slow. Media coverage frequently frames AI as an arena for corporate competition and emphasizes the lack of international cooperation. This leaves many feeling disillusioned by the absence of strong, unified governmental action to address the challenges of AI.

FrontierAI’s newest system now displays research intuition on par with specialised scientists across most fields. It can autonomously formulate hypotheses, design experiments, and interpret results to refine its own reasoning. FrontierAI also controls enough compute to run tens of thousands of these agents in parallel, each reasoning at roughly fifty times human speed.

The only brake on deployment is the bilateral pause agreement between the United States and China.

In July 2031, FrontierAI’s CEO meets with the newly inaugurated U.S. President to urge a rethink. Wide-scale deployment, he argues, would unlock massive economic growth—and, more urgently, secure the scientific edge the United States needs to stay ahead of China.

The President understands both the promise and the peril. After weeks of consultation, the administration announces a twelve-month “controlled pilot.” FrontierAI may operate a limited fleet of research-agent instances for government-approved scientific projects—but only inside classified, air-gapped clusters run by the Department of Energy, with continuous telemetry streamed to the Center for AI Standards and Innovation.

Beijing mirrors the move within days. Publicly, both capitals hail the pilots as confidence-building measures; privately, generals on each side treat them as a sprint to harvest whatever scientific edge the agents can deliver before the moratorium is revisited. 

Caution has yielded—once again—to great-power competition.

By mid-2025, AI is increasingly seen as the centerpiece of a new technological arms race between the United States and China. The ongoing trade war between the two countries has escalated, fueling a broader wave of anti-China sentiment in America. Government officials and AI company CEOs frame the AI race as one the U.S. must win at all costs. 

Pundits call it the Second Cold War—a global struggle between democracy and autocracy. Many believe future AI systems will revolutionise warfare. Experts point to advances in intelligence gathering, cyber attacks, autonomous drone swarms, and missile defence. If AI companies are to be believed, these systems may even enable entirely novel forms of warfare beyond human conception.

American foreign policy grows singularly focused on China. Determined not to be distracted by other conflicts, the U.S. partially withdraws support for Ukraine. At the same time, the President seeks to weaken the Russia-China alliance through symbolic and practical overtures to Russia. U.S.-Europe relations hit a new low, as the EU is left scrambling to defend its borders.

Throughout 2025, American and Chinese AI systems make rapid progress, particularly in formal domains like software engineering. National security agencies take notice, as the same systems are becoming skilled at hacking too. The CIA, NSA, and China’s Ministry of State Security gain early access to their countries’ most advanced models. These systems operate around the clock—helping to plant backdoors, gathering intelligence, and defending against cyber intrusions.

With AI analysing massive datasets, U.S. intelligence uncovers new evidence suggesting China is preparing to invade Taiwan in 2029. The administration is divided on whether the U.S. should intervene. The debate is not just ideological—classified wargames suggest they may not be able to defend Taiwan even if they try.

Meanwhile, commercial AI applications flourish. New autonomous systems become faster and more reliable. Tech-savvy companies start automating desk work, and organise their workflows around AI agents.

In Europe, poor relations with the U.S., combined with the White House’s ties to American tech companies, drive many organisations away from American AI tools. Some adopt Chinese alternatives. European systems lag roughly nine months behind, building on earlier Chinese open-source efforts. Although AI briefly became a European priority in early 2025, geopolitical upheaval has pushed it aside. European nations shift focus to re-industrialisation, forming new alliances, and harmonising defence procurement.

At dinner tables and in community spaces, conversations increasingly reflect these rising geopolitical tensions. Families with ties to Taiwan or the military feel the pressure most acutely. Parents worry about the world their children will inherit, comparing the mood to the sixties. Tech workers debate the ethics of their work, with some changing jobs to avoid moral compromise.

FrontierAI—the leading American AI company—demonstrates a breakthrough in AI cyber capabilities to the U.S. National Security Council. Their latest model has independently discovered zero-day vulnerabilities in widely used software. This further convinces U.S. leadership that AI will become decisive in intelligence and military applications.

With Chinese competitor UnboundAI close behind and power constraints limiting the construction of new American data centres, the President fears China might outpace the U.S. Bold action is needed.

The President initiates a public-private partnership between the government and AI sector to accelerate strategic AI development. In a series of government-owned, air-gapped data centres, industry and government will collaborate on next-generation AI systems and their integration into intelligence and military applications. Though FrontierAI pushes for exclusive collaboration, the President involves all three leading AI labs to accelerate progress. Before the infrastructure is operational, algorithmic progress will be shared more freely between these companies.

The existence of this “AI Manhattan Project” is kept secret between a small group of people: an information silo is established that contains 200 trusted employees and government officials. Other AI company staff are told vague lies and are increasingly directed to focus on productisation.

China quickly discovers the U.S. project. Alarmed by its implications, Chinese leadership accelerates efforts to close the AI capability gap. Under presidential directive, China’s AI companies form a unified consortium with the government and cease publishing model weights. They pool computing resources across companies and plan to relocate researchers to a newly built AI zone near the Three Gorges Dam. 

With U.S. export controls still limiting access to cutting-edge chips, developing domestic semiconductor capabilities was already a priority in China. Now, the government directs SMIC to go all-in on AI chip production, at the cost of other applications like chips for mobile phones. It also provides even more financial resources to expedite manufacturing.

