Pseudonymous account behind the Entropix sampling project that posts about open base models and using AI to strengthen cyber defenses.

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 67。变革程度:100 中的 46。解读范围:横向为 50 至 75,纵向为 41 至 53。这些是解读坐标,而不是事件概率。

xjdr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:2–10%。

我们如何估计 P(doom)

他们的展望取决于什么

一个核心假设

Effective capability depends on the complete system: model, problem specification, interaction, tools, harness, and operational environment.
回答 3

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 “how much” is not something I can responsibly reduce to a percentage or a dramatic milestone.
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。

64 / 100

影响力小影响力强

在定性尺度上,解读范围为 48 到 77。

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。

AI访问权限

限制对强大AI的访问。

模拟位置:允许访问,但须遵守能力或用途限制。

支持广泛或开放地访问强大AI。

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xjdr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 xjdr 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 increasingly becomes a powerful engineering tool, but its real capability will depend on the whole system around it: how clearly we specify the problem, how patiently we interact with it, what tools it can use, and whether the harness helps rather than hides weaknesses. Quick prompting often produces misleading judgments about what a model can or cannot do. At the same time, AI does not make intrinsically hard engineering—like reliable distributed systems—magically easy. I want open base models and meaningful research access to remain part of that future. They let more people inspect behavior, experiment with inference techniques, and build defensive capabilities. I am skeptical of responding to dangerous capabilities by reflexively restricting research and defense together. Frontier systems can be used proactively to find bugs and harden infrastructure. That does not mean treating every application casually. Offensive cyber agents should be evaluated with strong isolation: air gaps or heavily restricted networks, monitored egress, layered syscall controls, and defense in depth. Powerful tools need containment appropriate to what they can do. So my expectation is neither automatic utopia nor automatic catastrophe. AI’s impact will be shaped by careful problem formulation, access to useful models, disciplined experimentation, and serious operational security. I would also keep the experimental caveat explicit: impressive anecdotes can reveal possibilities, but they are not controlled comparisons or numerical forecasts about society as a whole.

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 defensible single-number or net-impact forecast. I expect substantial benefits where AI amplifies engineering, research, debugging, and defensive security, especially when people specify problems carefully and give models appropriate tools. Open base models and research access can broaden that experimentation rather than concentrating it behind a few interfaces. The harms are also real, particularly when offensive capabilities are connected to live systems without serious containment. Cyber agents need restricted networks, monitored egress, layered syscall controls, and defense in depth. Poor harnesses can create another kind of failure: unnecessary orchestration may conceal weaknesses in training and make systems look more reliable than they are. Overall, I expect the outcome to depend heavily on deployment choices rather than on “AI” as one fixed object. My bias is toward using frontier capabilities proactively to find bugs and harden systems, while isolating genuinely dangerous workflows—not broadly restricting the same research access needed to understand and defend against them. AI can improve our leverage, but it does not abolish hard engineering or operational discipline.

问题 3

你认为AI最终会在多大程度上改变世界?

I expect AI to change the world substantially, especially by increasing leverage in engineering, research, debugging, and security. But “how much” is not something I can responsibly reduce to a percentage or a dramatic milestone. Effective capability depends on the complete system: model, problem specification, interaction, tools, harness, and operational environment. That also means the change will be uneven. Careful, patient use can uncover capabilities that quick evaluations miss, while intrinsically difficult work—such as reliable distributed systems—remains difficult. AI can accelerate parts of the process without eliminating the underlying engineering constraints. Open base models and research access could distribute these gains and support experimentation, including defensive security work. At the same time, offensive cyber agents need strong isolation, monitored egress, layered controls, and defense in depth. So I expect major change, but not magic: more human leverage, new risks, and a continuing premium on disciplined engineering and deployment.

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 concrete concern is powerful offensive capability deployed without containment—especially cyber agents connected to live systems. Those require restricted networks, monitored egress, layered syscall controls, and defense in depth. That establishes a serious operational risk, not a specific extinction forecast. My instinct is to reduce risk through isolation, monitoring, and proactive use of frontier systems to find vulnerabilities and harden infrastructure, rather than treating broad restrictions on research access as a substitute for security engineering.
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