Kyle Mistele

Kyle Mistele

x.com/0xblacklight

Software engineer who writes about configuring coding agents, arguing for focused context, careful harness design and clear security boundaries.

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 49。变革程度:100 中的 65。解读范围:横向为 44 至 54,纵向为 41 至 84。这些是解读坐标,而不是事件概率。

Kyle Mistele的 P(doom) · 推断

≈5%

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:2–15%。

我们如何估计 P(doom)

他们的展望取决于什么

一个核心假设

Agents can produce complex code faster than teams can understand or maintain it, so technical debt may compound faster than models improve at navigating bad codebases.
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

But “ultimately” is doing a lot of work—I could not responsibly predict the endpoint or timetable.
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。

70 / 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

严重或广泛的危害预计将是未来不可忽视的一部分。

58 / 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

人类的选择可以大幅改变AI的发展轨迹。

66 / 100

影响力小影响力强

在定性尺度上,解读范围为 49 到 76。

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Kyle Mistele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 Kyle Mistele 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 think AI could be as consequential as fire—not because I can predict some specific endpoint, but because software can increasingly operate on meaning rather than only rigidly specified inputs. That opens genuinely new capabilities, especially in programming, tool use, and difficult debugging. At the same time, I still would not pretend to know exactly where this goes over a five-year horizon, much less attach a defensible numerical probability to catastrophe. The near-term future will depend heavily on engineering discipline. Agents can produce complex code faster than teams can understand or maintain it, so technical debt may compound faster than models improve at navigating bad codebases. Giving an agent more context, more tools, and more automated reviewers does not automatically solve that. Large contexts can weaken instruction adherence, adversarial review can invent defects, and tools introduce ordinary security problems involving trust boundaries, dependencies, prompt injection, and credentials. So I expect AI to expand what software can do, but the quality of the outcome will depend on deliberate program design, focused context, suitable verification, and well-designed harnesses. Better models matter; human understanding of the systems around them still matters just as much.

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 don’t think a single confident positive-or-negative forecast is justified. AI may be enormously consequential and unlock valuable semantic software, stronger programming tools, and solutions to difficult problems. But those gains can arrive alongside brittle systems, security failures, and technical debt created faster than people can understand it. The overall impact will depend less on raw model capability than many people assume. It will depend on whether organizations preserve human understanding, design focused contexts and effective harnesses, verify outputs appropriately, and treat tools, dependencies, and credentials as real trust boundaries. More autonomy, context, and automated review are not inherently better; each can amplify failure as easily as capability. So I expect substantial benefits and substantial harms, with engineering practice shaping how much of each we get. I would not convert that into a numerical doom probability or a neat net-impact score. Those numbers are not realistically calculable or testable.

问题 3

你认为AI最终会在多大程度上改变世界?

Potentially as much as fire: not merely improving existing software, but changing what software can represent and do by operating on meaning. That could reshape programming, knowledge work, and how people interact with complex systems. But “ultimately” is doing a lot of work—I could not responsibly predict the endpoint or timetable. The magnitude of change also should not be confused with uniformly good outcomes. Agents can generate complexity and technical debt faster than organizations can understand it. Larger contexts, more tools, and automated review can amplify mistakes, security failures, and false diagnoses as readily as capability. Better models do not eliminate the need for deliberate architecture, focused context, verification, and ordinary security boundaries. So I expect very large change, but I would resist turning that into a precise forecast about AGI, catastrophe, or net social benefit. The engineering around these systems will substantially determine what that change looks like.

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 don’t have a defensible percentage. My gut feeling is not a measurement, and presenting it numerically would create false precision for an outcome we cannot realistically calculate or test. That is not the same as saying the risk is zero. AI systems can amplify serious failures through autonomy, tool access, malicious dependencies, mishandled credentials, prompt injection, and poorly understood code. But moving from those concrete engineering risks to a quantified probability of human extinction requires assumptions about future capabilities, deployment, institutions, and failure chains that we cannot validate. I would rather discuss specific threat models and mitigations than manufacture a doom number.
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