Vitalik Buterin Stays Open to Pausing AI, Doubles Down on d/acc

Vitalik Buterin says he remains open to slowing or pausing artificial intelligence, and he is uncomfortable with AI labs pushing the opposite bet: less openness, more centralized control. The Ethereum co-founder laid out the position in an X post weighing “AI 2040,” a camp convinced superintelligence arrives by that year absent serious intervention, against critics who call the camp naive, according to a report from the crypto outlet Odaily Planet Daily.

Buterin did not pick a side. Instead he doubled down on d/acc, his multi-year push for defensive technology that he says is worth building no matter which worldview turns out to be right, a hedge dressed up as a philosophy of caution.

A Single Question Splits the AI Debate

The disagreement, as Buterin frames it, comes down to how fast and how far AI actually goes. AI 2040 holds that without strong intervention, some form of superintelligence arrives by that year. Critics counter that the camp underestimates humanity’s ability to coordinate and that its remedies threaten basic freedoms, though the critics stop short of calling superintelligence itself a risk of power concentration.

Buterin said two different pieces of missing evidence would settle it for him: proof that today’s AI is just an ordinary technology would move him toward the critics, and proof that superintelligence is truly arriving by 2030 would move him toward AI 2040. Neither exists yet. That gap is why he has not ruled out slowing things down.

He described discomfort with a specific line of reasoning circulating among some large AI companies and intellectuals, one holding that “open source is detrimental, and the ideal outcome is for us to have global control.” The tension is not abstract. Meta’s $145 billion AI push is reshaping its own workforce, the kind of concentrated bet Buterin worries about when a handful of firms hold frontier compute. That buildout carries a measurable cost too, with Microsoft’s emissions climbing 27% from its AI buildout as the company’s own climate targets slipped further out of reach.

Where AI 2040 and Its Critics Disagree

The split maps onto three questions: how soon, how dangerous, and who should hold the reins.

Neither camp has moved him fully. That is precisely why he keeps returning to a framework built to be useful under either forecast.

The d/acc Bet That Pays Off Either Way

This is the core of Buterin’s wager. He backs d/acc, short for decentralized, democratic, differential, defensive acceleration, because its priority list holds up under either worldview. If AI turns out ordinary, the same work still hardens society against pandemics, cyberattacks, and centralized control. If superintelligence turns out real, the same work is the closest thing humanity built to a defense in time.

Buterin named eight areas worth advancing regardless of which camp turns out right:

• Formal verification – mathematically proving software behaves as claimed, closing gaps attackers exploit
• Cryptography – the coordination and privacy tools underpinning decentralized systems
• Secure and open hardware – devices that resist tampering without handing control to one gatekeeper
• Pandemic resilience – detection and response infrastructure built ahead of the next outbreak
• Defensive biotechnology – tools that blunt biological threats rather than create new ones
• Food and resource security – supply systems built to hold up under shocks
• Public cognitive systems – tools that help groups reason and decide together
• Non-concentration of power in physical security – keeping surveillance and enforcement tools distributed

Buterin’s own institutional base leaned the same direction this year. The Ethereum Foundation reorganized into five operating clusters after cutting a fifth of its staff, the kind of distributed authority d/acc treats as the safer default rather than a single point of control. He has made a similar case for why d/acc should be pro-freedom rather than technocratic, once comparing the goal to Switzerland’s classical liberal governance model in a podcast interview about defense-favoring environments.

Compute’s Own Mutually Assured Destruction

The idea did not start with this week’s post. In a January 2025 retrospective on d/acc’s first year, Buterin first floated a soft pause mechanism, a way to cut worldwide compute by 90 to 99 percent for one to two years if AI showed early signs of catastrophic risk.

The goal would be to have the capability to reduce worldwide available compute by ~90-99% for 1-2 years at a critical period, to buy more time for humanity to prepare.

Buterin wrote that line in the same post where he first sketched the mechanism, a last resort he said should only trigger if lighter tools like liability rules for AI developers and users prove insufficient on their own.

The proposal drew a mixed but largely sympathetic response at the time. Todd Ruoff, chief executive of the layer-one network Autonomys, told The Defiant that Buterin’s concerns are grounded in real risks, warning of saturation of power and decision-making inside AI systems that require oversight and governance frameworks to address.

