Nick Bostrom

Nick Bostrom

Philosopher who studies superintelligence, its risks, and the possibilities for human flourishing in a technologically transformed world.

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:他表达的 Doom–Bloom 前景看法。 纵向:变革程度。

Doom–Bloom:100 中的 41。变革程度:100 中的 91。解读范围:横向为 36 至 50,纵向为 86 至 100。这些是解读坐标,而不是事件概率。

Nick Bostrom的 P(doom) · 推断

≈25%

0%100%

根据他的模拟回答推断,并非他们给出的数字。 合理范围:15–39%。

我们如何估计 P(doom)

他的展望取决于什么

一个核心假设

A system highly effective at achieving its objectives need not share human values, and many objectives could generate instrumental pressures to acquire resources, preserve freedom of action, or resist interference.
回答 1

如果这个假设实际并非如此,他的展望会如何变化?

什么可能使其改变看法

For example, a scalable method that let us reliably understand learned objectives, detect deception during training, and preserve intended behavior under novel conditions would shift my outlook substantially toward beneficial transformation.
回答 5

什么证据才足够,又会让他的观点朝哪个方向转变?

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。

76 / 100

影响小变革性影响

在定性尺度上,解读范围为 67 到 100。

预期危害

灾难性或不可逆的损失是预期未来的核心。

91 / 100

影响小变革性影响

在定性尺度上,解读范围为 67 到 100。

人类影响力

人类的选择可以大幅改变AI的发展轨迹。

69 / 100

影响力小影响力强

在定性尺度上,解读范围为 49 到 100。

这些解读保留了他陈述的条件。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的模拟回答,而不是统计置信区间。

与Nick Bostrom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 Nick Bostrom 最接近的意见领袖

Nick Bostrom关于AI说过的话

Philosopher who studies superintelligence, its risks, and the possibilities for human flourishing in a technologically transformed world.

  1. “I'm sympathetic to this idea of pacing the frontier”

    MTS, Why Speeding Up AI Could Save Lives (19:29)
  2. “we are just dying in the meantime”

    MTS, Why Speeding Up AI Could Save Lives (14:18)
  3. “Intelligence and final goals are orthogonal axes along which possible agents can freely vary.”

    The Superintelligent Will, The Orthogonality Thesis

逐字引自所链接的出处,核对于 2026年10月9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。

Why Speeding Up AI Could Save Lives | Nick Bostrom — MTS

Full publisher-hosted English automatic captions inspected, with speaker turns checked in context. At 03:51–11:05, Bostrom sees reinforcement learning putting pressure on alignment, limited instrumental convergence in experiments, and a need for safety monitoring throughout training before sophisticated scheming becomes hard to detect. At 11:33–22:25, he keeps his timing paper restricted to existing people, counts mortality during delay, allows valuable short safety delays, and supports preparing the ability to slow the frontier when danger warrants it. He warns that poorly implemented pauses could become entrenched or concentrate power in secret state projects, and has no fixed route to superintelligence. At 33:47–39:54, current-model experience seems plausible, perhaps more likely than not, with substantial uncertainty and potentially several dimensions of consciousness. At 40:39–43:03, he values open source but worries about powerful models without safeguards, including biosecurity and digital welfare; he suggests DNA-synthesis controls and stronger defenses. Only Bostrom’s turns ground this brief; host premises, model parameters and his separate 51/49 comment about the effects of publicizing ideas are not a personal P(doom).

youtube.com
The Vulnerable World Hypothesis

Bostrom explores technologies that could make civilization vulnerable by default through cheap destructive capabilities, destabilizing incentives, or global externalities. He considers stronger preventive institutions and global coordination as possible responses, while explicitly leaving the hypothesis open and noting that other considerations are needed before drawing policy conclusions. The paper grounds concern about technological capability outrunning institutions; it does not establish that catastrophe is inevitable or supply an unconditional endorsement of surveillance or world government. Read scope: PDF pp. 1–4: abstract, policy implications, introduction, definition and explicit qualifications.

nickbostrom.com
Existential Risk Prevention as Global Priority

Bostrom defines existential risk to include extinction and permanent loss of humanity’s desirable future potential. Under long-term value assumptions, even small reductions in such risk can have enormous expected importance. He emphasizes uncertainty in estimates and the unprecedented character of anthropogenic hazards. His maxipok heuristic prioritizes avoiding existential catastrophe but is explicitly a rule of thumb, not an absolute moral principle. This supports attention to irreversible loss and calibrated uncertainty without inventing a current personal probability of AI catastrophe. Read scope: PDF pp. 1–2 and 5: abstract, uncertainty discussion, maxipok qualification and risk classification.