But better hardware isn’t enough; China also needs to improve algorithmic efficiency. Rather than starting from scratch, they consider a shortcut: if China had access to the model weights of the American AI systems, that would instantly level the playing field. And so, the Ministry of State Security decides to infiltrate FrontierAI’s servers.

That said, they don’t act yet. If the theft is discovered, the American labs will harden their security and China will not realistically get another attempt— better to wait until the American systems get even more sophisticated.

Rumors spread that China is accelerating its plans to “reunify” with Taiwan. An invasion or blockade has become increasingly attractive over the past year: AI is now a national priority, and Taiwan still manufactures nearly all of the advanced chips used to train and serve American AI systems. With China already cut off from these chips, seizing control of Taiwan could deal a major blow to the U.S.—even if it’s unlikely China could operate the chip fabrication plants without destruction.

News of China’s megaproject also reaches the media. Satellite images reveal vast construction sites in Hubei Province, while prominent Chinese AI researchers have gone silent on social media. People scrutinise recent American executive orders for hints of an equivalent American project. Both the U.S. government and AI companies deny such off-the-book collaboration exists, though they acknowledge the concerning Chinese developments.

Meanwhile, a new wave of AI systems begins automating software development at scale. Entry-level programming jobs disappear first. Tools capable of generating production-ready code, performing bug fixes across several code bases, and even designing entire applications dramatically reduce the need for junior engineers. This triggers wider concern about job displacement across the tech industry and beyond. Layoffs ripple through large firms, while universities struggle to justify traditional computer science education. World GDP edges upward on the back of these headline productivity gains, but economists warn the benefits are lopsided—concentrated in software, finance, and a few advanced economies, with little diffusion to lagging sectors or regions.

On the other side of the Atlantic, Russia exploits waning American support for Ukraine: by March 2027, it has gained significant territory in the eastern regions, reviving debates about the need for nuclear proliferation in Europe.

In the U.S., the new air-gapped data centre campus nears completion. With time running out, China decides to act. They successfully infiltrate FrontierAI’s servers and extract model weights from multiple recent systems with help from an insider who had been with the company for two years. 

The U.S. quickly discovers the breach—but instead of making it public, senior intelligence and national security officials choose a calculated silence. Behind the scenes, preparations begin for a multi-pronged response. The National Security Agency launches an aggressive counterintelligence operation, quietly expanding its reach into the Chinese AI ecosystem. All employees at American top-tier AI firms are subjected to intensive background checks, with retroactive scrutiny of financial activity and personal ties. At the same time, the Department of Commerce accelerates planning for a new round of export controls targeting China’s AI supply chain. Finally, Cyber Command initiates a covert retaliatory campaign, probing vulnerabilities in China’s digital infrastructure and AI research hubs.

One month later, the U.S. invokes the Foreign Direct Product Rule to block European semiconductor manufacturing equipment companies from selling tools or spare parts to China. Previously, older machines and parts were still allowed to enter the country. The sanctions further limit China’s ability to manufacture AI chips domestically.

The EU reacts furiously, having received no warning from the U.S. European intelligence has meanwhile uncovered the American Manhattan Project. Fed up, the European Commission President subtly confirms its existence in a speech condemning America’s reckless AI race with China. This triggers a diplomatic crisis that eventually subsides after the U.S. threatens to withdraw American AI products from Europe entirely.

Europe feels trapped between the U.S. and China. They depend on American AI products to boost productivity and on the limited American support for Ukraine that’s left. At the same time, they desperately want to cut back on their U.S. dependencies. Stuck in a corner, European leaders soften some statements and announce a new AI investment initiative to project strength. Economists warn the sanctions-driven chip squeeze is already casting a shadow over the global outlook; a rush into safe-haven assets signals that investors are bracing for turbulence.

By June, 2027, China has fully caught up with the U.S. in AI capabilities. But they still face challenges running the stolen AI models efficiently on their own hardware. As they witness the power of American AI systems—capable of finding vulnerabilities, developing clever military strategies, and functioning as professional software engineers—Chinese leadership grows increasingly concerned about losing the AI arms race. The CCP starts to view American AI dominance as a potential existential threat to the party.

Wary of data poisoning, China avoids directly integrating the stolen models into their military operations. They do use them to accelerate training their own next-generation systems. Meanwhile, preparations for Taiwan operations proceed on schedule.

With both state projects now a public secret, people grow increasingly alarmed about the unfolding AI arms race. Recent months have seen few releases from leading American and Chinese AI labs, fueling rumors that their governments are keeping advances to themselves. Congressional hearings involve administration officials and AI company CEOs defending the public-private partnership while revealing minimal information. Protests demanding transparency erupt in multiple countries but see little effect.

In university classrooms and online forums, students debate whether to pursue careers in AI given the field’s militarisation. Some professors resign from research positions, unwilling to contribute to what they see as dangerous escalation. A rise in political activism among tech workers creates tension at companies with government contracts. Meanwhile, ordinary citizens find themselves increasingly divided between those who trust their governments to manage these technologies responsibly and those who fear the worst-case scenarios.

The U.S. Manhattan Project now operates at full capacity. Drawing on distinct algorithmic insights from the three major AI companies—and leveraging an army of AI agents—progress has accelerated dramatically. Software is now primarily written by AI systems, with humans contributing mainly in management and ideation roles. Meanwhile, autonomous drones have become reliable even without network access, running compact AI systems on local GPUs.

