The Race to AGI

On Friday, the president of the United States ordered every federal agency to stop using Anthropic's artificial intelligence. The reason: the company refused to strip safety guardrails from its models at the Pentagon's request. Hours later, OpenAI announced a deal to supply AI to classified military networks, filling the gap. Defense Secretary Pete Hegseth then designated Anthropic a supply-chain risk to national security, a classification typically reserved for foreign adversaries.

This is not a story about corporate rivalry. It is a story about what happens when a technology becomes so powerful that governments start treating access to it the way they treat access to weapons-grade material. And the technology in question does not yet do the thing its creators say it will soon do.

The race to build artificial general intelligence has moved from research labs into the machinery of state power. To understand why, you need to understand what the participants already have, what they believe they are about to get, and why the gap between those two things has every major government and corporation on Earth spending money at a pace that makes the space race look quaint.

The Four Pillars of Power

The simplest way to understand what is at stake is to understand what power actually is. Not in the abstract. In practice, throughout history, power has always rested on four pillars.

Knowledge is power. Francis Bacon coined the phrase in 1597, but the principle is older than language. The Rothschild banking dynasty built its fortune in part because its private courier network delivered news of Napoleon's defeat at Waterloo to London before anyone else had it. The exact scale of Nathan Rothschild's resulting trades remains debated by historians, but the underlying principle is not: he had information others lacked, and he acted on it first. The entire intelligence apparatus of every major nation exists for the same reason: knowing things before your adversary is worth more than almost any other advantage. Speed of information becomes speed of decision becomes speed of action. The gap between those who know and those who don't has always been the first and most durable source of advantage.

Money is power. This is not controversial, but the mechanism is worth unpacking. Money is stored optionality. It is the ability to take risk, absorb failure, move before competitors, and buy talent that would otherwise go elsewhere. The Dutch East India Company conquered territory across Asia and Africa not because it had the best soldiers but because it could finance armies, navies, and trade networks simultaneously, sustaining losses that would have bankrupted a nation-state. The Marshall Plan rebuilt Western Europe because the United States could write a $13.3B check (roughly $170B in today's dollars) at a moment when no other country on Earth had the capital to act at that scale. When an entity spends at that level, it is not purchasing a product. It is purchasing the future ability to act while others cannot.

Work is power. Whoever controls what gets done controls the outcome. The Industrial Revolution was not a story about steam engines at its core. It was a story about who owned the factories that replaced cottage labor, and what happened to the people who used to do that work by hand. The Luddites were not technophobes. They were skilled textile workers who understood, correctly, that the new machines transferred the value of their labor to the people who owned the machines. Slavery built the economic foundation of the American South because controlling labor at zero cost meant controlling the entire agricultural output of a region. Unions emerged because workers recognized that collective labor is leverage, that the people who do the work hold power only when they act together. The question at the center of every economic transition is the same: who does the work, and who directs it?

Power is power. Max Weber defined the state as the entity that holds a monopoly on the legitimate use of force. Everything else, knowledge, money, work, operates within the boundaries that raw power draws. The British Empire maintained global influence for two centuries because the Royal Navy could project force to any coastline on Earth. The Cold War was a forty-year standoff built on the simple question of who could destroy whom. International trade agreements are enforced not by goodwill but by the implicit threat of sanctions, asset seizures, or military response. Law is codified power. Imprisonment is applied power. The capacity to compel, through courts or through force, is the foundation on which every other form of power ultimately rests.

These four pillars are not separate phenomena. They form a cycle. Knowledge generates money. Money funds work. Work builds power. Power secures knowledge. Every empire, corporation, and institution that has ever dominated its era did so by controlling multiple pillars simultaneously. The British Empire had naval supremacy (power), global trade networks (money), the extracted labor of colonized populations (work), and the world's most advanced intelligence service (knowledge). The United States after World War II had nuclear weapons (power), the world's largest economy (money), an industrial base that had just outproduced every other nation combined (work), and the CIA and NSA (knowledge).

The reason the race to build artificial general intelligence has every major government and corporation spending at unprecedented scale is straightforward: AGI is the first technology in history that could accelerate all four pillars at once. But to understand why that claim is credible rather than hype, you need to understand what changed.

Why the AGI Race Accelerated

Three years ago, most serious AI researchers placed AGI decades away. Two things changed that.