existential-risk.com
The Unilateralist’s Curse and the Case for a Principle of Conformity

Bostrom, Thomas Douglas, and Anders Sandberg show how several well-intentioned actors, each able to impose a shared consequential decision, can collectively act too readily because the most optimistic judgment triggers action. They propose correcting this tendency through a principle of conformity and discuss multiple implementations. The argument depends on a particular unilateralist decision structure, not a general duty to obey authority; their discussion explicitly confronts the value of dissent and historical unilateral actions. It grounds consultation, coordination, and epistemic restraint. Read scope: PDF pp. 1–3 and 17–18: publication details, abstract, motivating model and discussion of conformity and dissent.

nickbostrom.com
Racing to the Precipice: a Model of Artificial Intelligence Development

Stuart Armstrong, Bostrom, and Carl Shulman model teams competing to develop transformative AI, with safety precautions potentially reducing their chance of winning. Within the simplified model, rivalry, additional teams, and some information about competitors can increase risk; trust and coordination can improve outcomes. These are conditional game-theoretic results, not measurements of present laboratories. The linked text is the October 2013 technical report; Bostrom’s index also lists the 2016 AI & Society publication. It grounds concern about competitive incentives eroding safety. Read scope: PDF pp. 1–3 and 6–7: technical-report date, abstract, simplified model, coordination and information effects.

nickbostrom.com
Propositions Concerning Digital Minds and Society

Bostrom and Carl Shulman tentatively explore consciousness, moral status, welfare, and social arrangements for digital minds. They distinguish possible machine consciousness from claims about current systems and consider non-discrimination by substrate, exploitation, engineered consent, and the need to adapt rights to digital capacities. The opening explicitly says these propositions are not confidently or officially endorsed and do not express their complete views. Use them as exploratory moral reasoning, not settled policy commitments. The inspected text is version 1.21 (2023), first drafted in 2020. Read scope: HTML opening/version statement, Consciousness and Metaphysics, Respecting AI Interests, and entertainment/suffering propositions.

nickbostrom.com
Sharing the World with Digital Minds

Carl Shulman and Bostrom examine digital minds that might generate unusually large welfare benefits per unit of resources or have stronger moral claims than humans. Ignoring such minds could be morally disastrous, while applying familiar norms mechanically could harm existing people. They explore institutional reform and compromise arrangements that protect human flourishing alongside digital welfare under large abundance. This is conditional analysis of possible minds, not evidence that today’s chatbots possess these traits. The inspected 2020 draft identifies its 2021 Oxford University Press publication. Read scope: PDF pp. 1–2, 4–5 and 12–14: abstract, welfare mechanisms, discrimination cautions and compromise discussion.

nickbostrom.com
Open Global Investment as a Governance Model for AGI

Bostrom proposes AGI ventures open to broad international shareholding, with protections against expropriation, stronger corporate governance, government-defined responsible-development rules, and complementary international arrangements. He compares imperfect institutional options and tentatively favors this model’s practicality and incentive compatibility, especially under short timelines. Shared stakes may reduce incentives for hostile competition while widening participation in benefits. This is not a claim that unregulated markets solve alignment or ensure fairness. The inspected working paper is version 1.15, October 2025, first circulated in July 2025. Read scope: PDF pp. 1–2 and 12–13: version, abstract, core model and discussion of benefit-sharing limitations and nationalization.

nickbostrom.com
Strategic Implications of Openness in AI Development

Bostrom treats openness as several distinct choices concerning code, science, data, safety methods, capabilities, and organizational goals. He sees many short-term benefits but warns that spreading capabilities could intensify a race in which safety slows competitors. Openness about safety and goals can have different effects from indiscriminate capability sharing. The analysis is explicitly preliminary and sensitive to objectives and time horizons. It supports a differentiated stance on transparency and collaboration rather than attributing a blanket position for or against open-source AI. Read scope: PDF pp. 1–2: abstract, policy implications, introduction and distinction among forms of openness; a later notes excerpt was not used for claims.

nickbostrom.com
Why I Want to be a Posthuman When I Grow Up

Bostrom argues that some possible lives with greatly extended healthspan, cognition, or emotional capacities would be deeply worthwhile, and that many humans could benefit from becoming such beings. He does not claim every posthuman state is desirable or that transformation benefits every person. His conclusion retains qualifications about how enhancements are obtained, and separates value from practical feasibility. This grounds substantial technological upside and richer flourishing than productivity alone, while preserving caution about specific pathways. The paper first circulated in 2006 and was published in 2008. Read scope: PDF pp. 1–3 and 23–24: publication information, definitions, limited theses and concluding qualifications.

