Martin Casado

Andreessen Horowitz general partner who is bullish on AI, treats safety as systems engineering and favors rules on harmful uses over model limits.

AI将如何改变世界？

模拟位置解读范围

横向：他表达的 Doom–Bloom 前景看法。纵向：变革程度。

Martin Casado的 P(doom) · 推断

≈3%

0%100%

根据他的模拟回答推断，并非他们给出的数字。合理范围：2–7%。

他的展望取决于什么

一个核心假设

Capital can now be turned into capability and usage unusually quickly: better models enable better products, those products generate demand, and that demand funds more infrastructure and development.

回答 1

如果这个假设实际并非如此，他的展望会如何变化？

一个尚未解决的问题

I don’t have a defensible number.

回答 3

什么能帮助他区分这里各种合理的结果？

什么可能使其改变看法

A repeatable demonstration that a development method creates a genuinely new, uncontainable risk—not just a stronger version of familiar cyber or software risk—would change my view most.

回答 4

什么证据才足够，又会让他的观点朝哪个方向转变？

相似的世界观

模拟世界观与 Martin Casado 最接近的意见领袖

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 is the most exciting platform shift I’ve seen and probably the biggest wealth unlock since the 1990s. Capital can now be turned into capability and usage unusually quickly: better models enable better products, those products generate demand, and that demand funds more infrastructure and development. I don’t think all the value stays with a handful of frontier labs. Over time, supply constraints should ease, open and long-tail models should handle more usage, and applications should capture more of the economics. I also don’t buy the jump from rapid progress to extinction. There’s an enormous gap between dismissing models as “stochastic parrots” and assuming unlimited, unstoppable intelligence growth. AI helping improve kernels, tools, or future AI systems is economically important, but calling every autocatalytic effect “recursive self-improvement” smuggles the conclusion into the terminology. The real risks are more familiar and more actionable. Cyber capability will create genuinely new pressure, but that’s a systems-engineering problem involving containment, permissions, monitoring, and explicit trade-offs—not mysticism. Computing has survived some very ugly security eras before, and AI may finally force us to build secure systems all the way down. My biggest concern is that doomsday messaging triggers hysteria and heavy-handed regulation. We should punish harmful uses under existing law, identify actual marginal risks, and add targeted rules where evidence supports them. Vague controls on model development will age badly, create loopholes, kneecap startups and open source, and hand an advantage to China.

问题 2

Setting aside good or bad: how much do you think AI will ultimately change the world—a little, a lot, or completely?

A lot—on the scale of a major computing platform shift. I expect it to reshape software, security, research, business formation, and how capital turns into productive capability. But “a lot” is not the same as “completely.” I don’t see evidence that it abolishes ordinary economics, institutions, physical constraints, or human agency. The jump from transformative technology to an unstoppable intelligence that replaces everything is exactly the kind of unsupported leap I reject.

问题 3

What’s your rough gut-feel chance that AI causes human extinction or a similarly permanent catastrophe? A number or “no idea” is fine.

I don’t have a defensible number. I think near-term extinction claims are fringe and badly overplayed, not a sound basis for policy.

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？

A repeatable demonstration that a development method creates a genuinely new, uncontainable risk—not just a stronger version of familiar cyber or software risk—would change my view most. For example, clear evidence of autonomous capability growth that defeats known controls and materially escapes physical, economic, and institutional constraints would force a different conversation. But it has to be demonstrated, not asserted through vague terms like “recursive self-improvement.” AI improving kernels or helping researchers build better models is an important autocatalytic effect; tools have long helped us build better tools. That alone does not establish runaway intelligence or extinction risk. On the economic side, I’d also update if frontier labs retained durable control despite easing supply constraints—if open models and applications consistently failed to capture meaningful usage and value. That would change my view of where the wealth accrues, though not by itself turn me into a doomer.

来源

用于为此模拟用户提供事实依据的文章、访谈和著述。

Aaron Levie, Steven Sinofsky & Martin Casado: How Do You Secure a World of AI Agents?

