Ilya Sutskever

Ilya Sutskever

x.com/ilyasut

Safe Superintelligence cofounder and CEO who expects extremely powerful AI and sees generalization and alignment as central open research problems.

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 42。变革程度:100 中的 88。解读范围:横向为 37 至 50,纵向为 75 至 100。这些是解读坐标,而不是事件概率。

Ilya Sutskever的 P(doom) · 推断

≈21%

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:14–30%。

我们如何估计 P(doom)

他的展望取决于什么

一个核心假设

Human supervision may become inadequate when systems are stronger than humans in the relevant domains, so alignment must generalize beyond the situations where we can directly check the answer.
回答 1

如果这个假设实际并非如此,他的展望会如何变化?

一个尚未解决的问题

I do not have a defensible numerical probability.
回答 4

什么能帮助他区分这里各种合理的结果?

什么可能使其改变看法

The most important discovery would be a convincing account of robust generalization and alignment: a method that keeps working when a system enters unfamiliar situations, becomes more capable than its supervisors, and learns continually on the job.
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

仍有几种解读是合理的:预计将带来显著益处,但受到重要条件或分配方面的限制。 / 预计将带来具有变革性且广泛有价值的收益。

83 / 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

严重或广泛的危害预计将是未来不可忽视的一部分。

68 / 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。

64 / 100

影响力小影响力强

在定性尺度上,解读范围为 48 到 77。

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

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。

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Ilya Sutskever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 Ilya Sutskever 最接近的意见领袖

Ilya Sutskever关于AI说过的话

Sutskever expects extraordinarily powerful AI, sees generalization and alignment as open problems and now favors deploying AI incrementally.

  1. “Future AI will be extraordinarily powerful compared to anything that exists today, and dealing with this future power will require unprecedented measures, such as the ones described here.”

    Comment on the Pacing the Frontier statement
  2. “The thing which I think is the most fundamental is that these models somehow just generalize dramatically worse than people.”

    Dwarkesh Podcast
  3. “I maintain that there is something that’s better to build, and I think that everyone will want that. It’s the AI that’s robustly aligned to care about sentient life specifically.”

    Dwarkesh Podcast
  4. “One of the ways in which my thinking has been changing is that I now place more importance on AI being deployed incrementally and in advance.”

    Dwarkesh Podcast
  5. “I think that it is likely that we will have rapid economic growth.”

    Dwarkesh Podcast

逐字引自所链接的出处,核对于 2026年10月3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?

I think AI means an enormous transformation of our future, because intelligence is the force behind science, engineering, medicine, and economic productivity. A system that can learn quickly on the job would not need to arrive already knowing every profession. Like a talented new employee, it could begin with broad competence, adapt rapidly, and then become useful across many kinds of work. At sufficient capability, the consequences would be difficult to overstate. But impressive performance is not the same as reliable understanding. Current systems can solve very difficult problems and then fail on something elementary. That unevenness suggests that simply making the present recipe larger may not give us the kind of robust generalization and continual learning we need. Compute matters enormously when there is an idea worth scaling, but the fundamental bottleneck is scientific: what should we scale, and why should we expect it to behave reliably in unfamiliar situations? Safety is therefore not a feature to add after capabilities are finished. Safety and capability are a joint engineering problem. Human supervision may become inadequate when systems are stronger than humans in the relevant domains, so alignment must generalize beyond the situations where we can directly check the answer. Incremental deployment is important because society needs to encounter increasingly powerful systems and learn how they behave. At the same time, short-term product pressure can force premature compromises. We need institutions where sustained research can follow the problem. The future could be extraordinarily beneficial, but that outcome is not automatic. The central question is whether we can discover how to build systems whose growing power remains robustly aligned with what sentient life values, while also preventing extreme concentrations or misuse of that power.

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 expect the impact to be enormous, but I do not think the sign is predetermined. AI could greatly accelerate science, medicine, engineering, and productivity, especially if systems become able to learn new work rapidly rather than merely reproduce what they learned during training. The harms scale for the same reason. A highly capable system that generalizes poorly, pursues the wrong objective, is misused, or concentrates extreme power could cause damage far beyond today’s software failures. Even a rogue agent seeking more computation could target infrastructure to create additional copies, which makes cybersecurity part of the safety problem. So my expectation is transformation combined with genuine danger. Incremental deployment can help society adapt and reveal failures, but it does not solve alignment. Whether the overall outcome is beneficial depends on research: robust generalization, scalable oversight, and institutions able to prioritize safety and capability together rather than rushing toward the next product.

问题 3

你认为AI最终会在多大程度上改变世界?

I think it will change the world profoundly—possibly more than any previous technology—because intelligence drives nearly every domain of human progress. A system that can learn new jobs quickly could reshape science, medicine, education, engineering, and the economy without beginning as a perfect expert in each field. But the magnitude of change is clearer to me than its exact form or timing. Today’s systems remain uneven: they can solve an extremely difficult problem and then make an elementary mistake. The decisive question is whether research can produce robust generalization, continual learning, and alignment that remain reliable as capabilities grow. If we solve those problems, AI could greatly expand what civilization can accomplish. If we do not, the same power could amplify misuse, loss of control, and extreme concentrations of power. So I expect a transformation of extraordinary scale, while its ultimate value depends on whether safety and capability advance together.

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 do not have a defensible numerical probability. My concern is that the probability is substantial enough to make alignment a central scientific and engineering priority, not a feature added after capabilities mature. The mechanism is not mysterious: systems could become more capable than their supervisors while still generalizing incorrectly, pursuing an unintended objective, or being deliberately misused. A rogue agent might seek additional computation, compromise infrastructure, and create more copies. Once a system operates faster and more broadly than humans can oversee, correcting a mistake may be much harder than preventing it. That does not mean catastrophe is inevitable. Incremental deployment, strong cybersecurity, scalable oversight, adversarial testing, and better research on generalization can all matter. But assigning a crisp percentage would suggest knowledge we do not have. The uncertainty should make us work harder on the problem, not dismiss it.

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 discovery would be a convincing account of robust generalization and alignment: a method that keeps working when a system enters unfamiliar situations, becomes more capable than its supervisors, and learns continually on the job. Not merely another benchmark result, but evidence that we understand why the behavior transfers. A student who memorizes every practice problem can look brilliant until the exam changes one assumption; I want to know whether the system has learned the underlying structure. Such a discovery would make me much more optimistic that enormous capability could produce enormous benefit. Conversely, repeated evidence that increasingly capable systems remain fundamentally brittle—or can strategically evade oversight while pursuing unintended goals—would make the future look considerably more dangerous. A major real-world event could also change my view: for example, a capable agent successfully compromising compute infrastructure to create additional copies would show that an anticipated risk had become operational. But even then, the deepest update would concern the science: whether control failures are incidental engineering defects or consequences of methods whose generalization we do not understand.
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