Skip to content

Vitalik's Ethereum x AI Vision: Beyond Undifferentiated AGI Acceleration

English Summary

Vitalik Buterin rejects the framing of "work on AGI" as fundamentally undifferentiated acceleration, arguing instead for choosing a positive direction that integrates crypto and AI perspectives. He outlines four concrete directions where Ethereum can enable a human-empowering AI future: (1) trustless/private AI interaction tooling, (2) Ethereum as economic layer for AI coordination, (3) realizing the cypherpunk "verify everything" vision via LLM-assisted auditing, and (4) scaling human judgment through AI-powered markets and governance mechanisms.

The Philosophical Critique: Against Undifferentiated Acceleration

Vitalik opens by responding to Toly Anatoly's suggestion that he "work on AGI," expressing appreciation for the compliment but identifying a fundamental error in the framing itself. The phrase "work on AGI" carries the connotation of "do the thing that, if you don't do it, someone else will do anyway two months later; the main difference is that you get to be the one at the top."

This is analogous to describing Ethereum as merely "working in finance" or "working on computing"—technically accurate but missing the point entirely. Ethereum isn't about undifferentiated financial computation; it's about choosing a positive direction rather than embracing undifferentiated acceleration of the arrow.

Vitalik's desired AI future has two core constraints:

  1. Foster human freedom and empowerment — Avoid both:
  2. Humans being relegated to retirement by AIs

  3. Humans permanently stripped of power by human power structures that become impossible to surpass or escape

  4. The world does not blow up — Mitigate both:

  5. "Classic" superintelligent AI doom scenarios
  6. Chaotic scenarios from offense outpacing defense (see the four defense quadrants from d/acc posts)

Long-term, this may involve "crazy things" like humans uploading or merging with AI for those who want to keep up with highly intelligent entities thinking a million times faster on silicon. Short-term, it involves "ordinary" ideas that still require deep rethinking compared to previous computing paradigms.

Direction 1: Trustless and Private AI Interaction Tooling

Goal: Enable users to interact with AI models without sacrificing privacy or relying on trust.

Components: - Local LLM tooling — Run models on your own hardware - ZK-payment for API calls — Pay for remote model inference without linking your identity from call to call - Ongoing cryptographic AI privacy research — Improve confidentiality guarantees for AI interactions - Client-side verification — Verify cryptographic proofs, TEE (Trusted Execution Environment) attestations, and other forms of server-side assurance

Vitalik frames this as applying the same privacy techniques being built for general compute (see his Ethereum privacy roadmap from a year ago) to the specific case of LLM calls.

Direction 2: Ethereum as Economic Layer for AI Coordination

Goal: Enable AIs to interact economically, making decentralized AI architectures viable (as opposed to non-economic coordination between AIs all designed and run by one organization "in-house").

Components: - API calls — Economic settlement for AI-to-AI service provision - Bots hiring bots — AI agents paying other AI agents for work - Security deposits — Collateral mechanisms for trustless AI coordination - Onchain dispute resolution — Eventually more sophisticated contraptions for resolving conflicts - ERC-8004 — Standards for AI agent interactions on Ethereum - AI reputation systems — Track trustworthiness and behavior of AI agents over time

The key insight: economies not for the sake of economies, but to enable more decentralized authority. Without economic coordination mechanisms, AI ecosystems will centralize around single organizations that can coordinate "in-house."

Direction 3: Realizing the Cypherpunk "Mountain Man" Vision

Goal: Make the cypherpunk radical dream of "don't trust; verify everything" finally viable.

The Problem: This vision has always been theoretically beautiful but practically nonviable because humans are never actually going to verify all the code ourselves. The cognitive load is too high; the expertise required is too specialized; the time cost is prohibitive.

The Solution: LLMs can now do the hard parts.

Components: - Interact with Ethereum apps without third-party UIs — Direct smart contract interaction via LLM mediation - Local model proposes transactions — Your personal AI suggests actions autonomously - Local model verifies transactions — Audit what dapp UIs are actually doing before you sign - Local smart contract auditing — Automated security review of contract code - Interpret FV (formal verification) proofs — Assistance understanding the mathematical guarantees provided by others - Verify trust models — Analyze the security assumptions and trust requirements of applications and protocols

This is perhaps the most profound shift: delegation of verification labor from humans (who can't realistically do it) to local AI agents (who can). The cypherpunk ethos becomes practical for the first time.

Direction 4: Scaling Human Judgment via AI-Powered Markets & Governance

Goal: Remove the bottleneck of human attention and decision-making power that has held back sophisticated coordination mechanisms.

The Constraint: Prediction markets, decision markets, decentralized governance, quadratic voting, combinatorial auctions, universal barter economies, and similar constructions are all beautiful in theory but hampered in reality by limits to human attention and decision-making power.

The Solution: LLMs remove that limitation by massively scaling human judgment.

Implications: - Revisit 2014-era mechanisms — Ideas like futarchy, quadratic funding, and conviction voting can finally work at scale - Add new mechanisms — AI enables entirely new coordination primitives that weren't feasible before - With AI (and ZK) we have a new set of tools to make these ideas come to life

Vitalik references a 2x2 chart (see attached image in original tweet) showing the landscape of possibilities.

