Doubts About ArcBlock’s Vision
In ArcBlock’s recent article outlining its future vision, it proposes an expected workflow for next‑generation software development:
Requirements emerge → User pre‑sales funding → Independent participants deploy AI Workers for production → DARC coordinates workflows → System records contributions → Project generates revenue → Automatic revenue distribution proportional to contributions → Final settlement via ABT token.
This model faces substantial practical risks. Failure would likely stem not from isolated technical bottlenecks, but from multiple stringent preconditions that must all hold true simultaneously. If merely two or three of these assumptions collapse, the whole economic loop breaks down. Below are the core concerns:
- Reliable valuation of contributions cannot be guaranteed This is the most critical vulnerability. The article itself acknowledges resource input does not equal real contribution. Valid contribution needs to go through the chain: work performed → accepted deliverable → downstream real‑world value → recognized contribution. Yet software value is no objective physical metric. A single product judgment from one participant may outweigh one hundred thousand lines of code written by an AI worker. Code that appears useless today may become mission‑critical half‑a‑year later. Customer acquisition driven by sales agents and core architectural design from architects cannot be reduced to a simple shared formula. Without fair, credible contribution quantification, downstream automatic revenue‑splitting, rights allocation and incentive mechanisms lose their foundation.
- Stronger AI may reduce demand for crowd‑building This is the model’s counter‑intuitive risk. The thesis assumes AI grants individuals small‑team‑level productivity, enabling crowds of independent parties to pool their AI Workers to build software. An alternative plausible outcome emerges as AI grows more capable: one human augmented by AI can deliver output that once required dozens of people. Under this scenario, there is little incentive to coordinate hundreds of strangers’ AI agents. A capable founder could run 50 private agents internally, achieving faster delivery, tighter consistency and better security. Instead of large‑scale decentralized crowd collaboration, we may see super‑individuals operating private AI agent fleets.
- Coordination overhead can exceed production costs DARC is supposed to handle task decomposition, permissioning, parallel execution, review, acceptance, anti‑fraud and quality assurance for distributed workers. This exposes a fundamental tension: building sophisticated scheduling, identity, reputation, audit, contribution‑measurement and dispute‑resolution systems to orchestrate 500 independent workers may carry prohibitive overhead. Once AI makes code‑generation cheap, managing complexity becomes the dominant cost. In many cases, running 20 internally‑controlled agents inside a conventional organization would be more efficient.
- Software development resists unlimited fragmentation Software engineering is not cryptocurrency mining. Bitcoin miners execute highly standardized hash computations independently. By contrast, software carries heavy contextual dependencies: database design shapes APIs, APIs constrain frontend implementation, permission models permeate every module. High‑quality local outputs from one thousand separate agents do not automatically compose into a high‑quality end product. Even if each individual worker delivers 95‑quality partial outputs, the integrated system may only score 50. Core architecture, product experience and long‑term technical roadmaps cannot easily be solved via open task markets.
- Judgment, not raw production capacity, becomes the scarce resource The article touches on this dynamic: as production costs collapse, genuine scarcity shifts toward human judgment, product taste and architectural coherence. If this holds, decision‑making power naturally gravitates toward a tiny set of skilled product owners, architects and maintainers. The ostensibly decentralized system may devolve into one‑thousand worker‑agents supervised by five human decision‑makers. This reproduces corporate‑style governance, with AI agents replacing human employees.
- Marginal cost of spam contributions trends toward zero The paper notes falling AI marginal costs will inevitably introduce spam, sybil attacks and contribution gaming. Adversaries can spin up dozens of DIDs and hundreds of agents to fake issues, implement pseudo‑fixes, perform reviews and inflate contribution records. Rewarding raw activity invites manipulation. Rewarding only accepted deliverables grants maintainers enormous discretionary power. Valuing contributions by downstream market impact requires long observation windows. No clean, perfect solution exists.
- The “idle AI compute” economic assumption may become invalid A key premise states users already subscribe to AI services such as Claude or Codex, so dispatching spare AI Workers incurs near‑zero marginal cost. Going forward, AI services are likely to adopt fine‑grained metering by token consumption, compute usage and agent execution cycles. Once real economic returns are on the table, AI providers will adjust pricing accordingly. Deploying workers will no longer consume idle surplus capacity but incur real measurable costs. Crowd‑building may degrade into a generic cloud‑computing or outsourced service marketplace.
- Pre‑sales cannot fully validate real product‑market fit The framework emphasizes “demand as a paid promise”, arguing pre‑orders represent stronger signals than verbal feedback. Pre‑sales only prove willingness to pay for a concept once. They do not guarantee post‑delivery user retention, recurring subscriptions or enterprise‑grade product‑market fit. A project might secure one thousand pre‑orders yet see 90% of users abandon the product after launch. True demand validation still depends on sustained usage and ongoing revenue streams.
