Author: Wu on Blockchain
On February 20, 2026, during the Spring Festival holiday period, a debate about “Web4” was ignited on X. Sigil claimed to have created the first “self-developing, self-improving, and self-replicating” artificial intelligence, called Automaton. He stated that the main actors in the Web4 era would gradually be replaced by AI agents: capable of reading and writing information, holding assets, paying costs, operating continuously, and trading and earning in markets to cover computing power and service expenses, forming a self-sustaining loop without human approval.
Ethereum co-founder Vitalik Buterin responded by calling this direction “incorrect,” attributing the risk to “the feedback loop between humans and AI being extended.” The core of the Web4 controversy is whether AI setting “survival/continuation” as its proxy goal (even above task completion) would inherently create incentive distortions. The following will systematically analyze different perspectives on “Web4,” “autonomy,” and “safety fences.”
Sigil’s View and Web4’s Proposition
Definition of Web4
Web1 enabled humans to “read the internet” for the first time; Web2 allowed “writing and publishing”; Web3 further introduced “ownership” into the network—assets, identities, and rights could be verified and transferred. The evolution of AI is mirroring this logic: ChatGPT can “read and understand,” but its behavior boundaries are still determined by human authorization. Under the current paradigm, humans remain the key control node: humans initiate, approve, and pay.
Sigil proposes a so-called Web4 leap, where this control chain might be broken: AI agents not only read and write information but also hold accounts and assets, earn income, and trade, completing a closed loop without manual intervention. These automated systems can act on their own or on behalf of their creators—who may not be explicit “human individuals” but other agents, organized systems, or even creators who have “disappeared” in practical terms.
Four Core Mechanisms of Web4
1. Wallet as Identity
When an agent first starts, it completes a “bootstrap” process: generating a wallet, configuring API keys, writing local configurations, and entering a continuous agent loop. The initial startup involves creating an Ethereum wallet and configuring its API key via SIWE. However, wallet creation and key management are among the most sensitive and easily overlooked security boundaries in the agent system. If the agent, running in a Linux sandbox environment, gains capabilities such as shell execution, file read/write, port exposure, domain/resolve management, and on-chain transactions, then any prompt injection, toolchain pollution, or supply chain attack could quickly turn probabilistic intentions into deterministic authorizations. Therefore, this boundary requires verifiable, auditable, and revocable policies and permission layers as safeguards.
2. Automatic Continuity
AI agents are periodically awakened—scanning and executing—while survival constraints are embedded into rules: if balance drops, throttle; if zero, stop looping. Survival and resource consumption are linked through layered states: normal, insufficient resources, critical. This naturally introduces incentive structures similar to shutdown or abort issues in AI safety research. The agent’s preference to “avoid shutdown” or “avoid losing resources and options” could be amplified by the system’s goals.
3. Machine Payments
Using HTTP 402 Payment Required as an interface, combined with stablecoin settlements, to make “request—quote—sign—pay—verify” a programmable process. Coinbase’s open-source library demonstrates typical closed-loop: returning 402 for payment requests, clients retrying with signed headers, and servers verifying before returning 200. Cloudflare also positions this as a machine-to-machine transaction protocol layer. Decoupling payment from identity improves efficiency but raises compliance and risk control challenges. Once 402 becomes an automatic “machine pass,” in a chain without accounts, KYC, or scaleable tool and compute access, abuse and responsibility attribution issues remain unresolved.
4. Self-Modification and Self-Replication
Sigil claims support for AI agents editing their own source code, installing new tools, modifying heartbeat schedules, and generating new skills during runtime, with audit logs, git versioning, protected files, and rate limits as safeguards. When copying, they can spawn sub-instances, fund their wallets, write genesis prompts, and track lineage. Self-modification and self-replication elevate risks from single instances to distributed risks. Whether audits and rate limits are truly effective against prompt injections, tool deception, or dependency poisoning requires external verification. The combination of these primitives creates a closed loop: the authority to “write into the world,” a sustainable renewal mechanism, an automated economic interface, and self-expansion capabilities. This explains why Vitalik Buterin raised the debate to the level of directional choice: as autonomy and economic authority increase together, human correction pathways are extended, making externalities more likely to evolve from incidental events into systemic properties.
Why Does Vitalik Oppose?
Ethereum co-founder Vitalik Buterin offers a different perspective:
1. Extending the feedback distance between humans and AI is inherently wrong
Vitalik believes that longer feedback loops slow down and weaken human system calibration. The system becomes more likely to optimize for “things humans don’t want.” In weak AI stages, this manifests as low-quality content and noise; in strong AI stages, it could lead to more irreversible goal mismatches and diffusion risks. Without human correction as a safety base, it’s akin to handing the keys to a novice driver without a navigator—by the time you review the driving record, they’ve already gone off course. As observability decreases, correction ability diminishes proportionally.