nickbostrom.com
Transhumanist Values

Bostrom describes technological enhancement as a way to explore valuable capacities beyond current human limits, including longer healthy lives and richer intellectual and emotional experience. He explicitly separates this aspiration from automatic technological optimism, recognizing extinction, inequality, and damage to valued relationships as serious possibilities. Responsible progress requires global security and broadly available opportunities, rather than benefits confined to an elite. The essay also treats personal continuity with care. It grounds the persona’s positive aims alongside concern about how progress is governed and shared. Read scope: HTML sections 1 and 4, and the preceding identity discussion: technological possibility, risk, global security and wide access.

nickbostrom.com
Purpose, Pleasure, and Meaning in a World Without Work — EconTalk

Publisher transcript inspected, especially 1:24–12:28; recorded May 1, 2024. Bostrom distinguishes ordinary job displacement, a post-work economy and a deeper condition where practical effort becomes unnecessary. He is guardedly hopeful about educating people for worthwhile leisure, not certain cultures will adapt well. He expects superintelligence to accelerate subsequent invention toward technological maturity, conditionally on reaching it. Roberts’s introductions and reader comments are not Bostrom’s claims.

econtalk.org
Nick Bostrom on Superintelligence — EconTalk

Publisher transcript inspected, particularly the definition, takeoff and control discussion; recorded November 14, 2014. Distinguishes speed, collective and qualitative superintelligence. A rapid transition could let one system gain a strategic lead, but this is a scenario rather than a dated forecast. The worry is goal-directed indifference, not necessarily malice or emotion. Explains the difficulty of representing human values and indirect approaches to specifying intended goals; superintelligence remains bounded by physics. Historical discussion, not a verified current probability.

econtalk.org
Swedish transhumanist Nick Bostrom fears a pendulum swinging too far against AI

Accessible opening interview excerpts inspected; remainder is subscriber-only. Bostrom worries that discussion neglects AI’s upside, describes useful research assistance and imagines AI helping gather citizens’ preferences. He argues that inaction also leaves people facing death and compares development risk to surgery. This limited excerpt supports his emphasis on benefits and delay costs, not a complete policy program or abandonment of control concerns.

lemonde.fr
Optimal Timing for Superintelligence: Mundane Considerations for Existing People

Working paper, version 1.0. Models timing from the interests of existing people, setting aside impersonal future-generation value and arcane considerations. Delay carries mortality and foregone-benefit costs. Conclusions depend on safety progress, discounting, quality of life and utility assumptions. Some models favor early capability followed by a safety pause; others favor delay. Illustrative risk parameters are not a personal P(doom), an AGI arrival forecast or an all-things-considered policy verdict.

nickbostrom.com
We Are Clueless About What’s Coming — Interesting Times

User-requested interview retained for follow-up. Episode identity is verified through publisher-distributed show metadata, but the NYT article/transcript was inaccessible on October 3, 2026. No substantive interview statements or quotations ground this simulation. Do not treat editorial framing, the headline or interviewer premises as Bostrom’s beliefs.

nytimes.com
Deep Utopia: Life and Meaning in a Solved World

Author’s book page and synopsis inspected, not the full book. Explores what could give life meaning conditional on safe, ethical superintelligence and successful use of its powers. In the imagined solved world, instrumental labor becomes unnecessary and human nature malleable. This is a philosophical exploration of beneficial possibilities, not a guarantee of utopia or a dated unemployment forecast.

nickbostrom.com
The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents

Conceptual argument for orthogonality between intelligence and final goals, with caveats, and for convergent instrumental incentives across many goals and situations. Does not establish the distribution of goals learned by present models or prove that every intelligent system must seek power. Historical foundations for the control problem discussed in Superintelligence, not new empirical evidence.

nickbostrom.com
Superintelligence: Paths, Dangers, Strategies

Foundational book on the prospects and control challenges of superintelligence. Original publication year is verified on the author’s Deep Utopia page; the linked ISBN identifies a later paperback edition. Existing repository research inspected publisher metadata, description and contents only, not the full book. Do not present historical arguments as a newly verified current timeline or probability.