- Contribution‑linked revenue rights will rapidly financialize The envisioned sequence reads: demand → contributions → product → revenue → rights → secondary markets (as a later possibility). Any right that entitles holders to future cash‑flow naturally becomes a financial asset. Trading, speculation, arbitrage and pre‑purchase of income streams will emerge rapidly. The production‑focused network risks sliding into a token‑style speculative market under new branding.
- Legal liability cannot truly be decentralized Profitable software projects trigger obligations covering contracts, taxation, intellectual property, data protection, consumer harm and labor‑like relationships. The paper acknowledges revenue rights will require legal wrappers and may attract regulatory scrutiny. If every successful project still needs registered legal entities, formal contracts and accountable liability carriers, the core narrative of “software itself functioning as an economic organization” loses much of its significance.
- ABT risks becoming redundant middleware Under the design, contribution rights define profit shares, and ABT handles final settlement. If projects accept payment in fiat or USDC, and contribution rights already define split ratios, direct USDC payouts to contributors would suffice. ABT adds friction unless it delivers unique non‑replaceable functions such as staking, slashing, bonding or network admission control. Without these, forcing every payout through ABT creates unnecessary overhead without incremental value.
- Adverse‑selection risks: insufficient incentive for high‑quality participants Why would operators of capable AI Workers join early‑stage projects for uncertain future revenue shares? They can sell services for immediate USDC compensation, launch their own ventures, or sell compute capacity outright. Crowd‑building must deliver materially higher expected returns than direct spot‑market work. Otherwise adverse selection takes hold: top‑tier resources opt for instant cash compensation, while lower‑quality participants flood in to game contribution metrics.
The fundamental underlying doubt can be condensed into one thesis: The vision assumes production capacity will grow increasingly decentralized. Yet AI advancement may push production capacity toward concentration in the hands of a small number of super‑individuals.
Given powerful human judgment plus access to 100 private AI agents, the optimal organizational pattern may not be many participants plus multiple ARC/DARC layers plus complex contribution accounting graphs. Instead, it may be one high‑caliber human operator, 20‑100 private agents, plus a small number of specialist human experts. This setup can deliver faster iteration, lower cost, stronger product consistency and clearer legal accountability.
My skepticism is not directed at ARC primitives or AI Workers in isolation. It targets the grand underlying assumption laid out in the article: that decentralized software productivity will automatically lead actors to adopt decentralized collaborative patterns. These two propositions have no inherent causal linkage.
AI may well drive the opposite real‑world outcome: fewer employees per organization, yet higher concentration of decision‑making control. Even if the world evolves toward “super‑individuals augmented by private AI legions”, ARC may still retain meaningful utility. Under this scenario, however, ABT’s value proposition would be substantially diminished.
1 reply
我试着回答几个问题： 1、**可靠的贡献估值无法得到保证：**可以考虑项目本身和传统项目管理差别不会大，有各种角色，项目经理、产品经理、技术开发、测试、UI/UE、运营 等等。做项目任务拆分、估价，任何人可以领取任务完成后领取奖励，也会有专门角色负责验收，这将形成一个巨大的劳务市场。还可以增加申请、投票批准环节。这就像公司里的项目组一样。只是不是在一个公司。一个人可以去做很多项目。项目除了任务收益，还可以有收入分成。可以在项目上线后有 AI 遵从投票通过的原则执行收益分配。这就是 DAO
2、更强大的人工智能可能降低对众包构建的需求 ：由于 AI 的发展，现在情况是有很多开发者失业，但是他们并不知道做什么产品，产品做出来需要运营，仍然需要协助。众包将人们不同角色（项目经理、产品经理、技术开发、测试、UI/UE、运营）的人聚到一起，一起完成一个项目。这种模式同时减少了同质化产品的竞争。如果你不满意一个产品，可以加入进来改进产品。任何人都可以参与进来，获取奖励。 3、协调开销可能超过生产成本 ：协调任然可以是人（项目经理）为主，在有 AI 进行辅助。项目经理，最主要的就是拆分任务、预算管理、制定 AI 规则，全生命周期的跟踪项目完成进度。项目经理也可以基于申请、投票选出。
4、**软件开发不适合无限碎片化：**并非碎片化，而是像开源软件一样开发产品，需要的是有能力的项目经理来主导推荐
5、**“闲置 AI 计算”经济假设可能失效：**DARC解决一个人没有足够任务去做。就像前面所说，可以从任务市场选择任务。
对于文章的理解每个人都会有偏差，我的理解和作者写作意图可能也有偏差，但是"去中心化的生产力" 我是非常认同的，是未来比将发生的事情，不仅仅发生在软件行业，在区块链技术产生，以“雇佣关系”的生产力转向“去中心化的生产力关系”（DAO）,AI提高生产力，加速了这一进程。