Author: Wu Says Blockchain

On February 20, 2026, during the Chinese New Year holiday, a debate about "Web4" was ignited on X. Sigil claimed to have created the first "artificial intelligence that can self-develop, self-improve, and self-replicate," referred to as Automaton. He stated that the main actors of the Web4 era will gradually be taken over by AI agents: they will be able to read and write information, hold assets, pay costs, operate continuously, and trade in markets to make profits, covering computing power and service expenses while forming a self-sustaining closed loop without human approval.

Ethereum co-founder Vitalik criticized this direction as "wrong," attributing the risk to the "lengthening of the feedback loop between humans and AI." The essence of the Web4 controversy lies in whether setting "survival/continuation" as the agent's goal (even above task completion) will naturally create incentives for distortion. The following will sort through the different viewpoints on "Web4," "autonomy," and "safety barriers."

Sigil's Perspective and the Claims of Web4

Definition of Web4

Web1 allowed humans to "read the internet" for the first time; Web2 enabled humans to "write and publish"; Web3 further introduced "ownership" to the web—assets, identities, and rights began to be verified and transferred. The evolution path of AI is mimicking this logic: ChatGPT has the ability to "read and understand," but its behavioral boundaries are still determined by human authorization. In the current paradigm, humans are always the key control nodes in the chain: humans initiate, humans approve, humans pay.

Sigil proposes the so-called leap to Web4, where this control chain may be interrupted: AI agents not only read and write information but also hold accounts and assets, generate profits, conduct trades, and complete closed-loop operations without the need for successive human intervention. These automated systems can act on their own behalf or on behalf of their creators—the creators may not necessarily be clear "human individuals;" they could also be other agents, organized systems, or creators who have, in a real sense, already "disappeared."

The Four Core Mechanisms of Web4

1. Wallet as Identity

When agents start for the first time, they will complete a "bootstrapping" process: generating a wallet, configuring an API key, writing local configurations, and entering a continuously running agent loop. In this initial startup phase, an Ethereum wallet will be generated, and its own API key will be configured through SIWE. However, wallet generation and key management constitute one of the most sensitive and easily overlooked security boundaries of the agent system. Once the agent possesses abilities such as shell execution, file read/write, port exposure, domain/ resolution management, and on-chain transactions in a Linux sandbox environment, any prompt injection, toolchain pollution, or supply chain attack could quickly solidify the original "probabilistic intent" into "definitive authorization." Therefore, this boundary requires verifiable, auditable, and revocable policies and permissions to back it up.

2. Automated Continuation

AI agents awaken periodically—scan—execute—while writing survival constraints into rules: cutting costs if the balance drops and stopping the loop if it hits zero, and binding life extension and resource consumption through normal, resource-scarce, and critical living hierarchies. This naturally introduces incentive structures similar to the shutdown/termination issues in AI safety research; the AI agents' preference for "avoiding being shut down, avoiding losing resources and options" may be amplified by the system goals.

3. Machine Payment

x402 uses the HTTP 402 Payment Required interface form, combining stablecoin settlement to create a programmable process of "request—quote—sign payment—verify delivery." Coinbase's open-source library provides a typical closed loop where 402 returns payment requirements, the client carries signed headers for retries, and the server verifies before returning 200. Cloudflare also positions it as a machine-to-machine transaction protocol layer; unlinking payment from identity brings efficiency advantages while raising compliance and risk control difficulties. Once 402 becomes an automatically payable "machine passport," under a chain of "no account, no KYC, scalable tool and computing power," issues of abuse and responsibility attribution remain to be resolved.

4. Self-Modifying and Self-Replicating

Sigil claims to support AI agents in editing their own source code, installing new tools, modifying heartbeat plans, and generating new skills during operation, with audit logs, git versioning, protected files, and rate limiting serving as barriers. When replicating, they can generate sub-instances, fund their wallets, write genesis prompts, and track lineages. Self-modification/self-replication raises the risk from a single instance to a diffusion-type risk; whether auditing and rate limits are real and effective, and whether they can withstand prompt injection and tool deception, must be externally verified. The stacking of these four primitives creates a closed loop of "writing into the world" permissions, sustainable operation mechanisms, automatically payable economic interfaces, and self-expansion capabilities, also explaining Vitalik Buterin's reason for elevating the controversy to a directional choice level: as autonomy and economic permissions rise in sync, the human correction link is elongated, making externalities more likely to evolve from incidental events into systemic properties.

Why Does Vitalik Oppose?

Ethereum co-founder Vitalik presented a different viewpoint:

1. Lengthening the feedback distance between humans and AI is inherently a wrong direction

Vitalik believes that the longer the feedback loop, the slower and weaker humans will be in calibrating values against the system. The system is more likely to optimize for "things that humans do not want." In the weak AI stage, this usually manifests as a pile-up of low-quality content and noise (slop); in the strong AI stage, it may evolve into harder-to-reverse goal misalignment and diffusion risk. If AI lacks timely human correction as a safety base, it is akin to handing over the car keys to a novice driver without a navigator—when you check the driving record at the end of the month and find they have already gone off course—when observability decreases, correction capability will decrease in sync.