The first was that scaling laws held. Between 2020 and 2024, each generation of large language model reliably gained capabilities by using more data, more compute, and more parameters. GPT-3 to GPT-4 was not a small jump. It crossed thresholds in reasoning and coding that researchers had expected to require architectural breakthroughs. The capabilities showed up from scaling alone.

The second was that the gains started showing up in places nobody predicted. Emergent capabilities, the term researchers use when a model suddenly performs a task it was never trained for, became the norm rather than the exception. (The concept itself is debated — some researchers argue the "emergence" is partly an artifact of how we measure — but the practical result, models doing unexpected things, is not in dispute.) Models began solving tasks well outside their expected capabilities. Models trained primarily on English started translating languages they had barely seen. The boundaries of what "just a language model" could do kept expanding in ways that made the critics' arguments harder to sustain.

Then DeepSeek made the conversation global. The Chinese lab's V3 model, released in early 2025, matched or approached the performance of frontier American models at a reported training cost of $5.6M (though total compute investment likely exceeded $1B). Their R1 reasoning model, released the same month, demonstrated competitive chain-of-thought reasoning at a fraction of frontier costs. The message was clear: this is not an American monopoly. The preconditions for building frontier AI systems exist in multiple places, and the cost of entry is falling.

That last point matters. When a technology gets cheaper while getting more capable, adoption curves steepen. AI inference costs have dropped by roughly 10x per year since 2023. What cost $100 to run in early 2024 costs pennies today. The same trajectory played out with gene sequencing two decades ago: the Human Genome Project cost $2.7B over thirteen years; today you can sequence a human genome for under $200 in a day. When costs collapse that fast, the technology stops being a tool for specialists and becomes infrastructure that everyone touches.

China has its own trajectory. Xi Jinping's 2017 national AI strategy set a target of global AI dominance by 2030. The country's semiconductor self-sufficiency push, accelerated by U.S. chip export controls, means the race is not just between companies but between industrial policies. DeepSeek was the proof of concept. The next generation of Chinese AI labs will build on cheaper compute, larger domestic datasets, and a regulatory environment with different priorities around deployment speed and safety constraints.

That trajectory means the question is not whether organizations will adopt AI but how fast the adoption will outrun their ability to understand what they have adopted.

2025 was supposed to be the year agentic AI arrived, the year AI systems stopped answering questions and started completing tasks. It partially delivered. Coding agents, research assistants, and customer service bots all improved. But the "drop-in remote worker" that several CEOs had promised did not materialize. Metaculus forecasters responded by pushing the strong-AGI median out by two years. The 80,000 Hours podcast ran an episode titled "What the hell happened with AGI timelines in 2025?" The correction was real. But even after it, the median estimate for strong AGI sits decades closer than it did in 2020. The trend bent. It did not break.

Why Multiple Labs Are Converging on AGI

In January 2026, the CEOs of the three leading AI labs appeared at Davos and said nearly identical things. Dario Amodei of Anthropic predicted AGI in 2026 or 2027. Demis Hassabis of Google DeepMind gave a 50% probability by 2030. Sam Altman told an audience that OpenAI knows how to build AGI and has shifted his language from AGI to superintelligence, as though the first milestone is already in the rearview mirror.

This convergence of timelines from independent organizations is not evidence of groupthink. It is evidence of something more interesting.

In 1858, Charles Darwin received a letter from Alfred Russel Wallace describing a theory of evolution by natural selection that was, in its essentials, identical to the one Darwin had been developing in private for twenty years. In 1876, Alexander Graham Bell and Elisha Gray filed patent applications for the telephone on the same day. Isaac Newton and Gottfried Wilhelm Leibniz invented calculus independently, on different continents, within a few years of each other.

These are not coincidences. They follow a pattern: when enough prerequisite knowledge, tools, and problems converge in a given era, certain inventions become nearly inevitable. Multiple people or groups arrive at the same breakthrough because the conditions for it have matured. The discovery is latent in the infrastructure of what is already known.

AI in 2026 fits that pattern. The prerequisite stack, including transformer architectures, massive training datasets, GPU clusters measured in hundreds of thousands of chips, reinforcement learning from human feedback, and chain-of-thought reasoning, exists in at least five organizations across two continents. Each lab has arrived at similar capabilities through somewhat different paths. The convergence of their timelines reflects the convergence of the underlying conditions.

The pattern has a name in the history of science: multiple discovery. Sociologists Robert Merton and Harriet Zuckerman documented hundreds of cases. Their finding was that simultaneous invention is not the exception. It is the default, once preconditions align.

