问题 1
Omar Khattab
x.com/lateinteractionAI researcher behind the DSPy framework who builds ways to program and optimize language-model systems and finds current models useful but brittle.
AI将如何改变世界?
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 69。变革程度:100 中的 51。解读范围:横向为 64 至 75,纵向为 4 至 100。这些是解读坐标,而不是事件概率。
Omar Khattab的 P(doom) · 推断
他们的展望取决于什么
一个核心假设
Factuality, consistency, and alignment are not properties of an isolated model call; they are properties of deployed systems containing retrieval, multiple calls, tools, control flow, and verification.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
The key uncertainty is whether we learn to manage these models much better.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。
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。
39 / 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。
69 / 100
影响力小影响力强
在定性尺度上,解读范围为 48 到 77。
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Omar Khattab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 Omar Khattab 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
I think AI’s future depends as much on how we program and optimize systems around models as on scaling the models themselves. Today’s models are extraordinarily knowledgeable and already create substantial everyday value, but they remain brittle on broad, multidimensional work. High benchmark scores and spectacular results on narrow, verifiable tasks should not be mistaken for generally superhuman competence.
The ambitious hypothesis is that we may be mismanaging capable models. Current scaffolds often rely on hand-written prompts, fixed workflows, and poor decomposition. If systems can learn how to break long-horizon tasks into useful subproblems, route information, invoke specialized components, and optimize the whole program against measurable objectives, existing models may support much stronger scientific and practical work. That is a research hypothesis, not an established forecast.
This also changes how I think about safety and reliability. Factuality, consistency, and alignment are not properties of an isolated model call; they are properties of deployed systems containing retrieval, multiple calls, tools, control flow, and verification. We should evaluate and optimize those complete systems. Sometimes a small specialized retriever can beat a much larger model on the task that actually matters, with better resource tradeoffs. So I expect progress to come not only from bigger models, but from better composition, specialization, optimization, and evaluation of AI programs.
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 have a confident net societal forecast. I expect substantial benefits because models already create real value, and better decomposition, retrieval, verification, and system-level optimization could expand that value considerably—possibly into long-horizon scientific and practical work.
But that upside remains partly hypothetical. Current systems are still brittle, insufficiently adaptive, and easy to overrate based on narrow, verifiable successes. Harms and failures also emerge from the complete deployment system, not just the underlying model. So my strongest expectation is rapid growth in useful but uneven capabilities, with the overall societal balance depending heavily on how well we design, evaluate, and govern the systems around them.
问题 3
你认为AI最终会在多大程度上改变世界?
Potentially a great deal, but I would not attach a confident magnitude to “ultimately.” Models already provide substantial everyday value, while remaining surprisingly brittle at broad, multidimensional work. That combination suggests large change without implying that current benchmark gains amount to general superhuman competence.
The key uncertainty is whether we learn to manage these models much better. If systems can learn effective task decomposition, coordinate specialized components, retrieve the right information, and optimize long-horizon programs against measurable objectives, then existing or future models could become far more consequential, including for scientific work. That is the mismanaged-geniuses hypothesis, not a settled conclusion.
So I expect meaningful and expanding change, but its scale will depend on advances in complete AI systems—not merely larger standalone models—and on whether those systems become reliable, adaptive, and governable in real deployments.
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 numerical probability. My work supports taking system-level safety seriously: catastrophic behavior would depend on the complete deployed system—models, tools, retrieval, control flow, permissions, verification, and institutions—not merely an isolated model’s alignment.
Current models are highly knowledgeable but still brittle and insufficiently adaptive on broad work. That observation does not establish that extinction risk is negligible, nor does narrow benchmark success establish that such systems are close to autonomous, generally superhuman capability. Better decomposition and scaffolding may unlock substantially stronger long-horizon performance, but that remains a research hypothesis. So I would not convert these claims into a gut-feel extinction percentage.
问题 5
你倾向于认为AI带来的益处多于危害,还是确实难以取舍?
I lean toward AI doing more good than bad, mainly because it already creates substantial practical value and better-designed systems could extend that value considerably. But this is not a confident net forecast: capabilities remain brittle, and deployment-level failures could be serious.
The decisive object is the whole system—models, retrieval, tools, control flow, verification, permissions, and institutions. If we improve decomposition and optimize these systems against meaningful objectives, the upside could be large. If we confuse narrow benchmark success with broad competence or treat standalone-model alignment as sufficient, outcomes could be much worse. So I lean positive, but conditionally rather than complacently.
你的立场在哪里?
描绘我的世界观

