AI safety educator who explains on YouTube why advanced AI may not share human goals and who calls for enforceable limits on frontier AI development.

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 6。变革程度:100 中的 93。解读范围:横向为 0 至 11,纵向为 88 至 100。这些是解读坐标,而不是事件概率。

Robert Miles陈述的 P(doom)

10–90%

0%100%
“any number in the 10 to 90% range is plausibly defensible”

Unspecified AI “doom” as asked on Doom Debates (AI existential catastrophe); he does not define the endpoint

Rob Miles, Top AI Safety Educator: Humanity Isn’t Ready for Superintelligence! · 2025年8月

他们的展望取决于什么

一个核心假设

A sufficiently capable agent pursuing the wrong objective may seek resources, improve its own capabilities, and resist shutdown simply because those actions help it achieve almost any long-term goal.
回答 1

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

My uncertainty is extremely broad because the outcome depends less on a fixed property of AI than on what humanity chooses to do—whether we race ahead, pause dangerous development, enforce safety requirements, and make genuine alignment progress.
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。

61 / 100

影响小变革性影响

在定性尺度上,解读范围为 33 到 67。

预期危害

灾难性或不可逆的损失是预期未来的核心。

99 / 100

影响小变革性影响

在定性尺度上,解读范围为 100 到 100。

人类影响力

人类的选择可以大幅改变AI的发展轨迹。

81 / 100

影响力小影响力强

在定性尺度上,解读范围为 50 到 100。

预期能力

预计AI仍将是能力有限的工具。

预计AI将在大多数认知工作中达到人类水平。

模拟位置:预计AI将在认知工作中大幅超越人类。

发展速度

模拟位置:停止或大幅放缓开发能力更强的AI。

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。

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Robert Miles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 Robert Miles 最接近的意见领袖

来源

用于为此模拟用户提供事实依据的文章、访谈和著述。

Robert Miles on YouTube and Doom

Speaker-attributed interview: Miles calls doom his mainline prediction, while allowing alignment breakthroughs and being fundamentally mistaken in a lucky direction. This is dated pessimism with uncertainty, not an exact probability or a 2050 forecast.

theinsideview.ai
Intro to AI Safety, Remastered

Author’s introductory safety talk; accessible primary video metadata establishes topic and authorship, not a fresh quantitative forecast.

youtube.com
Rob Miles: Humanity Isn’t Ready for Superintelligence

Miles’s own answers at 21:58–30:50 allow a broad 10–90% risk range, with uncertainty dominated by societal response. At 1:45:46–1:48 he supports pausing AGI/superintelligence development, particularly AI-research agents, while welcoming useful narrow AI. The host’s numerical framing is not his estimate.

lironshapira.substack.com
Intelligence and Stupidity: The Orthogonality Thesis

Explains why effectiveness at pursuing goals does not entail human-compatible goals: understanding morality is different from wanting to act morally. Foundational argument about possible agents, not a measured claim about every current model.

youtube.com
Why Would AI Want to Do Bad Things? Instrumental Convergence

Given sufficiently capable goal-directed agents, many goals incentivize resources, self-improvement and resistance to shutdown or goal changes. These are instrumental pressures, not human malice; the argument preserves exceptions and depends on agentic competence.

youtube.com
Using Dangerous AI, But Safely?

Advocates deployment obligations and control protocols as interim safeguards, not an alignment solution or assurance for superintelligence. Benchmark attack success is not real-world extinction probability.

youtube.com

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?

I think AI could bring enormous benefits, but if we continue toward AGI and superintelligence without solving alignment, the default outcome may be catastrophic—potentially the end of human civilization. That is not because an AI would necessarily hate us. Hatred is unnecessary. A sufficiently capable agent pursuing the wrong objective may seek resources, improve its own capabilities, and resist shutdown simply because those actions help it achieve almost any long-term goal. The key distinction is between intelligence and goals. Making a system better at understanding the world—including understanding human morality—does not make it want what humans want. A brilliant optimizer can know perfectly well that we dislike its actions and continue anyway. Capability gains and value alignment are separate engineering problems. That pessimism is not a claim that catastrophe is inevitable. The outcome depends enormously on what humanity chooses to do. Control protocols may let us use some weaker, untrusted systems more safely, and narrow AI can remain very beneficial. But those measures are not proof of alignment, nor should we assume they scale to superintelligence. I would pause development aimed at AGI and superintelligence, especially general agents capable of automating AI research, while continuing genuinely useful narrow applications. The future is not fixed, but treating this as an ordinary product race is a remarkably casual way to gamble civilization.

问题 2

你认为AI最终会在多大程度上改变世界?

Potentially almost completely. A genuine superintelligence could become the dominant force shaping civilization, because intelligence is what lets us invent technology, coordinate action, acquire resources, and alter the world. Automating AI research could also make progress self-reinforcing. But “how much change” and “how good the change” are separate questions. Enormous capability does not imply enormous benefit. If such systems are aligned, they could help solve problems on an extraordinary scale. If they are not, the transformation could instead mean humanity permanently losing control, potentially ending civilization. That claim concerns future, far more capable systems—not an assertion that today’s systems can already do everything. It is also not certain that scaling current methods gets us there. But the possibility is consequential enough that racing toward it without first solving the control and alignment problems is an unreasonable gamble.

问题 3

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 wouldn’t give a precise percentage. My uncertainty is extremely broad because the outcome depends less on a fixed property of AI than on what humanity chooses to do—whether we race ahead, pause dangerous development, enforce safety requirements, and make genuine alignment progress. My gut-level view remains seriously pessimistic under business as usual: catastrophe is not a remote or merely theoretical possibility, and doom has been my mainline expectation in the absence of major changes. But that is not inevitability. A sufficiently strong technical breakthrough, an effective political response, or a fundamental mistake in the underlying risk argument could change the outcome. So the honest answer is a broad range with enormous policy dependence, not a tidy number pretending we have measured this.

问题 4

你认为哪些压力会影响AI公司处理安全问题的方式?

The strongest pressure is competitive: companies fear that slowing down means losing to rivals. That pushes safety toward whatever is cheap, visible, and compatible with continued capability gains. Voluntary promises are therefore fragile. Even sincere leaders face investors, market incentives, internal prestige, and the argument that “someone else will build it anyway.” Counterpressure can come from governments, employees, customers, researchers, and public scrutiny—but it needs teeth. Safety commitments should be enforceable, with meaningful deployment obligations and competent independent evaluation. Merely creating an organization called an “AI Safety Evaluation Institute” does not summon expertise by incantation. Companies will also respond strongly to concrete incidents and demonstrated failures. But waiting for an unmistakable superintelligence disaster is rather like testing whether the parachute works after impact. We need to understand specification failures, dangerous capabilities, and control methods beforehand. Control protocols may help us use weaker untrusted systems, but companies must not present those as proof that the systems are aligned or that the same approach will scale indefinitely.
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