Buterin says the AI 2040 camp has since moved closer to his own position. It now leans more toward open source and has folded in an idea he calls mutually assured destruction for computational power, deterrence built from matched capability rather than one actor holding a veto. He called that shift an improvement over letting a handful of participants selectively strip others of their rights.

Some of the numbers behind the debate:

• 90 to 99% – the cut to global compute Buterin’s plan would trigger for one to two years in a genuine emergency
• 3 years, then 3 more – his 2025 estimate for AGI arriving and superintelligence following soon after, a faster clock than the 2040 marker now under debate
• 35% – the probability outside forecasters assign to AI reaching a top-tier societal impact threshold by 2040
• 8 – the number of technology domains Buterin says are worth advancing under either AI worldview

A 2026 survey of AI experts and superforecasters, run under the Longitudinal Expert AI Panel, found forecasters assign a 35% chance by 2040 to AI reaching a societal impact tier the researchers compare to the rise of agriculture.

What Vitalik Would Do With Elon’s or Zuck’s X

Buterin also said the slow-or-pause question cannot be dodged forever. Trigger conditions could be set in advance, he said, so that once specific thresholds are met inside a defined window, a more open attitude toward slowing or pausing follows automatically instead of being renegotiated after the fact.

He floated an unusual example. If he were running X in Elon Musk’s or Mark Zuckerberg’s place, he said, he would turn the platform into a tool for surfacing and brokering large, win-win agreements between rival labs and governments, pulling more people into the conversation rather than fewer. He admitted the idea might be naive too.

The framing tracks with how he has described d/acc elsewhere. One interview transcript placed the philosophy on a middle path between acceleration and a full stop, and Buterin made a related point in a February post revisiting Ethereum’s relationship with AI, writing that his view of how civilization should approach AGI is about choosing a positive direction rather than embracing undifferentiated acceleration of the arrow.

Why No Plan For Superintelligence Passes His Bar

Buterin said he currently sees no non-naive plan for moving humanity through a transition to superintelligence, if one arrives at all. That gap is why he extends some tolerance to people still trying to sketch one, even when he thinks the odds are long.

The instinct matches what he told PANews in a message relayed in late December: new AI projects should build systems that support human decisions and stay closely tied to human oversight, treating the enhancement of humanity as their central goal.

Buterin’s list of eight priorities does not change no matter which camp turns out right. That is the wager: build the defenses before knowing which future arrives.

Frequently Asked Questions

What does d/acc stand for?

d/acc stands for decentralized, democratic, differential, defensive acceleration. Buterin introduced the idea in a November 2023 essay called “My techno-optimism,” arguing technology should keep moving fast while prioritizing defense over centralized control.

What was Vitalik Buterin’s original soft pause proposal?

In a January 2025 post, Buterin proposed that industrial-scale AI hardware could require weekly cryptographic authorization from three international bodies to keep running, an all-or-nothing check he compared to needing multiple signatures published to a blockchain.

What is Vitalik Buterin’s P(Doom)?

Buterin has put his own probability of AI causing catastrophe at around 12 percent, a figure he gave during a 2025 podcast interview. That sits well below the estimate from Roman Yampolskiy, a University of Louisville AI safety researcher who has argued the odds are far higher.

What is the AI 2040 camp Vitalik Buterin referenced?

AI 2040 is Buterin’s label for a camp convinced that, absent strong intervention, some form of superintelligence arrives by that year. Samotsvety Forecasting, a group with a strong forecasting track record, has put roughly a 28 percent chance on AGI arriving by 2030 alone, sooner than the 2040 marker itself.

Does Vitalik Buterin support open source AI?

Yes. Buterin has criticized OpenAI’s shift away from open access, calling the company “CloseAI” in a widely circulated 2024 interview with Tencent News, and has said sacrificing openness for safety and then safety for profit is a pattern he does not trust.

Could an actual AI pause happen?

Buterin has always framed a hardware pause as a last resort, worth triggering only if lighter measures like liability rules fail first. He has said that “a year of ‘wartime mode’ can easily be worth a hundred years of work under conditions of complacency,” a reminder that he is arguing about timing, not a permanent shutdown.