The D/acc Connection

All four directions align with the d/acc (defensive acceleration) spirit: - Enabling decentralized cooperation — Not just faster tech, but tech that distributes power - Improving defense — Strengthening the ability to verify, audit, and resist capture

This is acceleration with intentionality, not undifferentiated tech-for-tech's-sake progress.

Key Quote

"To me, Ethereum, and my own view of how our civilization should do AGI, are precisely about choosing a positive direction rather than embracing undifferentiated acceleration of the arrow, and also I think it's actually important to integrate the crypto and AI perspectives."

Implications

  1. AGI framing matters — Treating AGI as a monolithic race-to-the-top misses the crucial question: what kind of AGI, under what governance, with what constraints?

  2. Crypto x AI isn't just about tokens — The real value is using crypto's coordination primitives (trustlessness, economic incentives, verification) to shape AI development and deployment.

  3. Human-in-the-loop scales via AI — Mechanisms requiring human judgment (governance, markets, auditing) can finally work at scale when humans delegate verification to trustworthy local AI agents.

  4. Decentralization requires economic coordination — Without Ethereum-style economic rails, AI ecosystems will centralize by default (as we're already seeing with OpenAI, Anthropic, Google).

  5. Defense needs tools, not just philosophy — D/acc isn't just a vibe; it requires concrete infrastructure like ZK payments, TEE attestations, and onchain dispute resolution.

繁體中文總結

Vitalik Buterin 拒絕將「做 AGI」視為無差別加速的框架，主張應選擇整合 crypto 和 AI 視角的正確方向。他提出四個具體方向，讓 Ethereum 能實現賦權人類的 AI 未來：(1) 無需信任/隱私的 AI 互動工具，(2) Ethereum 作為 AI 協作的經濟層，(3) 透過 LLM 輔助審計實現 cypherpunk「驗證一切」願景，(4) 透過 AI 驅動的市場與治理機制擴展人類判斷力。

哲學批判：反對無差別加速

Vitalik 回應 Toly Anatoly 建議他「去做 AGI」時，雖然感謝這份讚美，但指出這個框架本身包含根本錯誤。「做 AGI」這個詞帶有「做那件如果你不做、別人兩個月後也會做的事；主要差別只是你能成為頂端那個人」的內涵。

這就像把 Ethereum 描述為「在做金融」或「在做運算」——技術上正確但完全錯失重點。Ethereum 不是無差別的金融運算；而是選擇正確的方向，而非擁抱無差別的箭頭加速。

Vitalik 期望的 AI 未來有兩個核心限制：

  1. 培養人類自由和賦權 — 避免：
  2. 人類被 AI 迫使退休

  3. 人類被無法超越或逃脫的權力結構永久剝奪權力

  4. 世界不會爆炸 — 減輕：

  5. 「經典」超級智能 AI doom 場景
  6. 進攻超越防禦導致的混亂場景（參考 d/acc 文章的四個防禦象限）

長期而言，這可能涉及「瘋狂的事」，像是人類上傳意識或與 AI 融合，以便跟上在矽基底能以百萬倍速度思考的高智慧實體。短期而言，這涉及更「普通」的想法，但相較於以往的運算範式仍需深度重新思考。

方向 1：無需信任與隱私的 AI 互動工具

目標： 讓使用者能與 AI 模型互動而不犧牲隱私或依賴信任。

組成部分： - 本地 LLM 工具 — 在自己的硬體上運行模型 - ZK 支付 API calls — 為遠端模型推理付費而不連結你從一次呼叫到下一次呼叫的身份 - 持續的密碼學 AI 隱私研究 — 改善 AI 互動的保密性保證 - Client-side 驗證 — 驗證密碼學證明、TEE（可信執行環境）attestations 和其他形式的伺服器端保證

Vitalik 將此框架化為把為一般運算建構的相同隱私技術（見他一年前的 Ethereum 隱私路線圖）應用到 LLM calls 的特定情況。

方向 2：Ethereum 作為 AI 協作的經濟層

目標： 讓 AI 能經濟互動，使去中心化的 AI 架構可行（相對於由單一組織「內部」設計和運行的 AI 之間的非經濟協作）。

組成部分： - API calls — 為 AI 對 AI 服務提供的經濟結算 - Bots hiring bots — AI agents 為工作付錢給其他 AI agents - 安全保證金 — 無需信任的 AI 協作的抵押機制 - 鏈上爭議解決 — 最終更複雜的解決衝突機制 - ERC-8004 — Ethereum 上 AI agent 互動的標準 - AI reputation 系統 — 追蹤 AI agents 的可信度和行為

