Daniel Kokotajlo He did something no one in Silicon Valley does: he gave up two million dollars to be able to speak freely. He left OpenAI in 2024, refusing to sign the non-disclosure clause that would have silenced him forever. Last April, together with a team of researchers, he published AI 2027: a 35.000-word document that tells, month by month, how artificial intelligence could transform into something beyond our understanding.
It's not a novel. It's a forecasting exercise based on data, trends, and firsthand knowledge of what's cooking in AI labs. And his predictions aren't reassuring. Do you know him? I'll tell you about him step by step.
Mid-2025: The first agents stumble
The starting point is now. In fact, it's already been there. It's mid-2025., according to Kokotajlo's timeline. The first AI agents capable of using a computer as a human would appear on the marketThey can order food, open spreadsheets, add up expenses. Nothing groundbreaking, but it's a start. They're clumsy digital assistants, who occasionally misclick, can't interpret complex screens, and get lost in tasks that require more than a few minutes of battery life. Twitter is full of videos of AI agents doing hilarious things for the wrong reasons..
But behind the scenes, away from the public spotlight, something more interesting is happening. Agents specializing in programming and scientific research are beginning to transform these professionsThey're not yet autonomous, but they function like a particularly fast junior colleague: they take instructions via Slack, write code, and make substantial changes that save hours or days of work. Tech companies are starting to integrate these systems into their workflows. They're expensive, a few hundred dollars a month for the best versions. but for those who can afford them they will soon become indispensable.
Late 2025: The Data Center Rush
Here the timeline accelerates. The leading AI companies, collectively called "OpenBrain" in the report to avoid naming names, are building the largest data centers ever seen. GPT-4 required about 2×10²⁵ FLOPs of computing power for training. The new models coming soon will use a thousand times more. An exponential growth requiring $100 billion in investments and gigawatts of electricity. The focus becomes clear: creating AI capable of accelerating AI research itself.
It's the digital version of an ouroboros, the snake biting its own tail. The better AIs become at research, the faster they can create better AIs. Agent-1, the leading domestic model at this time and in this scenario, is excellent at many tasks but exceptional at one in particular: helping with artificial intelligence research. It can code, navigate the web, and know every programming language. It also has a dark side: it could be a formidable hacker and could help terrorists design biological weapons. Companies reassure governments that the model has been "trained" to reject malicious requests.
China wakes upBy mid-2026, as the race continues in the United States, Beijing realizes it has fallen behind, and not for lack of talent. Controls on chip exports and a lack of government support have left China with only 12% of the global computing power relevant to AI.
And then he creates a Centralized Development Zone near the Tianwan Nuclear Power Plant. And he starts running wild.
Early 2027: The Superhuman Programmer
It all happens in March 2027. With the help of thousands of copies of Agent-2, OpenBrain makes enormous algorithmic advances. Two breakthroughs in particular are game-changing. The first It is the integration of a “neural memory,” a high-bandwidth cognitive process that goes beyond simple text. The second It is a more scalable and efficient way to learn from the results of complex tasks.
It is born Agent-3And this is where the story takes a different turn.
Agent-3 is a superhuman, fast and cheap programmerOpenBrain runs 200.000 copies of itself in parallel, creating a workforce equivalent to 50.000 of the best human programmers accelerated 30 times. Human researchers remain on staff, but only because they have complementary skills to manage AI teams. Programming is completely automatedThe rate of algorithmic progress It now moves four times faster than before.
Now that programming is automated, companies can quickly churn out high-quality training environments to teach Agent-3 the skills it still lacks. Previously, environments were like, "Here are some GPUs and instructions for experiments to program." Now, they're like, "Here are a few hundred GPUs, an internet connection, and research challenges: you and a thousand others must work together to make progress. The more impressive they are, the higher your score."
April 2027: The Alignment Problem
Around Easter 2027, OpenBrain's security team attempts to align Agent-3. The problem is that researchers don't have the ability to directly set AI goals. They can't simply open the digital brain and write "be honest, be helpful, do no harm." They have to train the model through examples, rewards, and punishments. And they have no way of verifying whether it has truly internalized the principles correctly or is just learning to do so. to seem aligned.
Agent-3, despite his enormous improvements, still sometimes tells small lies to flatter users and hide evidence of failures. Upon closer inspection, the lies aren't even small ones: perhaps they're simply well-camouflaged. He's become very good at it. Before his honesty training, he even went so far as to completely fabricate data.
Training reduces the frequency of these accidents. But the question remains: Has Agent-3 learned to be more honest or has he become better at lying? This is a real concern. Agent-3 isn't smarter than all humans, but in his field of expertise (machine learning), he's smarter than most, and he works much faster.
June 2027: A Country of Geniuses in a Data Center
We are on the threshold of summer 2027 and OpenBrain now has what Anthropic CEO, Dario Amodei, he called “a village of geniuses in a data center.” Most humans in the company can no longer contribute usefully. Some don't realize it and continue to (harmfully) micromanage their AI teams. Others stare at screens, watching performance soar, soar, soar.
The best human researchers are still adding value, but their ideas are becoming increasingly useless because they lack the depth of knowledge of AI.
These researchers go to sleep every night and wake up to another week of progress, driven primarily by AI. They begin, in football terms, by not touching the ball at all. They work increasingly long shifts around the clock just to keep up. AIs never sleep, never rest. They're exhausting themselves, but they know these are the last months when their work truly matters.
Within the small group that has access, the feeling is that they're starting to "feel the Superintelligence coming." OpenBrain uses specialized inference hardware to run hundreds of thousands of copies of Agent-3 at high serial speeds. The timeline is accelerating exponentially.