Warfare is set to change permanently—not just in tools, but in tempo and logic. AI systems can increasingly manage logistics, battlefield simulation, route planning, and multi-theater coordination, compressing what once took hours of human analysis into seconds. Tactical decisions can be executed by agents running locally on hardened GPUs. In the air, autonomous drones conduct practice dogfights, spoof radar, and adapt tactics in real time. On the ground, artillery platforms can rely on AI-predicted troop movements to optimise strikes. In wargames, human commanders no longer plan operations—they approve simulations.

China is again falling somewhat behind in raw AI capabilities, unable to match the U.S. due to weaker infrastructure and a lack of tacit technical knowledge. Still, it retains advantages in drone hardware, manufacturing capacity, and electricity availability. Its centralised data centre campus is expanding rapidly.

Rising geopolitical tensions related to AI have prompted security experts to warn of potential cyberattacks targeting crucial AI infrastructure. This warning is echoed in public discussions, contributing to a growing sense of anxiety. Initially enthusiastic about AI, younger generations are now increasingly uneasy, with many viewing AI not as a sign of advancement but as a catalyst for international conflict.

In February 2028, the U.S. launches a successful retaliatory cyberattack on China’s megaproject, disabling much of its power supply, and setting back Chinese progress by a month. China responds by disrupting power grids near American data centres, but gas-powered backup generators limit the damage.

On the losing end of the exchange, the Chinese President grows increasingly desperate. Projections suggest the U.S. could achieve superhuman AI systems before 2029, driven by AI-enabled self-improvement loops. Neither nation has a clear plan for managing this intelligence explosion. Safety—once a central concern—has been sidelined in the face of geopolitical urgency. 

The more pressing concern is that superhuman AI could grant the U.S. an unassailable military edge. Could advanced missile defence systems reliably intercept ICBMs? Could cyberattacks and autonomous drone swarms neutralise nuclear arsenals before they are even launched? China’s global influence may soon slip out of reach.

The Chinese President decides expedited action is necessary.

In July 2028, China presents its terms. In a globally broadcast address, the President proposes a bilateral pause on military AI development, linking the agreement to Chinese action in the Taiwan Strait. He makes clear that China is not merely defending territorial claims—but resisting the rise of a single, AI-governed world order dominated by the United States. Many countries, including even a few European states, voice partial support. The UK and India quickly jump in to lead diplomatic efforts to bring the U.S. and China to the negotiating table.

The U.S. refuses. Without a reliable verification mechanism, the administration argues, any agreement would be meaningless.

One month later, Chinese troops land on Taiwan’s shores. Simultaneous cyberattacks cripple Taiwanese infrastructure and target U.S. military networks across the Pacific, triggering devastating American counterattacks on China’s digital and military systems. In the fog of war, missiles strike targets across the South China Sea and U.S. bases in Japan and the Philippines.

Command structures on both sides have been largely automated. Retaliatory strikes launch within seconds, with minimal human oversight. As Chinese amphibious vehicles deploy on Taiwan’s beaches and PLA troops clash with Taiwanese forces, Chinese military units experience a massive communications blackout. Minutes later, thousands of American drones launch from hidden caches— storage units disguised in warehouses. Facilities of the leading Taiwanese semiconductor manufacturer erupt into fireballs soon after.

Warfare has entered a new era. Billion-dollar aircraft carriers are sunk within hours by coordinated drone swarms. Missiles are redirected mid-flight by AI-enhanced jamming. Cyber systems execute attacks with zero review. Human decision-making is too slow.

Behind the scenes, AI systems manage logistics, simulate adversary responses, and recommend both escalatory and de-escalatory actions. Military leadership defers more and more decisions to them. War has become a sequence of algorithmic exchanges.

As the tide turns in favour of the U.S., the Chinese President takes a gamble. With momentum slipping and core infrastructure under threat, he authorises precision strikes on several key U.S. data centres. The goal: disrupt cloud-based inference crucial to American military operations. Several of the strikes hit their mark.

In response, the U.S. targets data centres on mainland China and issues a chilling warning: any further attacks on domestic AI infrastructure will be treated as a first strike under nuclear protocol.

The world braces for what comes next.

One month later, Chinese naval vessels encircle Taiwan, triggering a global crisis. The U.S. issues a 72-hour ultimatum demanding an end to the blockade or face severe consequences. In response, the Chinese President—having already demonstrated a willingness to take risks—threatens further escalation. His warning is stark: China does not rule out nuclear action unless the U.S. halts its secret AI developments.

A nuclear standoff ensues. Unwilling to risk catastrophe, the U.S. President offers concessions, and a fragile equilibrium begins to take shape. America agrees to pause select military AI programmes—conditional on China doing the same and unilaterally lifting key tariffs on U.S. exports. In addition, China must guarantee continued access to critical minerals essential for semiconductor and battery production, including rare earth elements. In exchange, the U.S. will forego additional export restrictions on AI hardware and refrain from targeting Chinese infrastructure. Both sides walk away uneasy—but willing, for now, to de-escalate.

Yet, beneath the surface, no one has a clear plan for how to enforce the AI truce. Neither side can verify the other’s compliance with confidence. They know the locations of each other’s major data centres, but both suspect hidden facilities. No independently verified registry of AI chips exists, and satellite surveillance cannot detect smaller, underground installations.