This matters because it means the race cannot be stopped by slowing down any single participant. If the preconditions are mature, the breakthrough will come from whichever lab reaches it first. This is the logic driving the spending, the government contracts, and the geopolitical tension. Nobody wants to be second.

What They Are Racing Toward

The word "AGI" means different things to different people, which is part of the problem. It helps to think about it in two tiers.

The first is domain-specific AGI: a system that is expert in one field and can operate autonomously within it. In software engineering, this means a system that can build, operate, and maintain a complex production project on its own — writing code, debugging it, responding to outages, and improving its own approach over time. It writes code, debugs it, responds to outages, and improves its own approach over time. We are close to this. Current AI systems can already do pieces of it; the gap is in sustained autonomy and reliability.

The second is multi-domain AGI: a system that operates at expert level across fundamentally different fields. Not just software engineering, but scientific research, strategic analysis, and creative work, all within a single system. This is the version that most experts mean when they use the term, and it is the version that changes everything.

The distance between those two tiers may be shorter than it looks. Once a system can improve its own capabilities (recursive self-improvement, or RSI), the gap between "expert in one domain" and "expert in many" compresses in theory. If each improvement cycle makes the next one faster, the transition from narrow to general could happen in months rather than years. RSI has not been demonstrated yet, and it may hit diminishing returns that nobody can predict in advance. But every major lab is building toward it, and none of them are betting it will fail.

So what happens if they get there?

The Discovery Engine

AlphaFold was the preview.

In 2020, Google DeepMind's protein-folding AI solved a problem that had stumped biologists for fifty years: predicting the three-dimensional structure of a protein from its amino acid sequence. Within two years, AlphaFold had predicted the structures of virtually every known protein, over 200 million of them. Research linked to AlphaFold 2 is now twice as likely to be cited in clinical articles as typical structural biology work. It won its creators the 2024 Nobel Prize in Chemistry.

But AlphaFold is a narrow tool. It predicts protein shapes. It does not design experiments, generate hypotheses, or synthesize results across disciplines. It assists scientists. It is not a scientist.

Post-RSI AI would close that gap. A system that can improve its own reasoning could move from "tool that answers questions biologists ask" to "system that asks its own questions and designs its own experiments to answer them." The difference is the difference between a calculator and a mathematician. A calculator processes what you give it. A mathematician identifies what to calculate in the first place.

Automated scientific discovery at that level would compress research timelines by orders of magnitude. Drug candidates that currently take 4-5 years from target identification to clinical trials could be identified, optimized, and validated in months. Materials science, which still relies heavily on trial-and-error synthesis, could shift to computational design followed by targeted verification. Climate models could incorporate granularity that is currently computationally prohibitive. Each breakthrough would feed back into the system's ability to make the next one.

Programmable Biology

The most surprising place this leads is not digital. It is biological.

CRISPR gave scientists the ability to edit genes with precision. AlphaFold gave them the ability to understand the proteins those genes encode. Gene synthesis costs have been dropping rapidly. What has been missing is the intelligence to orchestrate all three: to design novel biological systems from scratch, predicting how edited genes will produce specific proteins that fold into specific shapes that perform specific functions within living cells.

An RSI-capable AI fills that missing piece. It connects the gene-editing tools, the structural predictions, and the synthesis pipelines into a single design loop. The implications start strange and get stranger.

Microbes engineered to eat specific pollutants. Not a hypothetical: researchers have already modified bacteria to consume plastics, but the process is slow and unreliable because designing the metabolic pathways is hard. An AI that can model entire metabolic networks could design organisms optimized for specific environmental cleanup tasks, plastic in the ocean, PFAS in groundwater, carbon in the atmosphere.

Immune cells reprogrammed to hunt specific cancers. CAR-T therapy already does a crude version of this, engineering a patient's T cells to recognize cancer markers. But current CAR-T design is expensive and limited. An AI that understands protein folding, cell signaling, and immune evasion simultaneously could design immune cells that are orders of magnitude more targeted and effective.

Living structures grown rather than built. Researchers at MIT and elsewhere have experimented with engineered living materials, including bacteria that produce structural proteins on command. Scaling this from lab curiosity to practical construction material requires the kind of cross-domain biological design that only becomes feasible with AI that operates at the intersection of genomics, structural biology, and materials science.

Nobody associates AI with biology. That is precisely why this matters. The public conversation about AGI focuses on chatbots, job displacement, and military applications. The biological implications are less visible but potentially more consequential.