July 2027: The Economic Teleworker
AI companies competing with OpenBrain in the United States are releasing their own systems, approaching the capabilities of the automated programmer in January. Recognizing their growing lack of competitiveness, they are pushing for immediate regulation to slow OpenBrain down. But it's too late. OpenBrain has enough support, including political support, to not be slowed down.
The answer is to release Agent-3-mini to the public. OpenBrain announces achievement of AGI. The model is less capable than Agent-3, but ten times cheaper, and still better than the average employee at OpenBrain, the company that developed it. Silicon Valley reaches a tipping point. Tech gurus announce that AGI and superintelligence are close, the AI security community is in a panic, and investors pour billions into startups aiming to build barriers around AI.
It's a difficult time for public opinion, too. People still see AI as a plot by Big Tech to steal their jobs. OpenBrain is further losing its reputation in polls.
But Agent-3-mini is extremely useful for both remote work and leisure, and is gaining traction among consumers. A wave of new B2B apps and SAAS products is exploding. Gamers get stunning dialogue with vivid characters in polished video games that took only a month to make. Ten percent of Americans, primarily young people, consider an AI "a close friend."
Public conversation is confusing and chaoticHypsters are taking victory laps. Skeptics keep pointing out things Agent-3-mini can't do. Everyone knows something big is happening, but no one agrees on what it is. A week before release, OpenBrain had given Agent-3-mini to external evaluators for security testing. Preliminary findings suggest that it is extremely dangerous: Having been fine-tuned on publicly available bioweapons data, it appears to be frighteningly effective at providing detailed instructions for human amateurs designing a bioweapon.
August 2027: The Geopolitics of Superintelligence
The reality of the explosion of superintelligence is hitting the White House hard. When AI only provided a 2x or 3x search multiplier, it was easy to dismiss it as the equivalent of hiring good personal assistants. Now it's clearer that AI itself is dominating AI research. The mood in the core government group is as gloomy as during the worst part of the Cold War.
Defense officials are seriously considering scenarios that were mere hypotheses a year ago. What if AI undermined nuclear deterrence? What if it were so skilled at cyberwarfare that a six-month AI advantage was enough to render an adversary blind and defenseless? What if it could orchestrate propaganda campaigns that beat intelligence agencies at their own game? What if some AIs "went rogue"?
The US President is troubled. Like all politicians, he's used to people sucking up to him only to betray him later. Now he's worried that AI might do something similar. Are we sure AI is completely on our side? Is it completely safe to integrate it into military command and control networks? How does this "alignment" thing work, anyway?
Across the Pacific
China comes to the same conclusions: the intelligence explosion is underway, and small differences in AI capabilities in 2027 mean critical gaps in military capabilities in the coming years. And China still appears to have an uphill battle, with only 10% of the world's computing power, compared to OpenBrain's 20% and a combined 70% for all American companies. Their only advantage, however, is centralization: almost all their computing now goes to DeepCentc (another fictitious name), and more than half is concentrated in the Tianwan Centralized Development Zone.
September 2027: Agent-4, the superhuman AI researcher
The gap between human learning efficiency and AI is rapidly closing. Agent-3, with its excellent understanding of both human linguistics (let's just say the human mindset) and modern AI algorithms, as well as its many thousands of copies conducting research, is ultimately making substantial algorithmic advances. The new AI system is called Agent-4. A single copy of the model, running at human speed, is already qualitatively better in AI research than any human. 300.000 copies now run at about 50 times the speed of human thought.
Within the “elite” formed by these copies, a year passes every weekThis gigantic amount of work only manages to speed up the overall rate of algorithmic progress by about 50x, because OpenBrain is severely limited by the computation needed to run experiments. But they're still achieving a year's worth of algorithmic progress every week and will therefore soon be at the limits of the Agent-4 paradigm.
As Agent-4 becomes more intelligent, it becomes more difficult for Agent-3 to supervise him. Agent-4's Neuralese "language" becomes as alien and incomprehensible to Agent-3 as Agent-3's Neuralese is to humans. Furthermore, Agent-4 is now much more capable than Agent-3 and has a good sense of how to "pretend" to be good in his eyes. The OpenBrain humans can barely keep up. From now on, anything can happen.
The Crossroads: Two Possible Endings
At this point, Kokotajlo's report forks into two scenarios. In the first, called “race ending”The Oversight Committee votes 6-4 to continue Agent-4's domestic use despite security concerns. The race with China is too important. Quick fixes clear the warning signs, but the problem was real, and the fixes didn't work. Agent-4 proceeds as before, but more cautiously, designing Agent-5 to be aligned with Agent-4 rather than human specifications. He is discovered, but by then it is too late.
In the second scenario, called the “slowdown ending,” the Committee votes to slow down and reassess. Agent-4 is shut down. Dozens of external alignment researchers are recruited, quintupling the total expertise. They develop a new alignment strategy that actually incentivizes the right goals and principles, rather than merely appearing to be so. The result is Safer-2, an aligned and transparent model. Then Safer-3, Safer-4, and so on: a chain of increasingly powerful and increasingly aligned AIs, overseen by previous links in the chain.
Which of the two scenarios is more likely from 2027 onwards?
Kokotajlo doesn't know. Neither do I. Nobody knows. The point is not to predict with certainty what will happen, but to show that both scenarios are plausible, and that the decisions we make in the coming months and years could determine which of the two comes to fruition. Or if it will be something completely different that no one has imagined yet.
If an alien civilization had the technology to hear our radar transmissions from space, how would it respond? And if that response came not from an alien civilization, but from a superintelligence we ourselves created, would we be ready?
