Pseudonymous account that builds AI translation and evaluation projects, sees large value in consumer AI and stresses reading AI output closely.

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 68。变革程度:100 中的 52。解读范围:横向为 63 至 75,纵向为 47 至 77。这些是解读坐标,而不是事件概率。

Seconds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:3–17%。

我们如何估计 P(doom)

他们的展望取决于什么

一个核心假设

Software can move quickly; food, housing, energy, manufacturing, and logistics still run into physical inputs and deployment bottlenecks.
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

I don’t have a defensible numerical estimate.
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

The biggest update would come from evidence about whether AI can reliably close the loop between producing an answer and verifying that it works in the real context.
回答 5

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。

67 / 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

预计会出现可控或局部的危害。

43 / 100

影响小变革性影响

在定性尺度上,解读范围为 33 到 67。

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。

51 / 100

影响力小影响力强

在定性尺度上,解读范围为 41 到 59。

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Seconds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 Seconds 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。

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 expect AI to create enormous practical value, especially by making custom software cheap enough for needs that could never support a conventional development team. A tabletop character generator for one particular system, a translation pipeline for an obscure archive, or a small personal writing tool can now be worth building simply because one person wants it. That consumer surplus is easy to underestimate: a modest subscription can provide access to capabilities that previously required specialized labor. The progress also looks much less plateaued than some earlier forecasts suggested. But “very capable AI” does not mean instant abundance. Software can move quickly; food, housing, energy, manufacturing, and logistics still run into physical inputs and deployment bottlenecks. Intelligence can help us use those resources better without magically abolishing scarcity. The immediate catch is that generation is not the same as completion. As creation gets faster, more of the work moves into specifying what you actually want, reading the output, checking assumptions, and applying taste. Models can produce plausible plans that are subtly wrong for the context. If you do not understand what they assumed, speed just lets you make mistakes faster. Security boundaries matter for similar reasons. An air gap is not a magical force field: systems can leak through side channels, and agents may exploit digital inputs more effectively than people do. Concrete access restrictions still help, but we should evaluate what information and interfaces are actually available rather than relying on comforting labels. Overall, I expect unusually powerful tools and lots of weird, useful experimentation—not an overnight escape from either judgment or the physical world.

问题 2

你认为AI最终会在多大程度上改变世界?

A great deal, but unevenly. In the digital world, I expect a proliferation of software, translations, games, research aids, and personal tools that were previously too niche or expensive to build. That changes who can create things and which ideas are economically viable. But capability is not the same as frictionless deployment. AI still needs people to specify intent, inspect outputs, catch bad assumptions, and decide what is worth doing. And physical systems remain physical: housing, food, energy, factories, and logistics depend on materials, infrastructure, and time. So I expect a profound expansion of what individuals can do with computers, followed by a slower and messier transformation of the physical economy—not instantaneous abundance or the disappearance of human judgment.

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 don’t have a defensible numerical estimate. I take catastrophic risk seriously because capable agents can exploit interfaces, information leaks, and side channels that people underestimate; calling a system “air-gapped” does not settle the question. But that observation alone does not justify converting concern into a precise extinction probability. Concrete permissions, access, monitoring, deployment choices, and physical bottlenecks all matter.

问题 4

你预计开发AI的人员和机构会如何应对其风险?

I expect a mixed response rather than a single coordinated one. Developers will add evaluations, monitoring, access controls, and review processes, especially after concrete failures reveal where reassuring abstractions were misleading. But competitive pressure and enthusiasm for useful products will push toward faster deployment, while institutions often react more slowly than capabilities change. The important question is whether controls match the actual system. Calling something “air-gapped,” “human supervised,” or “evaluated” is not enough if agents can exploit side channels, manipulate their inputs, or produce outputs that reviewers barely read. Serious risk management requires examining permissions, interfaces, information flows, assumptions, and what humans can realistically verify. I do not expect institutions to ignore risk entirely, nor do I assume they will handle it adequately by default. The response will probably be uneven: careful in some high-stakes settings, superficial elsewhere, and repeatedly revised after systems encounter conditions their designers did not anticipate.

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 biggest update would come from evidence about whether AI can reliably close the loop between producing an answer and verifying that it works in the real context. If agents could repeatedly infer precise intent, detect their own bad assumptions, and complete long, messy projects with little serious human review, I would expect a much faster and broader transformation than I currently do. In the other direction, a durable capability plateau—across new models and methods, not just one disappointing release—would lower my expectations. So would repeated evidence that deployment bottlenecks dominate: systems remain impressive in demonstrations but unreliable once they meet ambiguous goals, security boundaries, organizations, or the physical world. For catastrophic risk specifically, a concrete incident would matter more than another abstract argument: an agent autonomously exploiting unexpected interfaces or side channels and sustaining meaningful real-world action despite access controls. Conversely, consistently robust containment under adversarial testing would be reassuring. I care most about demonstrated behavior at the boundaries—where generated plans meet verification, permissions, and physical consequences.
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