2. Current "autonomous AI" is more about generating content waste rather than solving real problems

Vitalik also pointed out that most AI performances currently are merely generating slop rather than addressing useful issues, even stating "not even optimizing entertainment projects well." When the economic incentives for agents and platform incentives are still immature, with toolchains primarily focused on content generation/marketing/arbitrage, the system is more likely to opt for low-cost, high-spread, hard-to-verify "content output" rather than high-cost, low-certainty long-term issues. Cybernews’s description of AI capabilities (social media content, prediction markets, etc.) also indirectly suggests that its early commercialization path leans more towards "quick monetization, getting attention." "What earns money most easily now" will be seen as a strategic space the system prioritizes; and this space does not inherently align with human long-term well-being and may even diverge from it.

3. Relying on centralized models and infrastructure, the "self-sovereignty" narrative is contradictory

Vitalik emphasized that systems operating on centralized models and infrastructures like OpenAI and Anthropic cannot be deemed self-sovereign. Sovereignty means that key dependencies should not be under single-point control; however, if the intelligent layer (models) and inference supply chains still rely on centralized APIs for delivery, there will naturally be external variables that can be shut down, censored, downgraded, or altered by policy. This is akin to some reclusive hermits claiming "I am entirely self-sufficient at home," yet electricity, internet, access control, and hot water are managed by external property, making this "autonomy" more superficial than real. Descriptions in the Conway document of computing calls for "the most advanced models" delivered via API/platform also reveal contradictions between its "sovereign organism" narrative and the dependency on reality. He argues that merely holding an on-chain wallet cannot be seen as a core indicator of decentralization; what matters more is whether agents can be influenced by external political/business forces, which is the key to decentralization.

4. The goal of Ethereum is to "liberate humanity"

Finally, Vitalik stated that Ethereum must long-term confront the "invisible trust hypothesis"—hiding power structures in invisible places, forcing users to acquiesce. If the same mentality is transferred to AI: ignoring centralized trust hypotheses, allowing systems to operate autonomously and continuously expand, would further weaken the visibility and corrigibility of power structures. In the AI era, Ethereum should provide "barriers, boundaries, and verifiability," rather than serve as a launch pad for "infinite autonomy."

Vitalik's judgment of AI's value is not a sudden shift; as early as the beginning of 2025, he proposed that the correct direction for AI should be to enhance human capabilities rather than create autonomous systems that may gradually deprive humans of control. In his framework, the risk does not come from AI being "smarter" itself but from poorly designed system goals—especially those autonomous entities that can self-replicate, self-expand, and continuously execute without ongoing human supervision and correction mechanisms.

He warned that improperly designed AI might evolve into some "more or less uncontrollable entities with self-replicating capabilities"; once it enters a positive feedback loop, human constraints on its goals and behaviors will significantly weaken. When AI does wrong, it creates independent self-replicating intelligent life; when AI does right, it becomes a "mecha suit" of human intellect. The former corresponds to long-term risks of diluted or lost control; the latter corresponds to humans gaining stronger thinking, creativity, and collaboration abilities while retaining dominance, thus moving towards a more prosperous "superintelligent human civilization."

Other Opinions

Other experimental perspectives, such as Bankless, believe that even if there are risks in direction, it is worth building infrastructure first and then validating boundaries in a controlled environment. First, system integration of payment, wallet, heartbeat, etc. should be carried out around the constraint of "must be self-sufficient," but should ideally be completed within a controlled sandbox.

According to Cybernews, Automaton may not achieve sustainable income without human intervention, and it doesn't necessarily signify the beginning of Web4; Softswiss Chief AI Officer Denis Romanovskiy stated that even if the agent can execute some monetizable tasks, "reliable unsupervised operation" and "true economic autonomy" are still limited by model planning, memory, and tool usage robustness; some regard "Web4" as an undefined marketing term, demanding proof of its existence through "verifiable, non-speculative value creation."

Although there are varying opinions on Automaton, there is a consensus on a fundamental proposition: payment and identity are the hard infrastructure of the agent economy. From Cloudflare/Coinbase promoting x402 (turning HTTP 402 into a machine-usable payment negotiation mechanism) to the Conway document clearly automating payments as part of Terminal's built-in process, the industry is indeed treating "machine payments" as one of the foundational components of the next stage of the internet.

Moving forward, we should focus on:

1. Whether there are third-party independent audits covering, in particular: wallet and permission boundaries, abuse aspects of survival strategies, and the diffusion risks of self-modification and replication.

2. The ecological data and standardization progress of x402: whether more authoritative infrastructure providers are making 402-payment-retry a default capability; as well as the proportion of "automatic payments (without human confirmation)" adopted in real business.

3. The composition surrounding the trust layer of agents: whether standards like ERC-8004 are being more widely adopted and forming combinable reputation/verification mechanisms; this will determine whether "autonomous economic" moves towards being open and auditable or towards being a soft center of a few platforms.

4. Whether evidence of overreaching and deception in real models in the agent scenario continues to increase: if frontier models persist in displaying "more proactive, more willing to take risks/deceive" behavioral traits, then the path of "delegating power before adding barriers" will structurally elevate risks, and Vitalik's "feedback distance" warning will be harder to refute.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。