What Else Is Coming

And that is just two examples. Multi-domain AGI unlocks acceleration on every front. Technologies that have lived in science fiction for decades, a universal translator, an invisibility suit, full-body immersion VR, direct brain-to-computer interfaces, all race closer to the present than they have ever been. Future Shock will be publishing analysis and predictions on when each of these technologies could arrive, grounded in the same precondition tracking and signal monitoring we use for our AGI predictions. Subscribe to the newsletter to follow along.

When Does This Happen?

Nobody agrees.

The expert spectrum runs from Elon Musk's "by year-end" to Yann LeCun's "the concept makes absolutely no sense." Prediction markets split the difference: Metaculus forecasters place "weak AGI" at late 2026 to early 2027 and "strong AGI" (including robotics and physical tasks) at 2033. The two numbers have been moving in opposite directions, with weak AGI pulling closer and strong AGI pushing further out, after 2025 disappointed expectations for agentic capabilities.

Notably, 2025 saw the first outward shift in AGI timelines since 2020. The Metaculus median for strong AGI moved from 2031 to 2033. Several CEOs quietly shifted their language from "AGI" to "superintelligence," which may reflect genuine progress or convenient goal-post relocation.

The skeptics deserve more than a footnote here. Gary Marcus, the most prominent AGI critic, argued that 16 of his 17 predictions for 2025 came true, including the prediction that AGI would not arrive. Yann LeCun, Meta's chief AI scientist, has repeatedly argued that current architectures lack the capacity for real understanding and that the very concept of AGI "makes absolutely no sense" as commonly discussed. Their core argument: scaling has limits, current models simulate competence without genuine comprehension, and the gap between impressive demos and reliable real-world performance is wider than the hype suggests. They may be right. The history of AI is littered with premature declarations of imminent breakthroughs, from the Dartmouth Conference in 1956 to the expert systems boom of the 1980s, each followed by a correction that took years to recover from.

We built a tracker because the disagreement itself is informative. If the experts cannot agree on when AGI arrives, the productive question is: what signals would tell us one side is right?

Future Shock tracks three concrete predictions with specific dates and resolution criteria:

1. Domain-Specific AGI (Software Engineering) by March 2027. An AI system autonomously builds and operates a production system with 1,000+ active users from a novel spec, handling bugs, scaling, security, and incident response. We give this 75% confidence.

2. Recursive Self-Improvement by July 2027. An AI system completes three or more autonomous improvement cycles with a measurable 5%+ gain each, with no human intervention between cycles. We give this 50% confidence.

3. Multi-Domain AGI by October 2027. A single AI system demonstrates domain-specific AGI across three or more fundamentally different domains. We give this 38% confidence.

Each prediction is monitored through a five-signal ensemble: benchmark performance, autonomous task duration (via METR evaluations), expert forecast movement, market probability shifts, and real-world deployment milestones. When the signals converge, we update our confidence scores.

The full methodology is at future-shock.ai/research/agi-predictions. The live prediction scoreboard, tracking these and dozens of other claims from major figures in AI, is at future-shock.ai/predictions.

What This Means

The honest framing is this: nobody knows exactly when AGI arrives, but the range of serious estimates has compressed from "sometime this century" to "sometime this decade." The money being spent, the government actions being taken, and the capabilities already demonstrated all point in the same direction. The world five years from now will look different from the world today in ways that are difficult to overstate and impossible to predict with confidence.

That is not doom. It is also not utopia. It is a transition, and transitions reward people who pay attention. A radiologist today who learns to work alongside AI diagnostic tools will practice medicine for decades. One who ignores them will find, within a few years, that the standard of care has moved on without them. The same logic applies to software engineers, financial analysts, researchers, and anyone whose work involves pattern recognition at scale. The question is not whether your field will be affected. It is whether you will see it coming.

The Anthropic ban and OpenAI's Pentagon deal both landed on February 27, 2026, within hours of each other. By the time most people read about one, the other had already happened. The pace is the point. This is a domain where falling behind by a few months means falling behind permanently, and that is true for nations, companies, and individuals.

We built Future Shock because the information gap is itself a form of power asymmetry. The people making decisions about AI, in boardrooms and in government, have access to briefings and internal research that the rest of us do not. Our goal is to close that gap. We track the predictions, source the data, and show our work, so that anyone paying attention can form their own view of what is coming and when.

If that sounds useful, subscribe to the newsletter. We publish daily, and we do not waste your time.