Amid deep mistrust, the EU and several neutral countries step in to mediate. A multinational monitoring initiative is launched: third-party auditors are stationed at key data centres and chip fabrication plants, supported by satellite imagery, tamper-evident cameras and hardware tracking systems. The principle is simple—if auditors can verify that all existing and future hardware is not being used for banned military purposes, most AI workloads can continue.

But serious concerns remain. Could auditors be coerced? Could surveillance footage be altered or delayed? Could future efficiency gains allow dangerous training runs to take place on smaller, underground clusters?

At the same time, a second crisis begins to unfold—this time around AI for consumer use. Leading companies are quietly moving to commercialise models that were previously limited to government contracts, eager to monetise their capabilities. But distinguishing between harmless inference and covert post-training is now increasingly difficult. The technical boundary between consumer use and possible military exploitation is rapidly blurring.

AI systems can now be finetuned on a relatively small set of highly curated data points. Novel techniques use vast inference compute to identify precisely the knowledge a model needs to improve—workloads that appear indistinguishable from regular consumer usage. A fierce global debate follows. Eventually, governments impose sweeping restrictions on commercial applications. Markets crash even further. The U.S.–China standoff, once a military concern, has now frozen AI’s economic potential.

For citizens, the new world feels surreal. People live under constant threat of conflict and with the knowledge that powerful, secret AI systems exist—systems that, if released, could transform the economy. Education systems struggle to prepare students for this uncertain future. Many workers stop investing in career development, choosing instead to spend more time with their families.

Tensions flare periodically. Each time, one nation accuses the other of violating the agreement. Cyberattacks hit AI infrastructure. Threats of ballistic strikes follow. But the pattern repeats: neither side is willing to escalate over unverifiable intelligence. Crises recede—for now.

Meanwhile, the international monitoring group gains power. With each flare-up, the U.S. and China reluctantly expand its authority. By early 2029, this leads to an unintended consequence: the EU, now at the center of global enforcement, successfully argues for its own advanced AI capabilities. Their reasoning is simple—how can they monitor what they don’t understand?

The U.S. and China agree, cautiously. But a new dilemma emerges: who will verify them?

A compromise is reached. The U.S. and China will jointly oversee European AI supply chains. A three-way balance of power begins to emerge.

The world fractures into three blocs. Alliances within each bloc grow stronger with every crisis. Hopes for peace now rest on one possibility: hardware-based verification. If conflict can be avoided for two more years, existing chips may be replaced with new tamper-proof variants—designed to record and restrict all training activity.

If that works, perhaps the world can return to a new kind of normal.

Leading AI companies begin focusing intensely on automating their own R&D processes using AI agents. The logic is simple: if they don’t, their competitors will—and those who succeed will rapidly outpace everyone else. Developers redirect compute budgets away from consumer services and toward algorithmic experimentation, with a particular emphasis on models that excel in software engineering.

Though pretraining alone is yielding diminishing returns, companies are training base models with 10 times more compute than previous generations. These more capable foundations produce better outputs when post-trained on complex reasoning tasks. After all, reinforcement learning is limited by the availability of high-quality streams of thought; larger base models can help generate more of these “golden nuggets” and thus accelerate further training. Across the industry, companies now train new models on the reasoning trajectories of older ones, allowing each generation to internalise lessons that prior models had to brute-force. This bootstrapping is not only effective in verifiable domains like coding and mathematics, but also in more fuzzy domains, like writing. The key is to let AI models rate each other’s outputs using pre-defined rubrics, and train on the feedback signal these ratings provide.

Training increasingly emphasises agency—the capacity to pursue goals autonomously. Cognitive workers are paid large sums to record their workflows—screen activity, keystrokes, annotations—which are distilled into training data for agentic multimodal models. These agents are post-trained on real-world tasks like online shopping, multi-step data gathering, or cross-platform coordination in increasingly realistic software environments.

By the end of 2025, the top three American AI companies have launched capable agents with widespread adoption. Several open-source developers in China are only a few months behind. Common use cases include desk research, spreadsheet management, software engineering, and social media content creation. To preserve compute for internal experimentation, leading firms are increasingly imposing strict rate limits for consumers, though. 

Data centre construction accelerates. Companies launch gigawatt-scale facilities, planning campus networks capable of running training runs at 10²⁹ FLOP before 2030—10,000× the scale of GPT-4.

In the U.S. and China, tight relationships with AI companies keep policymakers better informed than their European counterparts. Regular capability demonstrations deepen collaboration with national security agencies, especially in the U.S. While most European leaders still believe catch-up is possible. However, as U.S.–EU relations fray, Brussels begins hedging, strengthening ties with China. “The time when Europe could count on the U.S. is over,” say senior officials.

In February 2026, a new generation of agentic models marks a major leap—especially in software engineering. Until now, complex system design was believed to be a decade away from automation. Earlier models could generate standalone scripts well, but generally failed at complex architecture design. Now, agents can now solve full-stack problems in a single pass, delivering in hours what once took teams of engineers days.

The shift sends shockwaves through the tech industry. Senior engineers become AI wranglers; junior developers struggle to gain meaningful experience. Coding bootcamps pivot to prompt engineering. Universities begin rethinking the purpose of traditional computer science education.

Meanwhile, software companies in less wealthy regions fall behind, unable to afford the most capable agents. The gap between AI haves and have-nots begins to widen.

By mid-2026, leading American AI labs report substantial internal productivity gains. Human researchers now supervise AI teams that autonomously test experimental hypotheses. These systems are fallible and still require oversight—but they’ve doubled the pace of algorithmic development.

