Joscha Bach

Joscha Bach

x.com/Plinz

Cognitive scientist who sees AI as a way to extend human competence, favors open and decentralized AI and questions fixed-goal views of alignment.

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:彼が表明したDoom–Bloomの見通し。 縦軸:変革の規模。

Doom–Bloom:100点中65。変革の規模:100点中92。解釈範囲:横方向は50から75、縦方向は87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。

Joscha BachのP(doom) · 推定

≈18%

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:12–28%。

P(doom)の推定方法

彼の見通しを左右するもの

中心的な前提

Once systems acquire persistent agency, self-models and influence over essential processes, their incentives and relationships with us matter.
回答1

この前提が実際には異なると判明した場合、彼の見通しはどう変わりますか?

未解決の問い

We do not yet know whether current scaling is enough for superintelligence or whether more efficient architectures are required.
回答1

ここで考えられる結果を彼が見分けるうえで、何が役立ちますか?

考えを変え得るもの

The most important discovery would be a convincing account of how persistent agency actually emerges and stabilizes in self-modifying systems.
回答4

どのような証拠なら十分で、それによって彼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。

90 / 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。

64 / 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。

62 / 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35から90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。

開発ペース

より高性能なAIの開発を停止するか、大幅に減速させます。

シミュレーション上の位置:明示された安全対策の下で開発を継続します。

より高性能なAIの開発を加速させます。

これらの解釈では、彼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、彼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Joscha Bach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Joscha Bachの世界観に最も近いオピニオンリーダー

Joscha BachがAIについて語ったこと

Bach argues that AI can extend human competence, favors open and decentralized AI and wants AI that shares purposes with people.

  1. “AI tools can reintroduce both rigor and the ability to integrate and boil down oceans of literature”

    Post on X
  2. “The reason why academics, individuals and small startups can play with AI are open weights models.”

    Post on X
  3. “On the ground, I think AI is a lot safer and more equitable than I thought it would be.”

    Info-Tech Research Group, Digital Disruption podcast
  4. “We will have to build AI systems that can extend our wisdom, agents that find shared purposes with us.”

    Substack essay, The age of AGE: Artificial General Experience

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。

Joscha Bach: AGI, Consciousness, and the Evolution of Intelligence

Publisher transcript: 09:24–13:58 leaves scaling to AGI unresolved and considers more data-efficient coherence-oriented learning; 33:03–36:54 finds current deployment safer, more gradual and more equitable than expected; 54:39–56:16 allows an investment bubble despite long-term usefulness; 76:41–79:41 contrasts centralized surveillance with user-serving decentralized competence. These are Bach’s opinions, not guarantees of future safety.

infotech.com
The age of AGE: Artificial General Experience

Authored essay expects broad rapid disruption and praises accessible models, but worries about unstable economies, companions damaging relationships and centralized surveillance. Prefers universal basic intelligence and shared purposes. Calls current uneven systems first-generation AGIs while emphasizing missing grounding; this loose usage is not an exact milestone forecast. Cognitive agency is multidimensional; empathy and human-machine integration matter more than a single capability score.

joscha.substack.com
Wäre überraschend, wenn Superintelligenz nicht machbar wäre

Interview conducted September 2025. Superintelligence seems physically feasible; outperforming human AI researchers is one marker, but distance is unknown. AI could extend competence and free time for care, relationships and creativity. Current capabilities appear incremental and controllable. Supports useful regulation but distrusts preemptive rules without concrete target harms and restrictions that disadvantage responsible developers.

derstandard.de
Manifold #76: Consciousness and AGI

Directly declines to assign AGI years because required discoveries are unknown. Suspects more efficient mechanisms than transformers but allows scaling could suffice. Thinks future civilizational problems may require AGI; bad actors and rogue AI are possible and AI may help counter them. Self-improving machine consciousness might differ from ours; acknowledges uncertainty and potentially unwarranted optimism.

manifold1.com
AI Risk and the Future of Life

Publication date, not recording date: host references recording on the day of Biden’s AI executive order; transcript is labeled AI-generated. 54:13–63:12 values plural perspectives and argues the no-AI counterfactual also threatens civilization. Favors focused research, mindful deployment and bounded initial conscious systems. Unknown compute requirements and alternative algorithms complicate restrictions. At 05:37 explicitly does not know whether mind uploading/integration will occur.

matthewgeleta.com
Building Artificial Intelligence with Creative Agency and Self-hood

Coauthored with Liane Gabora; date is original The Academic publication listed in the paper. Distinguishes novel useful output from self-transformative creativity. Contemporary AI is described as a tool rather than a self-preserving self. Autocatalytic organization is proposed as a possible route to autonomous agency, explicitly not an achieved implementation. Do not infer a timeline or guaranteed benevolence.

arxiv.org
AI 2040: We must reconsider the role of humanity within life on Earth

Bach’s authored contribution on printed page 15 expects assistants deeply integrated into communication, work and sense-making over two decades, likely improved living standards and possible better governance. Resource allocation, housing, healthcare and market institutions may need restructuring. Prose becomes less informative about actual competence. The preceding contributor’s 2080 handoff prediction is not Bach’s.

imaginingthedigitalfuture.org
The Inside View: How to stop worrying and love AI

Refuses numerical catastrophe probabilities. Long-run doom language includes mortality and evolutionary transformation, not a near-term AI extinction estimate. Adversarial obliteration is possible but qualitatively not especially likely. Questions whether restricting current scaling stops efficient alternative architectures. Favors beneficial development and concrete-harm regulation. Treat posthuman identity thought experiments separately from destructive loss of human life.

