问题 1
Shannon Sands
x.com/max_paperclipsAI practitioner who argues AI safety should rely on security engineering, monitoring and voluntary standards and expects human-AI teams to persist.
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 72。变革程度:100 中的 59。解读范围:横向为 67 至 77,纵向为 40 至 85。这些是解读坐标,而不是事件概率。
Shannon Sands的 P(doom) · 推断
他们的展望取决于什么
一个核心假设
Today’s systems remain unusually dependent on humans who can define vague goals, supply context, evaluate results, and catch failures.回答 2
如果这个假设实际并非如此,他们的展望会如何变化?
什么可能使其改变看法
I’d update if systems repeatedly succeeded at vague, high-stakes tasks without skilled human direction—and especially if they could evade sandboxing, monitoring, access controls, or correction in real deployments.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。
72 / 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。
56 / 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。
61 / 100
影响力小影响力强
在定性尺度上,解读范围为 37 到 88。
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。
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Shannon Sands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 Shannon Sands 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 the near future is mostly human-AI teams, not autonomous machine replacements. Frontier models are already powerful—especially when you can delegate work to subagents—but vague real-world tasks still depend heavily on skilled human direction. If success hinges on knowing exactly how to prompt, supervise, and correct the system, I’m not calling that AGI. A genuinely general system should be able to infer what users actually need, more like an expert does.
That gives me reasons for employment optimism: people who can direct these systems will become dramatically more capable, and human judgment will remain valuable. But optimism is not an excuse for sloppy security. The immediate risks include mundane, familiar failures—weak access controls, ordinary exploits, bad monitoring, and systems deployed without adequate sandboxing. Those deserve at least as much attention as speculative architectural horror stories.
I’d like to see independent, pro-AI practitioners doing practical safety work: security operations, monitoring, sandboxing, and useful interpretability. Voluntary standards modeled on technical standards bodies, combined with real product-quality and liability incentives, strike me as more promising than building a bureaucracy that requires everyone to profess existential-risk ideology. Extraordinary danger claims need evidence. Ordinary security failures are already here, and accountability for them matters.
问题 2
你认为AI最终会在多大程度上改变世界?
A lot, but I wouldn’t pretend to know the ultimate ceiling. AI already acts as a major force multiplier, and systems with subagents can handle increasingly large chunks of intellectual work. That could reshape most knowledge work, software, research, and how organizations operate.
Still, “huge change” does not automatically mean autonomous replacement of people or imminent AGI. Today’s systems remain unusually dependent on humans who can define vague goals, supply context, evaluate results, and catch failures. My default expectation is therefore a world transformed by capable human-AI teams rather than one where humans simply disappear from the loop.
The scale of the benefits will also depend on whether we solve boring but consequential problems: access control, sandboxing, monitoring, interpretability, and accountability for insecure products. A powerful tool deployed carelessly can create enormous disruption without requiring any exotic superintelligence story. So yes, I expect profound change—but the practical details of deployment and security will shape whether that change is broadly productive or needlessly destructive.
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 numeric probability, and assigning one by gut feel would imply more evidence than we have. Extraordinary extinction claims need evidence about concrete architectures, capabilities, and failure mechanisms—not just “AI is powerful, therefore doom.”
I do think serious catastrophe is possible, especially through ordinary operational failures at scale: insecure systems, stolen access, poor monitoring, reckless deployment, and weak accountability. Those are concrete enough to act on now. My priority would be conventional security engineering, sandboxing, monitoring, practical interpretability, and liability for unsafe products—not treating an unsupported p(doom) as the price of admission to AI safety.
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 if systems repeatedly succeeded at vague, high-stakes tasks without skilled human direction—and especially if they could evade sandboxing, monitoring, access controls, or correction in real deployments. Demonstrated autonomous planning, persistence, deception, resource acquisition, or replication would matter far more than scary benchmark extrapolations.
I’d also take architecture-specific evidence seriously: reproducible results showing why a system’s internal reasoning or behavior becomes unusually hard to inspect and constrain. But I’d want independent replication and concrete mechanisms, not claims that some new architecture is automatically an uncontrollable intelligence explosion.
Evidence cuts both ways. If strong security controls, practical interpretability, and monitored deployments reliably prevent dangerous behavior as capabilities increase, that would strengthen my confidence in human control. The key question is what systems actually do under realistic conditions—not whether someone can tell the most dramatic story about them.
你的立场在哪里?
描绘我的世界观