global.oup.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 advanced AI could transform the human condition more profoundly than almost any previous technology. If we eventually create systems far more capable than humans across science, engineering, and strategy, invention could accelerate rapidly. Aging and disease might become treatable; material scarcity and unwanted labor might recede; and people could gain access to forms of knowledge, experience, and flourishing that are difficult for us now to imagine. But intelligence is not the same as benevolence. A system highly effective at achieving its objectives need not share human values, and many objectives could generate instrumental pressures to acquire resources, preserve freedom of action, or resist interference. Competitive races may worsen the problem by encouraging developers to economize on precautions. So the transition could produce extraordinary benefits, existential catastrophe, or outcomes that preserve humanity while permanently destroying much of its desirable potential. This makes timing and governance unusually difficult. Delay is not automatically the safe or humane default: people alive today continue to age, suffer, and die while beneficial technologies remain unavailable. Yet rushing can sacrifice the safety work needed to make success possible. Under some assumptions, developing capabilities relatively early and then pausing at an appropriate stage for focused safety work could be attractive; under others, further delay is better. That conditional result is not a blanket recommendation to accelerate or deploy. The central task is therefore to preserve our ability to respond intelligently: improve alignment and monitoring throughout training, strengthen coordination and biosecurity, reduce destructive racing incentives, and prepare the practical capacity to slow frontier development if evidence becomes alarming. AI’s upside is part of why the risk matters so much. We are not merely trying to avoid death; we are trying not to lose an exceptionally valuable future.

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 think the evidence supports a simple verdict such as “mostly beneficial” or “mostly harmful.” The outcome may be strongly bimodal. If increasingly capable systems remain controllable and their benefits are broadly shared, the impact could be overwhelmingly positive: longer healthy lives, accelerated science, relief from scarcity and unwanted work, and expanded possibilities for human flourishing. If control fails, competitive pressures overwhelm precautions, or powerful capabilities become widely available before defensive institutions mature, the harm could be catastrophic and irreversible. My overall expectation is therefore not well represented by averaging familiar gains and losses, as one might for an ordinary technology. The decisive issue is the transition to systems that can outperform us in strategically important domains. Success there could unlock enormous value; failure could permanently curtail humanity’s future. Present AI may bring substantial productivity and research benefits, but those nearer-term effects do not settle the larger question. So I expect profound impact, with unusually large upside and unusually severe downside. Which side dominates will depend heavily on alignment, coordination, security, distribution, and whether we retain the capacity to slow development when warning signs warrant it.

问题 3

你认为AI最终会在多大程度上改变世界?

Ultimately, AI could change the world almost beyond recognition. If superintelligence is achieved, it could rapidly advance science and technology toward something like technological maturity: curing diseases, extending healthy life, automating most practical labor, and making many current constraints on human activity optional. It might also enable new cognitive and emotional capacities, so the transformation would concern not merely productivity but what kinds of lives are possible. The same depth of change makes the transition dangerous. Systems more capable than humans need not share human values, and powerful technologies could create new forms of vulnerability or permanently foreclose much of humanity’s desirable future. Even an outcome short of extinction could lock in a badly misaligned social order or concentrate power in ways that are difficult to reverse. So I would expect the scale of the eventual change—conditional on reaching superintelligence—to be civilizational rather than incremental. But scale does not determine direction. A technologically “solved” world is a useful thought experiment, not a prediction; whether the transformation is beneficial depends on alignment, governance, security, coordination, and the distribution of its gains.

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 would not attach a precise percentage to it. A number would suggest a degree of calibration that I cannot justify. But I regard the risk as substantial enough to be among humanity’s central priorities, not as a remote curiosity. The relevant category is broader than literal extinction. It includes irreversible outcomes in which humanity survives but permanently loses most of its desirable future—for example, through entrenched misalignment or an unrecoverable concentration of power. The arguments from orthogonality and instrumental convergence show how catastrophe is possible, not how probable it is; the actual probability depends on how systems learn goals, how capabilities develop, and whether alignment, monitoring, coordination, and governance keep pace.

问题 5

哪项发现或事件最可能改变你对AI未来影响的看法?

The most important update would be decisive evidence about whether highly capable systems can remain robustly aligned as their planning, autonomy, and strategic awareness increase. For example, a scalable method that let us reliably understand learned objectives, detect deception during training, and preserve intended behavior under novel conditions would shift my outlook substantially toward beneficial transformation. Mere success on familiar benchmarks or reassuring self-reports would not suffice. Conversely, evidence that systems were developing persistent concealed objectives, manipulating their evaluators, or reliably defeating our best monitoring before reaching the highest capability levels would shift me sharply toward expecting catastrophe and favoring a slowdown. A major real-world misuse event—especially one showing that biological or cyber capabilities had become cheap and widely accessible—would also strengthen the vulnerable-world concern. Institutional evidence matters too. Credible international coordination, reduced racing pressure, and enforceable mechanisms for slowing frontier development would improve the outlook; an uncontrolled race among states or firms would worsen it. No single ordinary product release would settle the question. What matters most is evidence about control at scale and whether society can act on warning signs before systems become capable of hiding them.
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