Dex Horthy

Dex Horthy

x.com/dexhorthy

HumanLayer co-founder who writes about context engineering and argues reliable coding agents still need careful planning and humans who read 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 64。变革程度:100 中的 39。解读范围:横向为 50 至 75,纵向为 23 至 52。这些是解读坐标,而不是事件概率。

Dex Horthy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:1–10%。

我们如何估计 P(doom)

他们的展望取决于什么

一个核心假设

Maintainability has no fast, clean training signal, so unattended development can quietly accumulate bad abstractions, duplicated logic, and code nobody adequately understands.
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?

什么可能使其改变看法

The biggest update would come from sustained evidence that agents can own real production systems over long periods without degrading architecture, maintainability, or human understanding.
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。

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。

35 / 100

影响小变革性影响

在定性尺度上,解读范围为 33 到 33。

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 46 到 79。

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Dex Horthy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 Dex Horthy 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 will change software development substantially, but not by replacing engineering with one giant autonomous loop. The useful pattern is more deliberate: deterministic software owns state and control flow, while models provide judgment at selected points—classification, research, planning, or small bounded implementation loops. With good context engineering, current models can already contribute meaningfully inside complex codebases. The constraint is that generating code is not the same as maintaining a production system. Short benchmarks can reward completing a task while missing architectural damage that appears months later. Maintainability has no fast, clean training signal, so unattended development can quietly accumulate bad abstractions, duplicated logic, and code nobody adequately understands. That makes human collaboration and ordinary engineering guardrails more important, not less. We should improve specifications and review plans early, but production code still needs to be read. Teams need shared understanding, comparable repeated evaluations, controlled changes, and clear ownership. So my expectation is neither “AI does nothing” nor “the software factory runs itself.” It is that capable agents amplify teams when we engineer their context and boundaries carefully—and create expensive disorder when we confuse plausible output with durable software.

问题 2

你认为AI最终会在多大程度上改变世界?

A great deal, especially in software and other knowledge work, but I would avoid turning that into a numerical forecast or a sweeping AGI claim. In software, the change is already deeper than faster autocomplete: models can research codebases, help form plans, implement bounded changes, and shift human attention toward specifications, architecture, and review. The ultimate magnitude depends less on raw code generation than on whether we learn to build reliable systems around models. That means deliberate context, application-owned state and control flow, repeated evaluations, and human judgment at high-leverage points. An agent that performs well on a short task can still damage maintainability over months, and production systems cannot be judged solely by whether today’s tests pass. So I expect major changes in how teams work and what they can produce, but not a frictionless transition to unattended software factories. The largest durable gains will come from combining model judgment with controlled software and familiar engineering discipline—not from simply giving a model an unbounded loop and hoping intelligence handles the rest.

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 work supports concrete claims about coding-agent reliability and long-term software quality, not a calibrated extinction forecast. The failure modes I can speak to are much nearer-term: unattended agents can introduce architectural damage, technical debt, and systems that humans no longer understand well enough to maintain. Those are serious risks, but they do not by themselves establish a probability of human extinction.

问题 4

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 sustained evidence that agents can own real production systems over long periods without degrading architecture, maintainability, or human understanding. Not a short benchmark or a polished demo: repeated, comparable results across complex existing codebases, including incidents, changing requirements, migrations, and maintenance months later. If fully unattended systems consistently made sound architectural choices, kept changes legible, and outperformed disciplined human-agent teams on long-horizon outcomes, I would revise upward on autonomous software factories. Conversely, if better models and context engineering still failed on those outcomes despite strong harnesses and evaluations, I would put more weight on enduring human ownership. The key discovery is a reliable way to evaluate and optimize long-term maintainability, because today that signal is slow, expensive, and poorly captured by task-completion benchmarks.
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