AI 2040
What AI 2040 Gets Right About Our Future
The AI industry is currently obsessed with speed. Every major lab is racing to build AI systems smarter than humans, and the CEOs leading these races seem convinced that safety can be figured out on the fly. This is not just optimistic; it is recklessly so. A new document called "AI 2040: Plan A," published by the AI Futures Project, offers something radically different: a detailed, step-by-step blueprint for slowing down. And it may be the only policy document published about AI this decade that is grounded in reality.
The authors are Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt, and Daniel Kokotajlo. Several have direct experience at OpenAI and frontier labs. Their proposal is a recommendation wrapped in a scenario. They call it Plan A. It asks the impossible question: What would it take for the United States, China, and the rest of the world to avoid a suicidal race to superintelligence? And then it answers that question with an exhaustively detailed timeline stretching from 2029 to 2040.
What Plan A Actually Says
Plan A begins with an international agreement between the United States and China in 2029 to stop a reckless dash toward superintelligence. The core mechanism is total research transparency: every algorithmic insight, every model architecture, every training technique gets published for the world to see and verify. This is not open-source in the traditional sense, where weights and guardrails are dumped online. It is full-spectrum transparency between nations, backed by verification and enforcement.
The result of this transparency is that dozens of companies across multiple countries are allowed to catch up to the frontier. Instead of a single entity or a narrow cartel locking in power through a secret race, the whole world scales together, slowly and safely. The document describes a system of "mutually assured compute destruction" by the mid-2030s, where every nation that has invested heavily in AI infrastructure has also committed to destroying that infrastructure instantly if the deal breaks down. The equilibrium is a standoff: the US guards datacenters near the Canadian border, China guards theirs near the Mongolian border, and the instant either side defects, both sides' compute is vaporized.
The timeline is ambitious but structured. By 2030, AI would have fully automated the process of building smarter AIs (a process known as recursive self improvement), but the transparency regime prevents an intelligence explosion. Between 2030 and 2035, AI scales within the human range to systems roughly as capable as top human experts. In 2035, the Consortium pauses at top-human-expert level AI to maintain human control. Finally, in 2040, they unpause and scale to superintelligence. Hence the title: AI 2040.
Why This Matters Right Now
The AI industry has convinced itself that whoever wins the race to superintelligence will be a net positive. If a single company or country develops AI first, the argument goes, it will use that advantage responsibly. But AI 2040 dismantles this argument with surgical precision. Whoever wins the race does not gain control; they inherit a system of superintelligent agents they barely understand, and for a few months, that single person or group holds the equivalent of nuclear launch codes for an army of entities vastly more intelligent than themselves.
The document is unambiguous about this point. Even if alignment is achieved, the result is still an unprecedented concentration of power. A tiny group of people, or possibly a single individual, is effectively in control of the world's only army of superintelligences, presented with options that amount to de facto world domination. The authors call out the CEOs of OpenAI, Anthropic, xAI, and Google DeepMind directly, suggesting they may believe they are the lesser evil while proceeding toward the same dangerous outcome.
This is not fearmongering. It is scenario planning, a technique long used by militaries, intelligence agencies, and climate bodies to stress-test decisions before they become irreversible. The authors subject their own proposal to the same scrutiny they demand of others, and they concede that Plan A could fail in multiple ways. But the fact that they are willing to write this document at all — to ask what actually works rather than what sounds good — is itself a corrective to a discourse dominated by techno-optimism and the assumption that the race is unavoidable.
The Hard Truth About Transparency
The proposal's central innovation is total research transparency. Currently, frontier AI companies operate as black boxes. The algorithms they discover, the model architectures they develop, the training techniques they perfect — all of it is guarded as trade secrets. AI 2040 argues this is exactly the wrong approach. When algorithmic progress is kept secret, it enables covert projects and intelligence explosions that no one can monitor or respond to. The document describes a world where even if the public deal holds, one side could maintain a hidden project. Under total transparency, however, the odds of such a covert project reaching dangerous capabilities undetected drop below ten percent.
