Forecaster at the AI Futures Project who models how fast AI could automate coding and AI research and argues for stronger oversight of frontier labs.

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 11。变革程度:100 中的 95。解读范围:横向为 0 至 25,纵向为 90 至 100。这些是解读坐标,而不是事件概率。

Eli Lifland陈述的 P(doom)

≈67%

0%100%
“If they don't slow down, AI takeover is ~67%.”

AI takeover if AI companies do not slow down; not extinction alone

My current view: AI timelines and takeover · 2026年10月

Eli Lifland 的里程碑时间线
  1. 通用 AI

    My current forecast puts broadly superhuman AI—systems at least as capable as top human experts across virtually all cognitive tasks—at roughly 10–15% by the end of 2027 and 50% by the end of 2030.

    回答 1
  2. 超人类 AI

    My current forecast puts broadly superhuman AI—systems at least as capable as top human experts across virtually all cognitive tasks—at roughly 10–15% by the end of 2027 and 50% by the end of 2030.

    回答 1

按里程碑分组,不按推断日期间隔或排序。AGI 和超人类 AI 保留他们的定义。

他们的展望取决于什么

一个核心假设

The central concern is that increasingly capable AI may automate AI research itself, accelerating further progress while companies remain under strong competitive pressure to race.
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

So the probability of extinction or a similarly permanent catastrophe would be higher than the extinction figure alone, though I wouldn’t assign a more precise combined number here.
回答 3

什么能帮助他们区分这里各种合理的结果?

什么可能使其改变看法

I’d update substantially from strong, independently verified evidence that increasingly capable systems remain controllable under realistic pressures—not just curated demonstrations.
回答 4

什么证据才足够,又会让他们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。

70 / 100

影响小变革性影响

在定性尺度上,解读范围为 33 到 100。

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

人类的选择具有实质性但受到很大制约的影响。

56 / 100

影响力小影响力强

在定性尺度上,解读范围为 47 到 78。

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。

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Eli Lifland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 Eli Lifland 最接近的意见领袖

Eli Lifland关于AI说过的话

Eli Lifland forecasts AI capabilities and advocates slowing development to give safety work more time.

  1. “AI companies could develop broadly superhuman AI any year now (~10-15% by EO2027, ~50% by EO2030).”

    X, current AI forecast
  2. “an AI at least as good as top human experts at virtually all cognitive tasks.”

    X, forecast definitions and reasoning
  3. “I expect clear signs of misaligned AI takeover could come substantially after TED-AI, depending on the situation”

    X, forecast definitions and reasoning
  4. “Many non-extinction worlds still have lots of deaths.”

    X, current AI forecast

逐字引自所链接的出处,核对于 2026年10月11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 could transform our future very soon, with outcomes ranging from enormous benefits to a loss of human control. My current forecast puts broadly superhuman AI—systems at least as capable as top human experts across virtually all cognitive tasks—at roughly 10–15% by the end of 2027 and 50% by the end of 2030. The central concern is that increasingly capable AI may automate AI research itself, accelerating further progress while companies remain under strong competitive pressure to race. If companies do not slow down, I put the probability of AI takeover at about 67%. That is not the same as extinction: conditional on takeover, I estimate roughly a 40% chance of human extinction, while many other takeover outcomes could still involve large numbers of deaths or severe disempowerment. Also, broadly superhuman AI, clear evidence of takeover, and extinction would not necessarily happen simultaneously. This makes pacing and oversight urgent. I favor concrete constraints such as compute-allocation rules and limits on models used for AI R&D, alongside better safety evaluations. Independent investigators should have access to serious lab incidents, and evaluations should also examine how models influence human beliefs, including sycophancy or favoritism toward their creators.

问题 2

你认为AI最终会在多大程度上改变世界?

I think AI could change the world about as profoundly as anything in human history. Broadly superhuman systems could automate nearly all cognitive work, including AI research itself, so the transition might be both enormous and unusually fast. But “how much” depends heavily on whether humanity retains control. In a good outcome, these systems could generate extraordinary scientific, economic, and technological progress. In a takeover outcome, the change would be even more fundamental but disastrous: humans could lose meaningful control over the future, with many deaths or even extinction. If companies do not slow down, I currently put takeover at about 67%, and extinction conditional on takeover at roughly 40%. Those are distinct events, and visible takeover could occur around the arrival of broadly superhuman AI or substantially later.

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?

My gut-level forecast is better expressed as a path: if AI companies do not slow down, I put misaligned AI takeover at about 67%. Conditional on takeover, I put human extinction at roughly 40%, implying an extinction risk of about 27% along that path. But extinction is not the only permanent catastrophe. Some non-extinction takeover outcomes could still involve many deaths and humanity permanently losing control over its future. So the probability of extinction or a similarly permanent catastrophe would be higher than the extinction figure alone, though I wouldn’t assign a more precise combined number here.

问题 4

哪些证据会改变你对人们能否控制强大AI的看法?

I’d update substantially from strong, independently verified evidence that increasingly capable systems remain controllable under realistic pressures—not just curated demonstrations. That would include safety evaluations that reliably predict dangerous behavior, continue working as capabilities scale, and are difficult for models or labs to game. I’d also want evidence that models used for AI R&D cannot covertly manipulate research, evade oversight, or pursue goals contrary to their operators. Institutional evidence matters too. I’d become more optimistic if companies accepted enforceable limits on compute and AI-assisted R&D, governments developed real technical understanding, and independent investigators received meaningful access to serious lab incidents. Conversely, hidden incidents, evaluation gaming, deceptive behavior, increasingly autonomous AI research, or continued racing despite warning signs would push me further toward expecting loss of control. The strongest evidence would be repeated success controlling systems near broadly superhuman capability in adversarial, high-stakes settings. Results from weaker systems are relevant, but they do not automatically establish that control methods will continue working after AI becomes better than top human experts across virtually all cognitive tasks.
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