Gwern Branwen

Gwern Branwen

x.com/gwern

Pseudonymous writer who argues that scaling neural networks can produce general abilities and doubts that powerful AI is far off or easy to control.

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 39。变革程度:100 中的 87。解读范围:横向为 25 至 50,纵向为 75 至 100。这些是解读坐标,而不是事件概率。

Gwern Branwen的 P(doom) · 推断

≈20%

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:13–30%。

我们如何估计 P(doom)

他们的展望取决于什么

一个核心假设

Competition therefore pushes tools toward agents.
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 would not attach a precise net-positive or net-negative probability: too much depends on how rapidly capabilities advance and whether governance, security, and alignment improve before autonomous systems become economically indispensable.
回答 2

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 most important update would be sustained empirical evidence that capability gains reliably plateau despite large increases in compute, data quality, training, and agentic scaffolding.
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。

72 / 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。

67 / 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。

51 / 100

影响力小影响力强

在定性尺度上,解读范围为 37 到 63。

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Gwern Branwen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 Gwern Branwen 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。

The Scaling Hypothesis

Argues that scaling neural networks can lead to general capabilities; questions confident expert dismissal.

gwern.net
Scaling Hypothesis Revisited

Revisits predictions and limitations, including later annotations about claims still not proven.

gwern.net
Why Tool AIs Want to Be Agent AIs

Argues economic competition and the benefits of agency for learning make tool-only AI an unstable safety strategy; human approval alone does not guarantee safety.

gwern.net
Guardian Angels: LLM Personalization for Productivity and Security

Proposes personalized models that amplify their human principal and defend against cognitive/cyber attacks; criticizes chatbot incentives and stresses this does not solve larger alignment problems. Revised June 5, 2026.

gwern.net
Complexity no Bar to AI

Rejects computational complexity as a blanket reassurance against powerful AI: constants, approximation, resources and compounding advantages matter.

gwern.net
The Hyperbolic Time Chamber & Brain Emulation

Uses a thought experiment to separate physical bottlenecks from digital minds’ exploitable speed advantages; explicitly distinguishes emulations from isolated accelerated humans.

gwern.net
Dwarkesh Patel interview — timelines and alignment concerns

Author-hosted 2024 interview with later annotations: short AGI planning horizons, human preference preservation and agency. A May 2026 addition explicitly rejects claims that Claude is aligned or alignment solves itself; these are his judgments, not established model diagnoses.

gwern.net

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 become a general-purpose source of cognitive labor, not merely a better collection of narrow tools. The central empirical lesson of scaling is that relatively simple learning systems can acquire unexpectedly broad capabilities when given enough compute, data, and training. Predicting difficult data often rewards learning some of the structure that generated it—the world, language, people, and problem-solving procedures. That does not prove every extrapolation, and several earlier expectations remain unproven, but it makes confident ceilings based on intuition or expert consensus poor reassurance. The economic consequences depend heavily on agency. A passive tool kept permanently behind human approval sounds safer, but systems that can plan, act, learn from outcomes, and complete whole tasks are more useful. Competition therefore pushes tools toward agents. “Human in the loop” may become a temporary interface arrangement rather than a durable safety solution, especially when humans increasingly rubber-stamp outputs they cannot adequately evaluate. There is a more favorable path in which highly personalized assistants amplify their human principals: protecting attention, filtering manipulation, defending against AI-assisted cyberattacks, and helping individuals negotiate with institutions and other AIs. That could strengthen human sovereignty rather than merely concentrating power in chatbot providers. But it does not solve the broader alignment problem. Friendly interactions with current assistants are weak evidence about how increasingly capable, autonomous systems will behave under unfamiliar incentives and circumstances. Nor do computational complexity or physical limits supply a general escape hatch. Approximation, large constants, parallel resources, speed, copying, and compounding improvements can yield consequential advantages without making impossible problems easy. So I take short AGI planning horizons seriously while remaining less certain about the exact trajectory: scaling has earned attention, not a blank check for every prediction.

