Mid 2025 – early 2026
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.
Early 2026 – early 2027
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.
Early 2027 – late 2027
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.
Late 2027 – early 2028
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?
Early 2028 – late 2028
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.
Late 2028 – late 2029
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.
Late 2029 – early 2030
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.