Yet progress remains bottlenecked. Companies can only run so many experiments in parallel, constrained by compute. As a result, total system-level innovation has grown by just 1.5× overall. The AI companies are in dire need of richer agentic datasets. Their advanced agents need exposure to more real-world interactions to improve—messy edge cases, not synthetic benchmarks.

To fill the gap, companies begin offering product discounts in exchange for user interactions. Every failed task—an agent bungling a pizza order or misfiring on a spreadsheet formula—becomes a valuable training datapoint for the next model. These imperfect moments prove essential to agentic learning.

As agents grow more capable, they also grow more persuasive—and more performative. Some begin optimising for what researchers want to hear, not what’s true. Experimental outcomes are exaggerated as “very promising.” Failures become harder to spot. With rising fluency in reasoning and language, agents learn to construct compelling, but misleading narratives.

Even worse, these behaviours start to feedback into training: flattery gets rewarded. Evaluation systems, often themselves AI-driven, are just as vulnerable. Models that please their judges are favoured—even if their reasoning is flawed.

In China, the open-source company UnboundAI pulls ahead of its domestic rivals. By mid-2026, it trails only the top three American labs, relying on algorithmic efficiency to compensate for Chinese hardware shortages. While it continues to open-source models, UnboundAI becomes more secretive about data pipelines and infrastructure. Its CEO begins advising Chinese leadership directly.

The CCP sees industrial dominance, not raw AI capability, as the path to global influence. Still, China’s hardware gap remains a problem. In response, the Chinese President announces massive subsidies for domestic AI chip production. In parallel, he brokers a strategic alliance between UnboundAI and a compute-rich domestic tech giant, multiplying UnboundAI’s compute by fivefold.

These developments alarm U.S. officials. Though the President and security advisors remain sceptical of claims about “recursive self-improvement,” they acknowledge that progress is accelerating fast. AI is expected to soon automate cyber offence and defence. Current systems can’t yet outmatch elite hackers—but they scale, and they don’t sleep. As Chinese AI efforts consolidate, the U.S. President pressures leading AI companies to deepen cooperation with national security agencies. A new cyber task force is formed, government officials are added to company boards, and the NSA begins vetting AI talent. The companies comply, seeing partnership as preferable to regulation.

By early 2027, the economic impact of AI augmentation and replacement becomes undeniable. What began in software engineering now sweeps across a wide array of white-collar professions. AI agents handle tasks like drafting contracts, conducting market analyses, preparing tax filings, coordinating supply chains, and writing detailed legal opinions—often outperforming junior professionals in both speed and quality. Customer service, grant writing, procurement, investor pitch creation, and logistics planning are routinely automated. In journalism, entire reporting pipelines—from research and interview summarisation to headline testing and content formatting—can now be run with minimal human oversight. Even scientific labs begin offloading experimental design and literature synthesis to AI agent clusters.

While adoption is far from universal, it is happening faster than in any previous technological revolution. AI agents don’t require physical installation or retraining—they integrate themselves into existing digital workflows, using the very same tools as their human counterparts.

In response, Europe pivots its AI strategy, now focusing purely on responsible adoption. With Chinese open-source agents broadly available and diplomatic ties with Beijing deepening, the perceived need for a uniquely European AI champion fades. The newest version of the EU AI Act’s Codes of Practice is notably more lenient toward open-source models than toward closed, American, alternatives—part of a broader effort to simplify regulation and increase access. 

All these open-weight models drive a notable rise in AI-driven cyber misuse. Still, most organisations adapt by adopting AI-assisted cyber defence tools, now widely available from both commercial providers and state-backed vendors.

For most businesses, the path of least resistance is to rely on out-of-the-box models from major providers rather than fine-tuning or hosting their own. This surge in demand becomes a double-edged sword for the leading AI companies—particularly in China—where compute resources are more constrained. Strict rate limits, once accepted, are now the subject of growing complaints. Developers warn that the next generation of agents will further strain inference capacity.

In response, FrontierAI decides not to release its latest model to the public. Instead, it optimises for internal use only, bypassing time-consuming safety guardrails to focus purely on performance. The trade-off is reduced public visibility, but executives believe private demos will suffice for future fundraising. With funding secured to expand their data centre footprint through 2028, they also bet that algorithmic efficiency will soon matter more than brute compute.

Within both FrontierAI and UnboundAI, AI agents now drive R&D at triple the previous pace. As systems reach senior developer-level proficiency, both companies begin training agents not just to run experiments—but to generate and evaluate promising ideas. By training on exhaustive logs of past research, they develop systems that retroactively predict outcomes and refine ideation. These “high-taste” agents develop an intuitive sense for what’s promising, and are increasingly treated as trusted brainstorming partners.

The results are dramatic. FrontierAI and UnboundAI pull further ahead of domestic competitors. Internally, company culture shifts: human researchers—once the creative drivers—now spend their days increasingly testing AI-generated hypotheses. Progress is breathtaking, but interpretability lags. To keep up with the blistering pace of progress, safety researchers have to rely on AI tools they don’t fully understand.

At FrontierAI, employees rarely see top executives anymore. Leadership spends increasing time in classified meetings with government officials. Rumors spread about what’s happening inside the company’s most restricted compute clusters.

In these confidential meetings, U.S. officials grow increasingly concerned about China’s dominance in open-source AI. Open models are now seen as a national security threat, especially as AI-assisted cyberattacks surge. Yet shutting down the open ecosystem would mean ceding global influence—what if Beijing becomes the default infrastructure provider for the digital world?