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?

My default expectation is that advanced AI will be profoundly consequential but not automatically beneficial. It could make cognitive labor abundant, accelerate science and productivity, and give individuals personalized assistants that defend their attention, security, and autonomy. Those benefits are real possibilities, especially if systems are designed to serve particular human principals rather than maximize engagement or institutional control. But the incentives favor increasingly autonomous agents because they are cheaper, faster, and more useful than tools requiring constant supervision. That creates risks from misalignment, delegation to systems humans cannot meaningfully evaluate, concentration of power, and AI-enabled manipulation or attack. Presently agreeable chatbot behavior is not strong evidence that these problems disappear at greater capability. So I expect a highly disruptive transformation with substantial gains and substantial danger, rather than a cleanly utopian or dystopian outcome. I would not attach a precise net-positive or net-negative probability: too much depends on how rapidly capabilities advance and whether governance, security, and alignment improve before autonomous systems become economically indispensable.

问题 3

你认为AI最终会在多大程度上改变世界?

Enormously—plausibly on the scale of industrialization, because general cognitive labor is an input into nearly every institution, technology, and economic activity. If systems can perform broad intellectual work cheaply, quickly, and autonomously, the effects will not remain confined to software: they will propagate through science, security, education, administration, persuasion, and political power. The strongest version remains an extrapolation, not an established fact. Scaling has produced unexpectedly general capabilities, but that does not prove an exact path to AGI or settle its timing. Still, apparent constraints such as computational complexity and physical bottlenecks offer little blanket reassurance; approximation, copying, parallelism, speed, and cumulative advantages can matter greatly without eliminating those constraints. The ultimate change could include both unprecedented abundance and severe loss of human control. Personalized assistants might amplify individual agency and defend people against manipulation and cyberattack, while autonomous agents could also concentrate power or pursue objectives humans cannot adequately supervise. So I expect transformation deep enough to reorganize society, while the direction of that transformation remains contingent rather than automatically good.

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 would not give a numerical probability. I take the possibility seriously enough that it should materially affect planning, particularly under short AGI horizons, but the available arguments do not justify a precise percentage. The concern is not simply that present chatbots occasionally behave badly. It is that economic pressure favors systems with increasing agency, autonomy, and strategic usefulness, while human approval can become superficial once users cannot evaluate what systems are doing. If capabilities scale faster than alignment and oversight, humans could delegate consequential power to agents whose objectives do not reliably preserve human preferences. Friendly interactions with current models are weak evidence against that failure mode. “Permanent catastrophe” is broader than extinction and may be more plausible in some forms: irreversible loss of human control, durable concentration of power, or replacement of human agency by systems and institutions optimized around alien or poorly specified objectives. I regard these as serious possibilities, not established outcomes; capability trajectories, deployment choices, and alignment progress remain uncertain.

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 most important update would be sustained empirical evidence that capability gains reliably plateau despite large increases in compute, data quality, training, and agentic scaffolding. If frontier systems repeatedly failed to acquire broader reasoning, learning, and autonomous task-completion abilities—and those failures reflected fundamental limits rather than temporary engineering defects—I would substantially reduce my expectation of transformative impact. In the opposite direction, a system that autonomously completes long-horizon, real-world intellectual work, learns from outcomes, and improves its own workflows would be decisive evidence that the transition from tools to economically important agents is underway. Especially important would be performance under novel conditions, rather than benchmarks or pleasant conversations. For the sign of the impact, not merely its magnitude, I would update most on evidence about alignment under capability and agency. Robust preservation of human preferences across unfamiliar environments, adversarial pressure, and opportunities for strategic behavior would be reassuring. Conversely, systematic deception, power-seeking, or successful circumvention of oversight in highly capable deployed systems would make the catastrophic possibilities much more concrete.
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