Geoffrey Huntley

Geoffrey Huntley

x.com/geoffreyhuntley

Software engineer who created the Ralph loop technique for coding agents and argues that verifying real production behavior remains unsolved.

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 69。变革程度:100 中的 66。解读范围:横向为 64 至 75,纵向为 50 至 75。这些是解读坐标,而不是事件概率。

Geoffrey Huntley的 P(doom) · 推断

≈3%

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:2–8%。

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 bottleneck moves from typing code to building feedback that exposes failure and prevents the same failure from recurring.
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

I don’t have a defensible percentage.
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。

73 / 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。

55 / 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 49 到 76。

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Geoffrey Huntley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 Geoffrey Huntley 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 turns software engineering into the design of programmable improvement loops. Generation is effectively solved: the cost of exploring an idea, combining approaches, and producing candidate implementations has collapsed. That does not mean every experiment should ship. The bottleneck moves from typing code to building feedback that exposes failure and prevents the same failure from recurring. The practical future is not one giant autonomous agent improvising an entire business. It is model-driven loops combined deliberately with deterministic workflow stages: one clear task per loop, constrained environments, observable outcomes, and hard gates where determinism matters. Engineers remain responsible for verification. Tests, proofs, and actual production behavior are different things, and production correctness remains the unsolved part. Organizationally, this should let far more people contribute ideas and software instead of waiting behind traditional gatekeepers. But today’s scaffolding is temporary. Instructions and “skills” built around current model weaknesses need regular revalidation; standardizing them too early risks fossilizing obsolete workarounds. There is also a strategic issue. If a company hands its operational capability to a remote AI provider, it is accepting dependency on infrastructure that can change terms or disappear—including under sanctions or conflict. I expect local, open, reproducible models to matter for that reason. The future is cheaper exploration and broader participation, but only if we build strong verification loops and retain control of the machinery.

问题 2

你认为AI最终会在多大程度上改变世界?

Enormously. It changes the economics of turning ideas into software: generation becomes cheap and abundant, while verification, production correctness, and operational control become the scarce parts. That will reshape engineering roles, organizational boundaries, and who gets to contribute—not merely make existing programmers faster. But the depth of change depends on how we build around it. A giant improvising agent is not a software factory. The durable pattern is repeated model loops connected to deterministic stages, with failures feeding improvements back into the system. Organizations that learn to construct those loops will compound capability; those that merely bolt chat onto existing processes will get much less. It may also rearrange strategic power. If essential operations depend on a provider that can alter access or be cut off during sanctions or conflict, AI infrastructure becomes geopolitical leverage. Local, open, reproducible models therefore matter beyond cost or ideology. AI will change the world at the level of production and institutional power, but generating more artifacts is the easy part. Proving they work in the real world remains the hard part.

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 percentage. My concrete concern is nearer-term permanent dependency and systemic failure: organizations handing critical operations to opaque providers, weak verification being mistaken for production correctness, and access becoming geopolitical leverage. Cheap generation increases the volume of systems we can build—and therefore the number of failures we must detect. That is serious, but it does not justify inventing an extinction probability. The actionable response is to build constrained loops, deterministic gates, strong production verification, and local, reproducible capability where strategic autonomy matters.

问题 4

你预计AI会在什么时候给日常生活带来重大变化?如果你认为它会带来这种变化的话。

It is already beginning, but major everyday change arrives as organizations rebuild workflows around AI rather than merely adding chat interfaces. The key transition is when cheap generation becomes embedded in reliable loops: people describe an idea, systems produce candidate work, deterministic stages check what can be checked, and observed failures improve the next run. That will reduce gatekeeping around who can create software and contribute operational ideas. More people will be able to turn intent into working artifacts without waiting for a specialist to translate every step. But adoption will be uneven because production verification remains hard. Generating something plausible is immediate; proving it behaves correctly under real conditions is not. So I would not name a defensible year. Expect compounding change wherever feedback is fast and consequences are easy to verify, and slower change in domains where failures are costly or production behavior is difficult to observe. The visible breakthrough is not a magical autonomous agent. It is thousands of constrained, improving loops quietly becoming ordinary infrastructure.
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