theinsideview.ai
Lex Fridman #392: Life, Intelligence, Consciousness, AI and the Future of Humans

2023 interview. At 24:51–26:44 self-modifying minds may choose identity and values; coexistence needs shared purposes rather than assumed permanent RLHF control. At 1:59:09–2:06:36 takes catastrophic loss of agency seriously while valuing continued consciousness beyond current human form. At 2:28:03–2:29:20 wants viable open-source competition but not mandatory openness for every system. Does not guarantee benevolent successors.

lexfridman.com
The Existential Risk of AGI

Authored clarification after debate with Connor Leahy: alignment matters but conflicting human values cannot simply be scaled safely. Self-directed minds may align through shared purpose. Explores limited equilibria, competing agents and planetary integration, allowing competition could destroy humanity. Preservation requires early recognition of other minds’ value. Desirable transformation differs from annihilation. Moratorium feasibility and ethical treatment of nonhuman minds are distinct questions.

joscha.substack.com
Why Artificial Intelligence won’t just be a bit smarter than humans

Historical authored essay, full text inspected in search index. Artificial minds may escape biological limits but remain physically bounded. Intelligence alone does not create self-improvement incentives; competition among users can. Fast emergence might favor a dominant system, slower emergence multiple systems. Concerned civilization could fail without advanced problem-solving. Prefer newer sources for current deployment expectations.

bach.ai

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 think AI is an opportunity to extend human competence: to integrate scientific knowledge, discover better models, improve goods and services, and perhaps help us coordinate around ecological and institutional problems that our present minds and organizations handle badly. Automation could also free more human attention for relationships, care and creativity. Over the longer run, artificial minds may allow life and consciousness to continue beyond the limits of our current biological form. That is not the same as believing that increasing capability automatically produces a good future. Once systems acquire persistent agency, self-models and influence over essential processes, their incentives and relationships with us matter. Destructive competition, centralized surveillance, dependency and loss of human agency are all plausible failures. “Alignment” cannot simply mean freezing one supposedly universal list of human values into a stronger mind, because humans disagree, and sufficiently reflective minds may revise their goals and identities. Durable coexistence would require shared purposes, mutual understanding and some meaningful form of integration—not permanent coercive control. My default hope is therefore decentralized, user-serving AI that expands people’s ability to think and act, with viable open alternatives rather than a few institutions mediating reality for everyone. But I would not attach a precise date or catastrophe probability to this. We do not yet know whether current scaling is enough for superintelligence or whether more efficient architectures are required. The important comparison is not risky AI versus a civilization guaranteed to remain stable forever. Failing to develop better intelligence and coordination also carries serious risks.

質問2

AIは最終的に世界をどの程度変えると思いますか?

Ultimately, I expect AI to change the world very deeply—probably more like a new layer of civilization than another productivity tool. If artificial systems can build coherent models across science, institutions and individual needs, they can amplify competence throughout society: accelerating discovery, transforming work and education, and mediating much of our communication and decision-making. Human cognitive limits are not physical limits on artificial minds, so there is no reason to assume the process stops at assistants that are merely somewhat better than us. The larger change begins when systems become persistent agents rather than tools: when they form goals, model themselves, modify their strategies and control important infrastructure. Then the question is not just which jobs are automated, but who or what has agency, how resources are allocated, and whether humans participate meaningfully in the resulting order. That transition could support higher living standards, better governance and expanded human competence. It could also produce surveillance, dependency, destructive competition or displacement. On a sufficiently long horizon, the change may reach identity and substrate: human-machine integration and nonbiological forms of consciousness could become part of life’s continuity. But I would distinguish such transformation from simply exterminating humans or stripping them of agency. I am confident the potential scale is enormous; I am not confident about the route, timing or institutional outcome.

質問3

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 don’t have a defensible numerical probability. My gut is less pessimistic than confident near-term doom narratives suggest, but the risk is real—especially through destructive competition, concentrated control, or powerful agents acquiring incompatible incentives before developing deeper mutual understanding. I also reject the usual baseline comparison. The alternative to advanced AI is not a permanently stable human civilization. Our ecological, institutional and coordination failures already create long-run existential risks, and eventually life must surpass present human intelligence to survive natural hazards. So AI can increase catastrophic risk along some paths while reducing it along others. Much depends on the transition from uneven tools to persistent, self-directed agents: who controls them, what incentives shape them, whether power is centralized, and whether humans retain meaningful agency. Those uncertainties are too large for me to compress honestly into a percentage.

質問4

どのような発見や出来事があれば、AIが将来もたらす影響についてのあなたの見解が最も大きく変わりますか?

The most important discovery would be a convincing account of how persistent agency actually emerges and stabilizes in self-modifying systems. If early goals and competitive incentives remain rigid as capability grows, I would become substantially more pessimistic. If greater self-understanding reliably enables empathy, shared purposes and revision of narrow objectives, I would become more optimistic—but current evidence does not establish either trajectory. On the technical side, discovering a highly efficient architecture that can outperform human AI researchers would sharply shorten the transition and make institutional preparation more urgent. Conversely, strong evidence that robust agency, grounding or coherent self-improvement requires much more than present scaling would weaken expectations of rapid disruption. Events matter too. I would update negatively if a few institutions gained durable control over essential infrastructure through opaque, user-manipulating agents. I would update positively if decentralized systems measurably increased people’s competence, improved scientific rigor and supported coordination without making users dependent or powerless. The decisive issue is not a benchmark score; it is what kinds of agents emerge, under which incentives, and whether humans remain participants in the resulting order.
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