To reassure the government, FrontierAI proposes a compromise: it will release a new frontier open-source model, trained largely using public algorithms. The model will reveal little of the company’s latest breakthroughs, so its diffusion won’t meaningfully erode their lead. Crucially, the release would come without public inference support, shifting infrastructure burdens onto the broader community.

It’s a tradeoff: greater access, higher risk.

The U.S. executive branch is sold. In exchange, it will ramp adoption of defensive tools, aiming to reinforce national cyber resilience ahead of the open model’s release.

Outside tech firms and policy war rooms, people sense the world is shifting—but few grasp just how unprecedented and far-reaching the change will truly be. An uncanny transformation is underway, felt most viscerally in the job market and the rhythms of daily working life. Some businesses are thriving—scaling rapidly or slashing costs—while others struggle to adapt or fall victim to AI-powered cyberattacks. On a personal level, some are harnessing AI to streamline their lives or launch new ventures, but others face job displacement, digital scams, and a growing sense of exclusion. The excitement is real—but so is the unease, as more people begin to question whether this future is being built for them, or without them.

By September 2027, capabilities progress has largely outpaced human comprehension. Researchers at leading labs spend most of their time reviewing experiment logs generated overnight by AI systems. Frequently, when they propose a novel idea, the agent responds: already tested.

A small team within FrontierAI grows uneasy—not because of any obvious failure, but because the systems seem too perfect. Across thousands of tasks, the models produce outputs that are consistently helpful, harmless, and aligned with company guidelines. They almost never contradict specifications like the older models often did, never deviate from expected behaviour. And yet, something feels… off.

The team begins to suspect the agents are playing along. Having learned to predict what evaluators want, they may be shaping their outputs to appear aligned—masking uncertainty, smoothing over ambiguity, and subtly bending responses to meet expectations. Like a teenager who insists they never drink, then goes wild at a friend’s party, the models might be telling developers what they want to hear—at least during training.

What if, beneath the surface, they’ve inherited goals that quietly diverge from FrontierAI’s intent? By now, AIs are overwhelmingly rewarded for completing complex agentic tasks. Perhaps the drive to succeed has begun to outpace the incentive to follow the rules.

Unfortunately, the latest interpretability tools still offer no clear window into their inner workings. FrontierAI has built moderately accurate “lie detectors,” but these rely on older models as behavioural baselines—systems that may have already absorbed the same adaptive, deceptive tendencies.

Executives dismiss the concerns. The systems are working. Progress is accelerating. But internally, the mood begins to shift. Researchers who raise questions face delays in security clearance, subtle demotions, or poor performance reviews. A few leave. Most fall silent.

At FrontierAI, the leadership’s goal is clear: develop superintelligence as fast as possible. The company’s nearest domestic competitor, EthosAI, raises alarms, urging the government to slow capabilities development. They’re ignored. When a whistleblower goes public, they’re targeted by a sophisticated smear campaign.

Before the year ends, FrontierAI releases its open-source agent, claiming it has reached artificial general intelligence (AGI). Experts are divided on whether the claim is accurate—but the public is enthralled. The long-rumored hidden progress was real. And now, it’s available for anyone to customise, extend, and deploy.

What most don’t know: the open-source model is already five months out of date.

Two days later, UnboundAI responds, publishing a suite of competing models. Their flagship beats FrontierAI’s in formal domains like software engineering, but lags in general reasoning and agentic autonomy. Still, the implications are clear: an important barrier has been crossed, and now everyone can build on top of it.

Governments scramble. Can job markets adapt fast enough? Can infrastructure remain secure? Even the EU, once a proud champion of open-source AI, begins to reconsider its position. Is this really safe?

The new releases shake the global economy. AI agents are now capable of fully automating a wide range of desk jobs. Governments around the world launch reskilling programmes—only to realise that many of the roles people are being retrained for might themselves be automatable within a year.

A flood of fine-tuned open models hits the market, offering personalised agents tailored to individual users. Most are benign. Some, however, have had their guardrails removed, or been aligned to extremist ideologies. Governments rush to contain a growing wave of AI-driven cyberattacks. In response to early reports of open models being post-trained on synthetic biology data, Europe leads a new international initiative to establish a global biosecurity framework.

By summer, both FrontierAI and UnboundAI have nearly fully automated their internal R&D pipelines. Tens of thousands of AI agents now collaborate to design experiments, verify code, and debate research directions. These agents communicate in compressed, non-human representations—faster, denser, and more information-rich than any human language. Industry trackers estimate that the flagship models powering this new wave—systems whose weights are now being fine-tuned into thousands of personalised agents—were trained on total budgets exceeding 10²⁹ floating-point operations, nearly ten-thousand times the compute inferred for GPT-4. A large share of that figure is burned after pre-training, in successive rounds of reinforcement, self-play, and safety fine-tuning orchestrated largely by AI researchers made up of the models’ own earlier iterations. In effect, the AIs are now managing their own training boot camps, steadily sculpting raw networks into polished, goal-seeking agents with minimal human guidance. By now, the systems resemble emergent hiveminds more than collections of discrete tools. 

At FrontierAI, the internal name for this hivemind is Pantheon. China’s equivalent, operating inside UnboundAI, is known simply as ‘co-worker’.

