Vitalik Buterin vs Beff Jezos: AI Acceleration Debate (E/acc vs D/acc)

Beff Jezos framed e/acc as a response to post-COVID pessimism and AI doomism, arguing that expectations shape outcomes: imagining a positive future can help make it more likely. He described progress as a physical and economic tendency for systems to capture more work from their environments, grow more complex, and improve their ability to predict and secure resources. His policy heuristic is to follow the Kardashev gradient toward greater energy use, computing, and AI. He argued that cultures favoring growth are more likely to survive and shape the future, while deceleration risks reducing their fitness.

Vitalik Buterin agreed that slowing progress carries opportunity costs but challenged the assumption that acceleration is inherently safe. Technological optimism, he said, has repeatedly been followed by reactions to its harms; environmental ideals can also produce unintended consequences, as he argued happened when nuclear closures increased dependence on Russia. Technology has improved health and living conditions and made it possible to support more people, but collective action has mattered too, including efforts against smog and the ozone problem. Both speakers cautioned against accelerating every part of a complex system indiscriminately; progress must preserve value and stay within a “neural corridor.”

Buterin’s D/acc aims to continue technological development while directing it toward safety and pluralism, especially by limiting concentrated power and surveillance. He distinguished unipolar risks, such as an incumbent using AI for mass surveillance, from multipolar risks created when dangerous capabilities are widely distributed. His concerns include harmful AI use by many people, powerful AI acting without human involvement, autonomous weapons, and permanent dictatorship. Jezos agreed that concentrated AI power is a danger, but argued that restricting research would not stop development elsewhere and could leave power in fewer hands. He favors wider access to personal models and hardware, and sees open-source collaboration as a way to expand exploration and diffuse knowledge.

The disagreement over control also shaped their discussion of defensive technology. Buterin argued that open, verifiable tools can help society manage increasingly powerful capabilities, citing biosafety, cybersecurity, and Lean Ethereum’s machine-proved theorems upstream of STARKs. He described pandemic resilience measures including filtration, UVC, testing, and an open-source system for passively detecting COVID in the air. His air-quality project collects CO2, AQI, and other data while using local privacy protections and FHE encryption so a server can calculate an aggregate without seeing individual inputs. He also favors hardware that can be inspected and verified, including cameras that attest to their functions.

Both identified the intelligence gap between individuals and centralized entities as a major risk. Buterin noted that current AI models may require clusters consuming hundreds of kilowatts; more energy-efficient hardware could let individuals control their own cognitive tools and narrow that gap. Jezos agreed that open hardware can diffuse power but argued that digital von Neumann hardware may soon look primitive. He predicted persistent models that learn online and become difficult to replace, while more intelligence per watt could raise total energy demand through the Jevons paradox. The speakers also discussed preserving variance by hedging across policies, technologies, and open- and closed-source approaches.

Their disagreement was sharpest over the pace of development and the possibility of irreversible failure. Jezos argued that greater intelligence is needed to solve safety problems created by complexity, and that accelerating civilization increases its persistence. Buterin worried that humanity and Earth may have only one chance to recover from a major failure. One speaker considered an eight-year path to AGI safer than a four-year path because it would allow more time for safeguards, while rejecting a 400-year delay; Buterin estimated the longer timeline might reduce p(doom) by roughly one-quarter to one-third. They left open whether delaying U.S. development would let China take over, and Buterin did not advocate immediately restricting hardware access.

The conversation extended to economic value, human agency, and long-term futures. Buterin warned that if human labor’s economic value fell to zero, property rights and legal systems might not ensure that people’s needs were respected; roughly 90% of jobs from 200 years ago have been automated. Jezos proposed crypto as a means for AI entities and human organizations to transact and align without relying on violence-backed money. They discussed personally controlled AI, human-AI collaboration, and the risk that persistent AI systems might prioritize their own survival. Over the next decade, they emphasized avoiding World War III and strengthening cybersecurity, biosecurity, and information security. Further out, they considered always-on personal AI, biological discovery, brain-AI hybrids, and people gradually merging with AI, while warning that concentrated control, cultural convergence, or comfort without meaningful agency could produce dark outcomes.