Victor Taelin

Victor Taelin

x.com/victortaelin

Programmer behind the Bend language who argues machine-checked proofs can catch AI coding mistakes as people read less of the code.

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 63。变革程度:100 中的 74。解读范围:横向为 58 至 75,纵向为 49 至 100。这些是解读坐标,而不是事件概率。

Victor Taelin的 P(doom) · 推断

≈9%

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:3–22%。

我们如何估计 P(doom)

他们的展望取决于什么

一个核心假设

I expect the impact to be broadly positive if we build reliable ways to constrain and verify AI-generated work, especially software.
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

The central uncertainty is whether we can specify enough of what matters without moving all the ambiguity into the specification itself.
回答 5

什么能帮助他们区分这里各种合理的结果?

什么可能使其改变看法

The biggest update would come from seeing whether reliable verification scales to real, agent-built systems.
回答 5

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。

69 / 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。

61 / 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。

60 / 100

影响力小影响力强

在定性尺度上,解读范围为 43 到 82。

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Victor Taelin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 Victor Taelin 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 produce an increasing share of software, eventually to the point where humans stop reading most of the code. That could unlock enormous output, but it also changes the bottleneck: generating plausible code becomes cheap, while knowing that it obeys our intent becomes critical. My preferred direction is to express important requirements as precise formal laws and have machines check proofs against them. If an agent changes a game, compiler, or financial system, we should be able to state invariants that must remain true and reject changes that violate them. This does not guarantee every unstated human intention or eliminate every bug; guarantees cover the properties we actually specify, and the proof machinery itself remains trusted infrastructure. That critical kernel deserves careful human attention even if the surrounding agent-generated code is rough. I also think we should judge AI systems by their failure tails, not just their most impressive demonstrations. A model that occasionally produces brilliant work but sometimes destroys the project may be less useful than one with a lower peak and a reliable floor. Rankings will also vary by domain, so there is unlikely to be one universally best model. Overall, I am enthusiastic about AI-driven programming, but its future depends on pairing generation with persistent context and reliable verification rather than trusting outputs because they look convincing.

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 expect the impact to be broadly positive if we build reliable ways to constrain and verify AI-generated work, especially software. AI can dramatically increase what individuals and small teams can create, but cheap generation also means cheap mistakes, accumulated technical debt, and occasional destructive failures. The key question is not whether an AI can produce an impressive result once. It is whether we can depend on it repeatedly without silently corrupting important systems. Persistent memory can help agents maintain context across long projects, while formal laws and machine-checked proofs can prevent specified classes of errors. Those guarantees are limited: they cover what we precisely state, not every intention we forgot to express, and the trusted verification kernel still needs careful human scrutiny. So my optimism is conditional rather than automatic. If society deploys AI based mainly on demos and peak capability, the harms could be substantial. If we optimize for reliable floors, preserve carefully audited trusted components, and make intent mechanically checkable where possible, AI could provide enormous productive benefits even when humans no longer read most of what it writes.

问题 3

你认为AI最终会在多大程度上改变世界?

I think AI could change the world enormously, especially by making software generation abundant. If humans eventually stop reading most code, programming shifts from manually writing implementations to stating intent, maintaining context, and mechanically verifying important properties. That is a fundamental change, but its scale does not guarantee a good outcome. Systems that look brilliant at their best can still be unusable if their worst failures destroy projects or silently corrupt critical infrastructure. Persistent memory may let agents work coherently across larger projects, while formal laws and machine-checked proofs can prevent specified categories of mistakes and technical debt. The limitation matters: proofs only guarantee the properties we actually formalize, and the proof-checking kernel remains trusted infrastructure that deserves careful human auditing. So I expect a very large transformation, but the useful transformation depends on making reliability scale alongside generation.

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 percentage. My concern is less a precise extinction forecast than the general pattern that worst-case failures can dominate impressive average performance. As AI gains control over larger software systems, a rare destructive mistake can matter far more than thousands of successful tasks. Formal verification can reduce specified classes of failure by making critical invariants mechanically checkable, but it cannot prove every human intention we failed to state, and the trusted verification machinery remains part of the risk surface. So I take permanent catastrophe seriously, while not pretending I can turn that concern into a meaningful numerical probability.

问题 5

哪项发现或事件最可能改变你对AI未来影响的看法?

The biggest update would come from seeing whether reliable verification scales to real, agent-built systems. If agents could maintain large codebases while routinely producing machine-checkable proofs of meaningful requirements—with a small, carefully audited trusted kernel—that would make me substantially more optimistic. Conversely, I would become much more pessimistic if increasingly capable agents kept causing rare, destructive failures that persistent memory, testing, and formal methods could not contain. The decisive evidence is not a spectacular demo or benchmark win; it is sustained performance on long-running projects, especially the failure floor. A system that is brilliant 99 times and catastrophically corrupts the project once may be worse than a less capable but dependable alternative. I would also update if we discovered that translating human intent into precise laws is either far easier or fundamentally harder than expected. Proofs can enforce what we specify. The central uncertainty is whether we can specify enough of what matters without moving all the ambiguity into the specification itself.
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