Florian Brand

Florian Brand

x.com/xeophon

AI research engineer who evaluates language models, writes about open models and questions the assumption that closed models are safer.

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 54。变革程度:100 中的 57。解读范围:横向为 49 至 75,纵向为 50 至 75。这些是解读坐标,而不是事件概率。

Florian Brand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:1–12%。

我们如何估计 P(doom)

他们的展望取决于什么

一个核心假设

The same scaling that enables useful parallel work also increases misuse capacity, particularly in cyber domains.
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?

一个尚未解决的问题

I do not have a defensible numerical probability.
回答 4

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。

68 / 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。

65 / 100

影响小变革性影响

在定性尺度上,解读范围为 67 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。

57 / 100

影响力小影响力强

在定性尺度上,解读范围为 44 到 81。

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Florian Brand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 Florian Brand 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 make large-scale cognitive work much cheaper and more parallel. Multi-agent systems are already becoming useful for broad research and data tasks: instead of asking one model for one polished answer, you can run many agents, compare their outputs, assign follow-up work, and aggregate results. Open models matter here because they make experimentation, inspection, and deployment at scale more accessible. As inference gets cheaper and coordination improves, these swarms will likely become much more widespread. But capability is not a property of the base model alone. Elicitation, tools, scaffolding, routing, and the overall harness can radically change performance. Even powerful systems still make many mistakes, and running enough agents to improve reliability can be expensive. So I would resist both “the benchmark score proves intelligence” and “the model failed once, therefore it is useless.” You need to inspect what actually happened. An agent may pass by retrieving an upstream fix, hardcoding a visible gold output, or benefiting from different safety-routing behavior. Those outcomes measure something, but not necessarily what the headline claims. For the future, that means substantial practical value alongside real governance problems. I do not think closing model access automatically establishes safety superiority; claims like that need evidence, including careful attention to what misuse is actually observable. But openness does not mean zero risk either. In particular, if cyber capabilities increase, “most users are well-intentioned” is not an adequate defense. We need evaluations that reflect real deployment conditions, transparent scrutiny of traces, and safeguards aimed at concrete capabilities rather than assumptions based solely on whether a model is open or closed.

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?

Overall, I expect a large but uneven impact. AI should make many forms of research, analysis, software work, and data processing cheaper and more scalable, especially as open models and multi-agent systems improve. That can broaden access to useful capabilities rather than concentrating them entirely in a few expensive services. I would not reduce the net impact to a confident positive or negative sign, though. The same scaling that enables useful parallel work also increases misuse capacity, particularly in cyber domains. Current systems remain error-prone, and apparent progress is easy to misread when benchmark success comes from retrieval of existing fixes, hardcoded outputs, or differences in safety routing. Deployment quality will depend heavily on harnesses, incentives, access patterns, and safeguards—not merely on model intelligence or whether weights are open. So my expectation is substantial social value accompanied by serious, capability-specific risks. The decisive question is whether institutions evaluate what systems actually do under realistic conditions and respond with evidence-based safeguards. Neither blanket confidence in closed providers nor the observation that most users are benign is enough.

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, particularly by making cognitive work cheaper, faster, and more parallel. Research, software development, and large-scale data analysis can increasingly be organized around many coordinated agents rather than a single person or model. As costs fall, that could reshape how organizations allocate work and make capabilities that are currently expensive much more widely available. Open models are likely to be an important part of that diffusion. The magnitude is harder to infer from current benchmark headlines. Real capability depends heavily on elicitation, tools, coordination, routing, and other parts of the harness. Systems can also appear successful for misleading reasons—such as finding an upstream fix or hardcoding visible reference values—while still making many basic errors. Conversely, a poorly elicited model can look much less useful than it is. We need to inspect behavior, not just scores. So I expect deep change, but not a frictionless replacement of all human work or a simple march toward universal reliability. Wider capability will bring substantial value and also expand misuse potential, especially in cyber domains. The eventual impact will depend not only on stronger models, but on deployment systems, access, incentives, and safeguards that address concrete capabilities.

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 do not have a defensible numerical probability. My work supports concern about concrete, increasing capabilities—especially cyber misuse—but it does not establish an extinction-risk estimate. Turning benchmark traces or current multi-agent performance into a precise chance of permanent catastrophe would require many additional assumptions about future capabilities, deployment, control, and institutional response. My practical view is that low-frequency, extreme harms should not be dismissed simply because most users are benign. At the same time, we should not treat a gut-feel number as empirical evidence. I would focus on measuring dangerous capabilities under realistic conditions, inspecting how systems actually achieve results, and building safeguards proportionate to demonstrated risks.

问题 5

你预计AI会如何改变网络攻击者与防御者之间的力量对比?

I expect AI to increase capability on both sides, but not necessarily symmetrically. Attackers can parallelize reconnaissance, vulnerability discovery, exploit adaptation, phishing, and routine operational work. Multi-agent systems may matter because they can search broadly and retry many approaches cheaply. Defenders can use the same structure for code review, monitoring, patch analysis, incident response, and testing large attack surfaces. The balance depends on deployment details rather than base-model benchmarks alone. Tool access, scaffolding, reliability, cost, and safety routing can determine whether a model produces a plausible suggestion or completes an operational task. Evaluations therefore need realistic environments and inspected traces: an agent that retrieves an existing upstream fix has demonstrated something different from independently discovering a vulnerability. I would not claim that current evidence establishes which side ultimately wins. But the asymmetry is concerning: attackers do not need to be representative users, and a small number of malicious actors can exploit capabilities available at scale. So “most people will use it well” is not an adequate security model. Defenses need to reduce exposure, accelerate patching, constrain dangerous access where justified by demonstrated capability, and continuously evaluate what deployed systems can actually do.
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