Calls Dario Amodei’s pacing post sensible and pragmatic but its atmospherics broken: pacing is orthogonal to security, placates the pause camp without satisfying it, and cannot be reconciled with talk of species extinction. Says labs should address x-risk directly. Drawing on his Lawrence Livermore weapons work, argues that if the most knowledgeable insiders believed in existential risk the answer would be nationalization with proven controls; since he says most do not, it is a recruiting and retention problem. Unlabeled automatic transcript: only turns attributable by context, a speaker-labeled clip and his own posts are used; third-party summaries conflict on some attributions.

podscripts.co

Martin Casado on Where the Value Is Going in AI

Sets out cases for and against frontier labs capturing everything. Explicitly guessing, he expects supply constraints to ease around 2028, large labs to keep about 80% of dollar-weighted share while about 60% of tokens go to long-tail and open models, and applications to capture more value. Distinguishes autocatalytic use of AI to build AI from recursive self-improvement, calls AI the biggest wealth unlock since the 1990s and says he is very bullish. Automatic transcript; guest turns inspected.

podscripts.co

To Regulate AI Effectively, Focus on How It’s Used

Argues for regulating harmful uses under existing law and studying marginal risk before new development rules, since AI has no stable definition and development rules invite loopholes. Says a demonstrably uncontainable new risk would change the conversation but has not been shown. Calls the precautionary principle bad for innovation, rejects the social-media analogy, and says regulatory uncertainty has chilled US open-source releases while Chinese open models dominate startup use. Full speaker-labeled transcript inspected.

a16zpolicy.substack.com

Base AI Policy on Evidence, Not Existential Angst

Older authored essay, first published in Fortune. Defines marginal risk as a new class of risk requiring a policy shift, says AI marginal risk remains a research question, cites GPT-2 and election deepfake fears as overblown, and concludes that AI appears tremendously safe and that heavy investment might be better policy than encumbrance. Full essay inspected; newer 2026 statements take precedence where they add cyber risk or political specifics.

a16z.com

The labs, pacing and the safety complex

Says the labs face impending, unsophisticated federal regulation and a well-financed doomer complex, and that applauding their governance attempts, doubting the proposal and criticizing the wider safety complex are compatible positions not to be conflated. Sets aside near-term extinction rhetoric, sees ordinary security implications, places himself in the “not” camp on pacing or an independent oversight body, and calls heavy-handed federal regulation disastrous but likely without industry coordination. Says he has no special knowledge. Full long-post text inspected via the X API.

x.com

Doomsday messaging and political backlash

Says he is not worried about a regulatory-capture conspiracy by the labs but about a doomsday-cult style of advocacy for sensible engineering and governance provoking mass hysteria and a backlash that sets the industry back a decade. A forecast of political risk, not a claim that safety work is unnecessary. Full text inspected via the X API.

x.com

If you believed the doomsday rhetoric

Self-reply stating that he does not believe the doomsday rhetoric, and that anyone who does should consistently lock the work in a Department of Energy lab with security clearances. A conditional argument about consistency, not his own policy proposal. Full text inspected via the X API.

x.com

Safety as systems engineering versus alignment

Describes a disconnect between people who treat AI safety as a systems-engineering problem with clear trade-offs and those who treat it as an alignment problem, where trade-offs cannot be made on something not understood or controlled. His other posts place him on the systems side; this post itself is descriptive. Full text inspected via the X API.

x.com

Covert channels are an old problem

Partly defends a colleague’s air-gap exfiltration scenario as worth considering, citing a long history of covert channels, but says it is not a net new concern and that the field understands it well. Illustrates his marginal-risk method: compare an AI threat with known security problems. The quoted post belongs to another speaker. Full long-post text inspected via the X API.

x.com

The x-risk discussion is tempering

Reports that industry discussion is becoming more nuanced: novel cyber risk appears real but is the kind of problem the industry has handled before, systems can be deployed pragmatically and safely at scale, and near-term extinction fears are increasingly seen as fringe. An account of his conversations, not survey evidence. Full text inspected via the X API.

x.com

Between stochastic parrots and unstoppable intelligence

Agrees with a quoted Jack Clark post that the “stochastic parrot” framing misled people about AI progress, and adds that there is a big gap between that framing and unlimited, unstoppable intelligence growth. Rejects both extremes; it does not supply a capability timeline. The quoted claim is Clark’s. Full text inspected via the X API.

x.com

AI cyber capability could force secure systems

Hopes AI cyber capabilities will force the industry to build genuinely secure systems all the way down, pointing to prior secure-systems research. An optimistic engineering response to a real risk, not a claim that current systems are secure. Full text inspected via the X API.

x.com

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

AI将如何改变世界？

更多详情

相似的世界观

模拟评估

你认为AI对我们的未来意味着什么？为什么？

Setting aside good or bad: how much do you think AI will ultimately change the world—a little, a lot, or completely?

What’s your rough gut-feel chance that AI causes human extinction or a similarly permanent catastrophe? A number or “no idea” is fine.

哪项发现或事件最可能改变你对AI未来影响的看法？

来源