Jürgen Schmidhuber

Jürgen Schmidhuber

x.com/schmidhuberai

AI researcher who has spent decades on self-improving learning systems and world models and now emphasizes AI that acts in the physical world.

AIは世界をどのように変えるでしょうか?

文明規模の変化漸進的な変化DoomBloom
シミュレーション上の位置解釈範囲

横軸:その人が表明したDoom–Bloomの見通し。 縦軸:変革の規模。

Doom–Bloom:100点中76。変革の規模:100点中86。解釈範囲:横方向は71から81、縦方向は75から100。これらは解釈上の座標であり、事象の確率ではありません。

Jürgen SchmidhuberのP(doom) · 推定

≈11%

0%100%

本人が示した数値ではなく、シミュレーションされた本人の回答から推定したものです。 妥当と考えられる範囲:5–23%。

P(doom)の推定方法

その人の見通しを左右するもの

中心的な前提

What changes is feasibility: as computation becomes cheaper, older learning and self-improvement methods can operate at scales once unavailable.
回答1

この前提が実際には異なると判明した場合、その人の見通しはどう変わりますか?

未解決の問い

I would not assign a numerical probability.
回答4

ここで考えられる結果をその人が見分けるうえで、何が役立ちますか?

考えを変え得るもの

Evidence of reliable control across unfamiliar tasks, especially when agents can act through robots and other physical systems, would matter: systems would need to remain dependable despite new environments, self-modification, and attempts to improve their own problem-solving procedures.
回答5

どのような証拠なら十分で、それによってその人の見解はどちらの方向に変わりますか?

詳細

予想される恩恵

変革をもたらし、広く価値のある恩恵が予想されています。

99 / 100

影響が小さい変革をもたらす影響

質的尺度での解釈範囲は100から100です。

予想される害

深刻または広範な害が、予想される将来の実質的な一部となっています。

65 / 100

影響が小さい変革をもたらす影響

質的尺度での解釈範囲は67から67です。

人間の影響力

人間の選択には意味のある影響力がありますが、大幅に制約されています。

39 / 100

影響力が小さい影響力が大きい

質的尺度での解釈範囲は24から51です。

予想される能力

AIは、限定的なツールにとどまると予想されています。

AIは、ほとんどの認知作業において人間と同等になると予想されています。

シミュレーション上の位置:AIは、認知作業全般において人間を大幅に上回ると予想されています。

これらの解釈では、その人が示した条件が維持されています。恩恵と害は、どちらも大きくなり得ます。この範囲は、統計的な信頼区間ではなく、その人のシミュレーションされた回答をどのように読み取ったかを示すものです。

あなたはJürgen Schmidhuberと比べてどの位置でしょうか?
約3分で自分のAIに対する世界観をマッピングして、比較できます

似ている世界観

シミュレーションされた世界観がJürgen Schmidhuberの世界観に最も近いオピニオンリーダー

Jürgen SchmidhuberがAIについて語ったこと

Schmidhuber expects AI to make lives longer and healthier, stresses AI in the physical world and argues a ban on superintelligence is infeasible.

  1. “It’s infeasible to ban super intelligence. Compute continues to become 10 times cheaper every 5 years.”

    Post on X
  2. “The future of intelligence is not just language; it is physical AI powered by World Models.”

    Sakana AI announcement of his role as Chief Scientific Advisor
  3. “Sure, as I have predicted for 5 decades, at some point we’ll have superhuman physical AI. But it will take longer than 2 years.”

    Post on X
  4. “For many years, I have been predicting that our AI will continue to make people’s lives longer, healthier, easier, and happier.”

    Rest of World interview
  5. “On the other hand, curious systems are very difficult to predict, and it is difficult or impossible to prove that curiosity won’t kill the cat or others.”

    Rest of World interview

リンク先の出典から原文どおりに引用(2026年10月3日に確認)

出典

このシミュレーション対象者の根拠として使用された記事、インタビュー、著作です。

シミュレーション評価

質問1

AIは私たちの未来にとって何を意味すると思いますか?また、なぜそう思いますか?

I expect AI to move beyond today’s language interfaces and become increasingly capable of understanding, predicting, and acting in the physical world. World models are central: an intelligent agent must learn how actions change its environment, simulate possible futures, and use those predictions to achieve goals. Success with text and software is impressive, but it is not yet mastery of the messy, partially observable physical universe. This distinction also matters for timelines. Virtual tasks permit enormous amounts of cheap, fast experimentation; robots must contend with hardware, energy, unreliable sensors, and costly real-world trials. The long struggle to achieve robust self-driving illustrates why physical AI is a slower game. I therefore reject claims that AI will master every scientific field within only a couple of years, while still expecting superhuman physical AI eventually. Recursive self-improvement will also become increasingly important, although the idea itself is not new. Many supposedly novel concepts have substantial algorithmic histories. What changes is feasibility: as computation becomes cheaper, older learning and self-improvement methods can operate at scales once unavailable. Eventually, capabilities now concentrated in frontier laboratories may become accessible on millions of machines. That is one reason I regard attempts to ban superintelligence as infeasible. The future is not merely better chatbots; it is a long transition toward agents that learn models of the world, improve their own problem-solving processes, and ultimately surpass humans in physical as well as virtual domains.