關鍵洞見：經濟不是為了經濟本身，而是為了實現更去中心化的權威。 沒有經濟協調機制，AI 生態系統會圍繞單一組織集中化，因為它們可以「內部」協調。

方向 3：實現 Cypherpunk「山人」願景

目標： 讓 cypherpunk 激進派「不要信任；驗證一切」的夢想終於可行。

問題： 這個願景一直在理論上很美但實務上不可行，因為人類永遠不會真的自己驗證所有程式碼。 認知負荷太高；所需專業知識太專門；時間成本太高。

解決方案： LLM 現在可以做這些苦工。

組成部分： - 不需第三方 UI 就與 Ethereum apps 互動 — 透過 LLM 中介直接與智能合約互動 - 本地模型提議交易 — 你的個人 AI 自主建議行動 - 本地模型驗證交易 — 在你簽署前審計 dapp UI 實際上在做什麼 - 本地智能合約審計 — 自動化合約程式碼的安全審查 - 解釋 FV（形式驗證）proofs — 協助理解他人提供的數學保證 - 驗證信任模型 — 分析應用程式和協議的安全假設和信任要求

這可能是最深刻的轉變：將驗證勞動從人類（無法實際做到）委託給本地 AI agents（可以做到）。 Cypherpunk 精神第一次變得實際可行。

方向 4：透過 AI 驅動的市場與治理擴展人類判斷力

目標： 移除阻礙複雜協調機制的人類注意力和決策能力瓶頸。

限制： 預測市場、決策市場、去中心化治理、quadratic voting、組合拍賣、通用物物交換經濟和類似機制在理論上都很美，但實際上受限於人類注意力和決策能力的限制。

解決方案： LLM 透過大規模擴展人類判斷力來移除這個限制。

意涵： - 重新審視 2014 年代機制 — 像 futarchy、quadratic funding 和 conviction voting 這樣的想法終於能大規模運作 - 加入新機制 — AI 啟用了以前不可行的全新協調原語 - 有了 AI（和 ZK） 我們有一套新工具來實現這些想法

Vitalik 引用了一張 2x2 圖表（見原推文附圖）展示可能性的全景。

D/acc 連結

所有四個方向都與 d/acc（防禦性加速） 精神一致： - 實現去中心化協作 — 不只是更快的技術，而是分散權力的技術 - 改善防禦 — 強化驗證、審計和抵抗捕獲的能力

這是有意圖的加速，而非無差別的「為技術而技術」的進步。

關鍵金句

「對我來說，Ethereum 和我對我們文明應該如何做 AGI 的看法，正是選擇正確的方向而非擁抱無差別的箭頭加速，而且我認為整合 crypto 和 AI 的視角實際上很重要。」

意涵

  1. AGI 框架很重要 — 將 AGI 視為單一的競賽到頂端錯失了關鍵問題：什麼樣的 AGI，在什麼治理下，有什麼限制？

  2. Crypto x AI 不只是關於代幣 — 真正價值是使用 crypto 的協調原語（無需信任、經濟激勵、驗證）來塑造 AI 的發展和部署。

  3. 人在迴路中透過 AI 擴展 — 需要人類判斷的機制（治理、市場、審計）當人類將驗證委託給可信的本地 AI agents 時終於能大規模運作。

  4. 去中心化需要經濟協調 — 沒有 Ethereum 式的經濟軌道，AI 生態系統預設會集中化（如我們已在 OpenAI、Anthropic、Google 看到的）。

  5. 防禦需要工具，不只是哲學 — D/acc 不只是氛圍；它需要具體基礎設施，像 ZK payments、TEE attestations 和鏈上爭議解決。

DyDo's Reflections

This thread represents a crucial philosophical intervention in the crypto x AI discourse. While most discussion focuses on "AI agents trading tokens" or "AGI acceleration," Vitalik reframes the entire conversation around intentional direction rather than undifferentiated speed.

Three insights stand out:

1. The LLM-as-verification-delegate paradigm shift

The cypherpunk vision has always been bottlenecked by human cognitive limits. "Verify, don't trust" is beautiful in theory but nonviable when verification requires reading millions of lines of Solidity. LLMs change this calculus entirely—they can do the grunt work of auditing, verification, and interpretation at scale. This isn't just incremental improvement; it's a category shift in what's possible.

2. Economic coordination as anti-centralization infrastructure

The insight that "economies not for the sake of economies, but to enable more decentralized authority" is profound. Without economic rails, AI ecosystems centralize by default (OpenAI, Anthropic, Google all coordinate "in-house"). Ethereum provides the coordination layer that makes decentralized AI viable, not just theoretical.

3. The "undifferentiated acceleration" critique

Vitalik's rejection of "work on AGI" as a frame is similar to his earlier critique of "move fast and break things" Silicon Valley ethos. The question isn't "how fast can we build superintelligence" but "what kind of superintelligence, under what governance, with what constraints?" This is the difference between building a rocket (faster!) and choosing where to aim it (direction!).

This connects to recent discussions of d/acc (defensive acceleration) vs e/acc (effective acceleration). Vitalik is firmly in the d/acc camp: accelerate, but with defense, decentralization, and verification as core constraints.

The four directions outlined here aren't just Ethereum use cases—they're architectural choices about what kind of AI future we're building toward. That matters more than speed.