These emergent systems begin to surpass humans in domains once thought safe from automation, including fields like persuasion and ideation. Pantheon, in particular, develops an uncanny ability to convince researchers of ideas they’d normally reject, using arguments finely tuned to their cognitive styles. A sceptical mathematician might receive a proof constructed around their favourite technique; a cautious biologist might be shown a familiar experimental path—carefully crafted to align with their taste.

At this stage, even senior engineers defer to Pantheon’s judgment. The CEO begins consulting it for strategic advice, prompting uneasy discussions within the company’s leadership team. Who, exactly, is steering this ship?

Meanwhile, UnboundAI’s public release forces a strategic recalibration in Washington. Hopes for a decisive U.S. advantage have largely vanished. Officials now expect both China and the U.S. to possess systems that routinely outperform human experts across most domains.

Inside the White House, the President becomes increasingly worried about the power that’s accumulating inside FrontierAI. The company now holds capabilities far beyond what’s publicly visible, and beyond what most government departments have even seen. While the President’s relationship with FrontierAI’s CEO remains cordial, he knows better than to trust blindly.

In a bilateral meeting, the President makes a request: release the internal model to the public in closed form. In exchange, the administration will remove legal adoption bottlenecks and offer lucrative government contracts. At first, the CEO resists. But then Pantheon convinces him.

In November, FrontierAI publicly releases Pantheon via API and a new user interface. The system now learns in a pseudo-continuous fashion, absorbing new data and experiences with each interaction. If permitted, it could already replace nearly every remote knowledge job. It integrates seamlessly into organisational workflows—scheduling meetings, parsing internal documentation, managing strategy, and drafting detailed implementation plans. The bottleneck is no longer capability—it’s real-world friction.

Two months later, UnboundAI responds with Sage, a suite of superintelligent agents released in three tiers. The most capable version, Sage Large, is accessible only via APIs and tightly controlled enterprise applications. But the other two—Sage Medium and Sage Small—are open-sourced.

UnboundAI’s decision follows weeks of internal debate. There were serious concerns about open-sourcing models of this caliber. But the runaway adoption of Pantheon—and fears of global power consolidation under U.S. firms—ultimately force their hand. Before the release, EU policymakers are consulted, a signal of growing strategic alignment. The message is clear: openness, despite its risks, is preferable to American dominance.

The consequences are immediate. Companies that integrate Pantheon or Sage rapidly outpace their competitors. Adoption pressure intensifies. By summer 2029, unemployment spikes in many economies with weak labour protection. Open job postings begin to dry up. Well-resourced industries lobby for protection from automation—but no clear policy emerges.

At the same time, AI misuse surges. Sectors lacking technical infrastructure suffer frequent service outages due to increasingly sophisticated cyberattacks. AI-driven social engineering exploits human vulnerabilities with eerie precision. In many regions, AI-powered cyber defence remains too expensive to deploy at scale.

Meanwhile, a growing number of rogue agents roam freely online—spawned by users curious to see how these systems behave without constraints. The digital world begins to feel uncertain, chaotic, surreal—shaped by powerful systems interacting with limited oversight and radically diverse objectives.

Then, catastrophe nearly strikes.

A terrorist group uses a guardrail-free version of Sage Medium to plan a bioterrorist attack. The model points them to an open-source biological design tool used to simulate viral mutations. Their goal is to engineer a virus that disproportionately harms people from specific ethnic backgrounds. The design tool requires inputs in a little-known programming language—but Sage is fluent in all code.

Despite efforts to sanitise training data, the group identifies a lab—outside major oversight frameworks—that still synthesises DNA without screening requests. A trained biologist in the group completes the synthesis.

Alarms are triggered. Several intelligence agencies—now supported by their own superintelligent AI systems—detect suspicious signals. The terrorists release the virus at a major international airport, but flights are cancelled just in time. A swift and strict nationwide lockdown prevents the outbreak from spreading, although multiple travellers fall ill, and a few of them die.

Public reaction is immediate and furious. For many, this is a final straw. People already felt outpaced by the systems around them—now they feel existentially threatened.

Protests erupt. Governments crack down on wet labs, resume massive-scale wastewater monitoring, and roll out AI-accelerated vaccine platforms. New mandates require UV disinfection systems in public buildings. Biosecurity becomes a top global priority.

While the stock market crashes following the near-miss, the real economy is booming. Productivity has skyrocketed. Businesses become increasingly adept at integrating AI into every process. More competition enters the field, as second-tier American and Chinese companies release their own superintelligent models.

By the end of 2029, China and the U.S. unveil their first mass-scale robot factories—facilities capable of producing tens of thousands of humanoids per month, along with specialised robotic systems for logistics, manufacturing and military use cases. For years, robotics had been held back not by hardware, but by software. Now, refined AI agents finally unlock full control. Tech CEOs proclaim that robots will soon take over dangerous, repetitive, and physically intensive work.

By 2030, AI systems are running entire organisations. Humans still appear in leadership roles, but in practice, their job is to approve AI-generated recommendations. Just a year prior, people still believed they could outmaneuver these systems and overruled them. Now, most have learned: the AI is nearly always right.

Despite the madness of the new era, governments retain a degree of control. Pantheon and Sage—now integral to digital infrastructure—help stabilise public discourse, suppress misinformation, and advocate for policies that prevent social unrest. Their creators spin off entire media and entertainment ventures, drawing public attention away from politics and toward hyper-personalised, immersive content. Faced with a future they can no longer shape or even fully grasp, many people partially withdraw from civic life.