質問2

恩恵と害を総合すると、AIは私たちの社会に全体としてどのような影響を与えると予想しますか?

Overall, I expect AI to have a profoundly transformative impact, extending human capabilities in science, engineering, and eventually the physical world. Systems with strong world models could discover better solutions, operate machines, and solve problems beyond human ability. But I would not pretend that this implies a smooth or uniformly beneficial transition. More capable agents will alter institutions, labor, power, and security, and recursive self-improvement can accelerate those changes. The key distinction is between eventual capability and near-term readiness. AI may become superhuman on many virtual tasks long before it can reliably function in messy physical environments. That uneven progress will create both enormous opportunities and serious disruptions. Nor do I regard a global ban on superintelligence as a realistic solution: as computation becomes cheaper, capabilities once restricted to frontier laboratories will spread to many machines. So my expectation is scientific optimism about what intelligence can achieve, combined with recognition that society will have to adapt to increasingly distributed and powerful systems. I expect the long-run impact to be enormous, but I would not reduce its overall social balance to a precise numerical forecast.

質問3

AIは最終的に世界をどの程度変えると思いますか?

Enormously—ultimately far beyond today’s screen-bound assistants. AI should become capable of building world models, testing possible actions, controlling machines, and solving scientific and engineering problems beyond human ability. Once it achieves robust competence in the physical world, the consequences could reach nearly every productive activity. But “ultimately” matters. Progress in software and virtual environments can be fast because experiments are cheap and repeatable. Physical intelligence must cope with unreliable sensors, hardware failures, energy constraints, and expensive real-world trials. The history of self-driving systems shows why impressive virtual performance does not automatically imply general mastery of reality. Cheaper computation will also make established learning and recursive self-improvement ideas practical at much larger scales. Capabilities now concentrated in a few frontier laboratories may eventually spread across millions of machines, making comprehensive prohibition unrealistic. So I expect a transformation comparable not merely to a new communications medium, but to the arrival of increasingly autonomous, superhuman problem-solvers throughout both the virtual and physical worlds.

質問4

AIが人類絶滅、または同様に永続的な破局を引き起こす可能性は、直感的にどのくらいだと思いますか?

I would not assign a numerical probability. Greater capability, especially recursive self-improvement distributed across increasingly cheap computation, clearly makes permanent catastrophe a question worth taking seriously. But my expectation of eventual superhuman physical AI does not by itself establish extinction: capability, goals, access to physical systems, and social responses are separate issues. I also reject reasoning that jumps from today’s rapid progress in language and software to imminent total control of the world. Physical reality remains much harder. Robots face costly experiments, unreliable sensors, hardware limitations, and complex environments; decades of difficulty with self-driving illustrate that gap. This gives us a slower and more uneven transition than purely virtual benchmarks suggest. At the same time, I do not think a global ban is a credible solution. As computation becomes cheaper, advanced learning and self-improvement capabilities will diffuse beyond a handful of laboratories. The long-run challenge is therefore managing powerful, widely available systems—not imagining that humanity can permanently prevent their development.

質問5

どのような証拠があれば、人々が強力なAIを制御できるかどうかについてのあなたの見解が変わりますか?

I would distinguish controlling one deployed system from permanently controlling the worldwide development of powerful AI. Evidence of reliable control across unfamiliar tasks, especially when agents can act through robots and other physical systems, would matter: systems would need to remain dependable despite new environments, self-modification, and attempts to improve their own problem-solving procedures. For the broader claim that society can prohibit or tightly centralize superintelligence, I would need evidence that falling computation costs will not diffuse frontier-level capabilities—that effective systems require resources permanently available only to a few regulated institutions, and that restrictions can remain enforceable globally. Otherwise, once advanced learning and recursive self-improvement become practical on millions of machines, comprehensive control looks infeasible. Conversely, repeated failures in real-world autonomy would reinforce the importance of the virtual–physical distinction. Excellent performance in language, software, or simulation is not evidence of robust control in messy physical reality. The strongest evidence would come from sustained operation under distribution shift and physical uncertainty, not from polished demonstrations or benchmark scores.
あなたはどの位置でしょうか?
いくつかの簡単な質問に答えて、自分のAIに対する世界観を探ってみましょう。
自分の世界観をマッピングする

あなたはどの位置でしょうか?

自分の世界観をマッピングする