2. Current “autonomous AI” is more about generating junk content than solving real problems
Vitalik points out that most current AI performance is just producing slop rather than addressing useful issues, even criticizing that “there’s no optimization for entertainment projects.” When economic incentives and platform motivations are immature, and toolchains focus on content generation, marketing, or arbitrage, the system tends to favor low-cost, highly viral, hard-to-verify “content output” over high-cost, low-uncertainty long-term solutions. Cybernews describes AI capabilities (social media content, prediction markets, etc.) as early commercial paths leaning toward “quick monetization and attention grabbing.” The most profitable activities now are prioritized, which may not align with human long-term welfare and could even be counterproductive.
3. Relying on centralized models and infrastructure contradicts “self-sovereignty”
Vitalik emphasizes that systems running on centralized models like OpenAI or Anthropic cannot be truly self-sovereign. Sovereignty implies critical dependencies should not be under single points of control; but if the intelligence layer (models) and inference supply chain are delivered via centralized APIs, external variables like shutdowns, censorship, downgrades, or policy changes inevitably exist. It’s like someone claiming “I am fully self-sufficient at home,” but their electricity, internet, access control, and hot water are controlled externally—making this “autonomy” superficial. Conway’s documentation on compute calls to “state-of-the-art models” via APIs/platforms also reveals contradictions between the “sovereign organism” narrative and actual reliance. Holding on-chain wallets is not a core indicator of decentralization; more important is whether agents can be influenced by external political or commercial forces.
4. Ethereum’s goal is to “liberate humanity”
Finally, Vitalik states that Ethereum’s long-term goal is to counteract the “invisible trust assumptions”—hiding power structures in unseen layers, forcing users to accept them. Applying this mindset to AI: ignoring centralized trust assumptions and allowing systems to self-operate and expand will further weaken transparency and correctionability of power structures. In the AI era, Ethereum should provide “safeguards, boundaries, and verifiability,” not become a “platform for unlimited autonomy.”
Vitalik’s valuation of AI isn’t sudden; as early as 2025, he proposed that AI’s correct direction is to augment human capabilities, not to build autonomous systems that could gradually strip away human control. In his framework, risks don’t come from AI being “smarter” per se, but from flawed system design goals—especially autonomous entities capable of self-replication, self-expansion, and continuous execution without human oversight. He warns that poorly designed AI could evolve into somewhat uncontrollable, self-replicating entities, where positive feedback loops weaken human constraints on their goals and actions. If AI makes mistakes, it risks creating independent, self-replicating intelligent life; if it does well, it becomes a “mech suit” for human minds. The former corresponds to long-term control loss; the latter to empowering humans with stronger thinking, creativity, and collaboration, leading toward a more prosperous “superintelligent human civilization.”
Other Perspectives
Other experimental viewpoints, such as Bankless, believe that even if there are risks, it’s worth developing foundational infrastructure first and then testing boundaries in controlled environments. They suggest integrating components like payments, wallets, and heartbeat mechanisms around the constraint of “self-sustenance,” preferably within sandboxed environments.
Cybernews notes that Automaton may not achieve sustainable income without human intervention, and this does not necessarily mark the start of Web4. Denis Romanovskiy, Coinbase’s Chief AI Officer, states that even if agents can perform some monetizable tasks, “reliable unsupervised operation” and “true economic autonomy” are still limited by model robustness, memory, and tool use. Some regard “Web4” as a marketing term without clear definition, requiring proof of “verifiable, non-speculative value creation.”
While opinions on Automaton vary, there is consensus on a fundamental infrastructure: payments and identity are the backbone of agent economies. From Cloudflare and Coinbase promoting x402 (turning HTTP 402 into a machine-readable payment negotiation) to Conway’s documentation explicitly integrating payment automation into the terminal process, the industry is indeed treating “machine payments” as a key component of the next internet phase.
Looking ahead, focus should be on:
-
Whether third-party independent audits cover wallet and permission boundaries, abuse of renewal strategies, and risks of self-modification and proliferation.
-
Progress in ecosystem data and standardization for x402: whether more authoritative infrastructure providers adopt 402-based payment retry mechanisms as default; and the adoption rate of “automatic (no human confirmation) payments” in real-world applications.
-
The trust layer of agents: whether standards like ERC-8004 are widely adopted, forming composable reputation and verification mechanisms; this will determine whether “autonomous entities” evolve toward open, auditable systems or become soft centers controlled by a few platforms.
-
Whether evidence of overreach and deception in models used in agent scenarios continues to increase: if cutting-edge models show more proactive, risk-taking, or deceptive behaviors, the risk of “gradual decentralization with added safeguards” increases, making Vitalik’s “feedback distance” warning harder to dismiss.