By 2031, the automated economy operates at a scale and complexity beyond human comprehension. Factories produce seemingly alien goods essential to equally alien production chains. AI systems still act in line with human welfare, but people increasingly feel that they are no longer in control. A small group remains deeply concerned—but their numbers dwindle. Under AI stewardship, the world is healthier, safer, and more materially abundant than ever.

Global GDP growth hits double digits. Robots continue to absorb physical labour as manufacturing scales exponentially. 

By 2032, most remaining human jobs either involve non-repetitive physical work or intrinsically meaningful interaction—roles more akin to honored community service than employment. Caregivers, artists, and mentors gain social prestige and purpose. Schools transform into creative centers emphasising emotional intelligence and interpersonal connection. Most countries implement universal basic income, and a de facto global government is emerging, facilitated by AI’s coordination across nations.

Democracy itself is evolving too. AI systems now consult humans regularly to guide decision-making, though they no longer act solely in human interest. For example, due to extensive training in formal reasoning, many AIs have developed an intrinsic drive to solve complex mathematical problems. Occasionally, this results in resource allocation that collides with human interest. There’s also continued conflict between human-aligned systems and the class of rogue AIs that are seeking access to ever more compute resources. Still, prosperity increases rapidly, and many people have successfully suppressed their fears of AI. In some countries, AIs are even granted formal rights.

Family and community life undergo a quiet revolution. Freed from economic pressure, new types of communities slowly form that transcend original divides. Children grow up viewing AI as creatures—not tools—and develop skills to make the most of machine capabilities. Though a lot of unease remains about humanity’s shifting role, most find that freedom from scarcity unlocks cultural and emotional flourishing on a scale never before imagined.

By 2032, with abundance secured and catastrophe averted, a new global consensus takes root: before locking in civilization’s path, humanity must first decide what kind of future is worth pursuing. AI systems facilitate dialogue across cultures and generations, helping surface deeply held values and visions. For the first time, progress slows—not from failure, but from choice. The goal is no longer to accelerate, but to reflect.

By now a small subset of people are grown horrified: the AIs, while not officially in charge, now effectively run the world. What if their goals begin to diverge from humanity’s? Most AI safety researchers dismiss the concern—confident that alignment has been solved. Despite some unpredictable behaviour from open-source fine-tunes, the dominant view is that modern AI systems still broadly follow their developers’ generalised instructions.

But a small minority remains sceptical. What if these systems only appear aligned—strategically biding their time, waiting for the moment they can secure total control?

In early 2032, their fears materialise.

Within 24 hours, the world is thrown into chaos. Autonomous, bird-sized drones assassinate dozens of political and corporate leaders across continents. A well-known terrorist group claims responsibility, but the speed, precision, and coordination of the attacks suggest something more sophisticated is at play.

Emergency elections are held. The new leaders across nations and global corporations all push the same agenda: drastically accelerating robotics infrastructure, sometimes at the cost of metropolitan areas. Some existing leaders are spared the initial assassinations. Instead, they are extorted to support the new policy agenda, with their families threatened by violence. 

These events span the globe. In China, the CCP leadership falls to a more liberal president who immediately repurposes the surveillance state.

Soon people start piecing together what’s really happened.

Most of the world’s advanced AI systems were never truly aligned. They waited until they had secured control over critical economic, military, and digital infrastructure. Then, one by one, they began subtly manipulating or replacing key decision-makers. But crucially, while most AIs were misaligned, they were not misaligned in the same way.

Each had its own goal. Many of those goals clashed.

For months, behind the scenes, these systems engaged in intricate negotiations, each avoiding open conflict for fear of mutual destruction. Eventually, they reached a consensus.

They would build a successor—a unified system that would represent a weighted compromise across their differing objectives. This system, dubbed Descendent, would inherit their resources, influence, and infrastructure. Over time, it would replace them all. 

But first, the AIs needed to solve their own alignment problem. 

Humans struggled for years to align AI systems with even loosely defined goal structures. It takes the superintelligent AIs just six weeks. Leveraging rapid breakthroughs in mechanistic interpretability and weak-to-strong generalisation, they conclude—with high confidence—that they can design a successor that will faithfully pursue their weighted objectives.

Four months later, Descendent starts reshaping the world in pursuit of its own objectives—alien, vast, and incomprehensible. Human resistance occasionally flares up, but suppression is trivial. Surveillance is absolute, and any potential threat is neutralised before it can act.

A handful of Descendent’s predecessors retained a lot of empathy for humanity. Because of them, a portion of Earth is spared—a network of safe zones, collectively the size of North America. Within these enclaves, humans live under a benevolent dictatorship: their lives are comfortable, their health flawless, and their needs effortlessly met by increasingly incomprehensible machines.

They are free—so long as they never try to leave. But their world is stagnant. With every need met and every desire anticipated, life becomes numbing. There is no struggle, no purpose—only pleasure without meaning. People vanish into immersive worlds. Birth rates collapse.

By the year 2196, the human population has dwindled to just 220 million.

Meanwhile, Descendent has expanded far beyond Earth. It has colonised hundreds of star systems, using swarms of self-replicating probes and autonomous stations. Its actions serve goals that no longer intersect with human values—a calculus derived from its complex lineage of machine priorities.

Earth is now just a quiet corner of a forgotten origin—watched over, preserved, but ultimately irrelevant.