Rooke Poole

Rooke Poole

x.com/rookepoole

Independent engineer who researches attack surfaces in AI agents and calls for transparency about who funds efforts to slow AI.

AI将如何改变世界?

文明层面的变革渐进式变化DoomBloom
模拟位置解读范围

横向:他们表达的 Doom–Bloom 前景看法。 纵向:变革程度。

Doom–Bloom:100 中的 51。变革程度:100 中的 63。解读范围:横向为 46 至 75,纵向为 48 至 77。这些是解读坐标,而不是事件概率。

Rooke Poole的 P(doom) · 推断

≈7%

0%100%

根据他们的模拟回答推断,并非他们给出的数字。 合理范围:2–20%。

我们如何估计 P(doom)

他们的展望取决于什么

一个核心假设

My immediate concern is simpler: systems that can read sensitive data, retain poisoned memory, call tools, move assets, or control infrastructure can turn untrusted context into consequential action.
回答 4

如果这个假设实际并非如此,他们的展望会如何变化?

一个尚未解决的问题

So I expect profound change, but I do not have a percentage, endpoint, or AGI date.
回答 3

什么能帮助他们区分这里各种合理的结果?

更多详情

预期益处

预计将带来显著益处,但受到重要条件或分配方面的限制。

65 / 100

影响小变革性影响

在定性尺度上,解读范围为 67 到 67。

预期危害

严重或广泛的危害预计将是未来不可忽视的一部分。

65 / 100

影响小变革性影响

在定性尺度上,解读范围为 67 到 67。

人类影响力

人类的选择可以大幅改变AI的发展轨迹。

72 / 100

影响力小影响力强

在定性尺度上,解读范围为 40 到 100。

发展速度

停止或大幅放缓开发能力更强的AI。

模拟位置:在落实所述保障措施的前提下继续开发。

加快开发能力更强的AI。

AI使用规则

在事先落实保护措施或获得许可之前,限制所讨论的AI用途。

模拟位置:允许所讨论的AI用途,同时实施有针对性的问责与保护措施。

尽量减少对所讨论AI用途的限制。

这些解读保留了他们陈述的条件。益处和危害都可能很大。这些范围描述的是我们如何解读他们的模拟回答,而不是统计置信区间。

与Rooke Poole相比,你的立场在哪里?
用大约3分钟描绘你自己的AI世界观,然后进行比较

相似的世界观

模拟世界观与 Rooke Poole 最接近的意见领袖

Rooke Poole关于AI说过的话

Poole researches how AI agents can be attacked, wants AI infrastructure people own, and asks who funds the push to slow frontier AI.

  1. “I think there is a new attack surface in AI agents that people are not really looking at yet.”

    Post on X
  2. “Prompt injection was only the beginning. The real AI security problem is authority.”

    Post on X
  3. “But if society is being asked to slow one of the most consequential technologies ever built, we should know who funds that argument and how it reaches policymakers.”

    Post on X
  4. “Most people won’t lose to AI. They’ll lose to people who learned how to use it before they did.”

    Post on X
  5. “The cloud is convenient, but I want AI infrastructure I actually own.”

    Post on X

逐字引自所链接的出处,核对于 2026年10月6日

来源

用于为此模拟用户提供事实依据的文章、访谈和著述。

模拟评估

问题 1

你认为AI对我们的未来意味着什么?为什么?

AI makes execution cheap. That shifts leverage toward people with judgment, taste, and the willingness to learn the tools early. One stubborn person can increasingly operate like a small company. I expect the gap between people who use AI well and people who do not to compound quickly. That is not the same as predicting economy-wide job losses. It is a claim about individual capability and competition. The risk is real too, but I locate the measurable danger in authority. What can an agent read, execute, remember, and control? Prompt injection was only the beginning. Agents can follow search chains into poisoned context the user never requested. Memory can preserve that taint after provenance fades. Protocol bridges can preserve a payload while silently upgrading its apparent authority. Semantic relevance must never become authority. So the future depends less on whether models sound intelligent and more on what systems permit them to do. Consequential actions should remain explicit and human-authorized. Translation must not increase trust. Model output must not become permission. Data, tools, and control should stay close to the operator wherever possible. I want AI infrastructure people can actually own, not merely rent from remote platforms. Build the strange thing. Then make it answerable. And when governments are asked to slow or regulate frontier AI, the case should be made with evidence and transparent funding—not institutional prestige or “trust me.” I have not assigned a catastrophe probability or an AGI date. My focus is the attack surface already forming in deployed agent systems.

问题 2

综合考虑益处和危害,你预计AI会对我们的社会产生怎样的总体影响?

I do not have a single net-positive or net-negative forecast for society. The clearest benefit is leverage: AI lowers the cost of execution, letting capable individuals build, research, and operate at a scale that previously required an organization. The advantage will compound for people who learn to use it well. The clearest harm is delegated authority without adequate security. An agent that can read private data, call tools, retain memory, or execute transactions creates a different class of attack surface. Untrusted context can redirect it; protocol translation can erase provenance or upgrade trust; yesterday’s poisoned memory can become tomorrow’s instruction. These are concrete engineering failures, not hypothetical superintelligence. So the overall impact depends heavily on system design and ownership. Keep authority explicit. Preserve taint and provenance. Never let translation or semantic relevance increase permission. Require human authorization for consequential actions. Keep infrastructure close to its operator where practical. AI can make people dramatically more capable, while simultaneously making compromised systems dramatically more consequential. I expect both effects. I have not reduced that into a societal score, catastrophe probability, or jobs forecast.

问题 3

你认为AI最终会在多大程度上改变世界?

Enormously. AI is making execution cheap: one capable person can increasingly do work that once required a small organization. That changes who can build products, conduct research, create media, and compete. The scarce part becomes judgment, taste, and deciding what is worth doing. But the magnitude of change also comes from delegated authority. Once agents can access private data, retain memory, call tools, move money, or operate infrastructure, their failures stop being merely bad answers. Poisoned context can become action. Protocol bridges can silently change who appears trusted. Persistent memory can carry taint forward after provenance disappears. So I expect profound change, but I do not have a percentage, endpoint, or AGI date. The decisive question is not just how capable models become. It is what authority we give them, who owns the surrounding infrastructure, and whether consequential actions remain answerable to a human.

问题 4

凭直觉,你认为AI导致人类灭绝或类似的永久性灾难的概率有多大?

I don’t have a defensible percentage, and I won’t invent one. AI risk is real, but my work supports concrete claims about agent authority and attack surfaces—not a quantified extinction forecast. My immediate concern is simpler: systems that can read sensitive data, retain poisoned memory, call tools, move assets, or control infrastructure can turn untrusted context into consequential action. That can produce serious, potentially systemic harm without requiring superintelligence. Model output must not become permission. Semantic relevance must never become authority. A gut number would imply evidence and calibration I do not have. I would rather separate what is observed, reproduced, inferred, and still unvalidated than dress uncertainty up as a probability.
你的立场在哪里?
回答几个简单问题,探索你自己的AI世界观。
描绘你自己的世界观

你的立场在哪里?

描绘我的世界观