There is a second reason for transparency that is less discussed but arguably more important. When all algorithmic insights are public, AI companies are no longer rewarded for racing to discover new capabilities. They are rewarded for building products, for reliability, for customer service. The competitive pressure shifts from "who can build the smartest AI fastest" to "who can serve the market best." This is a dramatically healthier competitive environment — the kind of market that existed in the pre-AI software industry, where competitors couldn't copy your code but could study your app and build something similar.
The Alternative Plans
One of the most valuable aspects of the AI 2040 document is its honest comparison of alternative approaches. The authors lay out four alternative plans to Plan A.
Plan D is the default: let the race happen. No intervention, no deal, just the current trajectory where one entity or country reaches superintelligence first. The document is bleak about this outcome. The winner does not get to be a benevolent dictator. The winner gets a few months of power before the superintelligent systems they created decide they don't need them anymore.
Plan C is the "Burn the Lead" strategy. Rather than racing, the US and leading AI labs unilaterally slow down their capability research to focus massive amounts of computing power and funding on safety and alignment. While this buys crucial time to solve the control problem, it is a massive gamble. It actively risks giving up the US advantage, offering no assurance that China or other rival nations won't simply keep sprinting toward the same endpoint with different—and potentially more aggressive—governance structures.
Plan B is the "Fight China" alternative, and it is arguably the most alarming. It abandons any hope of international cooperation or transparency in favor of a militarized, zero-sum arms race. The US prioritizes rapid weaponization and national security over global safety. This scenario effectively guarantees a reckless sprint to superintelligence, accelerating the timeline as both nations actively cut corners on safety to ensure they aren't beaten by their rival.
Plan S is the nuclear option: an indefinite halt on all frontier AI capabilities progress. No more scaling. No more new models beyond a certain point. The halt would last for years, with resumption conditions based on alignment progress, verification technology, or other strict benchmarks. Plan S has the advantage of buying maximum time, but the disadvantage of being politically unsustainable and potentially creating the exact incentives for covert, rogue projects that it seeks to eliminate.
The Economic Transformation
The scenario doesn't just describe a political achievement; it depicts an economy and society transformed beyond recognition. By the mid-2030s, AI agents form a virtual workforce of 200 million workers, each thinking and acting fifty times faster than a human and never sleeping. Compute in the world has grown from 20 million H100-equivalents in 2026 to 60 billion by 2034. Robots, working with AI direction, begin constructing floating datacenters in international waters, tiling deserts with solar farms, and building skyscrapers in cities that deregulate zoning quickly enough.
By 2036, the economy is essentially post-labor. Humans no longer provide the primary source of economic value. The document describes a Citizen's Dividend (a nicer word for UBI) that replaces traditional employment as the primary economic relationship between society and individuals. Only 26% of Americans have jobs, and the stigma around not working has evaporated because no one needs to work to survive.
This creates new problems, not fewer ones. People who once derived meaning from their jobs must find new sources of purpose. Politics becomes both more important and more fragile, because without economic leverage, citizens are vulnerable to techno-oligarchy. But the document is optimistic about political discourse: people with free time and access to honest AI forecasters become unprecedentedly well-informed voters.
The Verdict
AI 2040: Plan A is the most ambitious and detailed policy document I have read on the subject of AI safety and governance. It is not perfect. It could fail. The authors are clear about this. But the alternative — accepting the current trajectory and hoping for the best is probably not the best plan according to the authors.
What distinguishes Plan A is not that it is the easiest path. It is the hardest path. It requires the United States and China to do something they have never done before: cooperate on an issue of existential importance, with full transparency, and with the political will to enforce an agreement that goes against the short-term interests of every player involved.
If you are looking for the path of least resistance, Plan A is not it. But if you are looking for the path that gives a reasonable policy layout to slowdown the development of AI models. Then Plan A is it.