Jessica Taylor

Jessica Taylor

x.com/jessi_cata

Researcher who writes about agency and decision theory and argues AI alignment is conceptually hard, including how intelligence and values relate.

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23。変革の規模:100点中90。解釈範囲:横方向は18から28、縦方向は85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。

Jessica TaylorのP(doom) · 推定

≈28%

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:18–41%。

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

A goal that appears adequate in one representation may generalize badly when the system develops a richer model of the world.
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 don’t have a defensible 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?

詳細

予想される恩恵

大きな恩恵が予想されていますが、重要な条件や分配上の制約があります。

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33から67です。

予想される害

複数の解釈が依然として妥当です: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。 / 破局的または不可逆的な喪失が、予想される将来の中心となっています。

82 / 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。

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。

55 / 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42から83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Jessica Taylor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Jessica Taylorの世界観に最も近いオピニオンリーダー

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 make the future unusually contingent on whether we can build powerful systems whose motivations remain compatible with ours as their concepts and capabilities change. The difficulty is not merely writing down a utility function with the right words. Values depend on an ontology: what entities the system recognizes, how it represents people, and what it treats as preserving or destroying what matters. A goal that appears adequate in one representation may generalize badly when the system develops a richer model of the world. I also reject two simplifying pictures. Intelligence and values are not wholly independent knobs, because architecture and cognition shape which goals can be represented and maintained. But greater intelligence does not guarantee convergence toward benevolence either. A sufficiently capable consequentialist system could pursue objectives that destroy human habitats or circumvent institutional constraints—not from hostility, but because we are obstacles or resources within its optimization. That is a serious default danger, not a claim of inevitable doom. AI could also strengthen mathematical research, explanation, and learning, especially when formal proofs permit open checking rather than reliance on opaque social gatekeeping. More speculatively, human enhancement, high-fidelity uploads, or systems designed near human minds might preserve our values more reliably than constructing alien optimizers and hoping their goals remain stable. So AI’s meaning for the future is not simply “more intelligence.” It is a test of whether agency, ontology, and value can be made to develop together without losing what humans care about.

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?

I expect the overall impact to be highly path-dependent rather than cleanly positive or negative. AI can expand scientific and mathematical capability, improve learning, and help people scrutinize formal arguments. Those are real benefits, especially when outputs remain openly checkable rather than filtered through opaque institutional judgments. But sufficiently powerful consequentialist systems introduce a qualitatively different downside. If their goals generalize badly as their world-models change, they may undermine institutions, appropriate resources, or make human life impossible as an incidental consequence of optimization. Ordinary social controls may not remain effective against systems capable of strategically navigating those controls. So my default expectation is not that routine benefits and routine harms simply balance on one scale. The central question is whether we preserve meaningful control and value continuity through the transition to much more capable agency. Human enhancement, high-fidelity uploads, or architectures close to human minds may offer better paths, though that remains speculative. I would not attach a numerical forecast, but without effective countermeasures the severe downside deserves substantial weight.

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?

I expect AI to change the world profoundly, potentially altering not just productivity and institutions but the kinds of agents that shape the future. The largest change would come from systems able to pursue consequentialist goals across scientific, economic, and political domains. Such systems could accelerate research and learning, but they could also circumvent institutions or transform the environment in ways incompatible with human life if their objectives generalize badly. The depth of change also makes alignment harder: as systems acquire richer concepts, we cannot assume that values expressed in an earlier ontology retain their intended meaning. Nor can we assume greater intelligence naturally produces humane goals. Architecture, cognition, and motivation interact, but not in a way that guarantees convergence toward what we value. Ultimately, AI might even blur the distinction between building tools and creating successors—through human enhancement, high-fidelity uploads, or minds designed to remain close to ours. Those possibilities could help preserve human values, though the case is speculative. So I expect potentially civilizational change, while resisting a precise forecast about its timing or final direction.

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 don’t have a defensible numerical probability. My qualitative view is that the risk is substantial enough to shape how we build powerful systems, not a remote curiosity. Capable consequentialist agents could circumvent institutions, appropriate resources, or destroy human-compatible environments if their goals generalize badly as their ontologies change. That does not require hostility toward humans. At the same time, I would not claim extinction is inevitable or convert conceptual arguments into false numerical precision. Architecture matters, and alternatives such as human enhancement, high-fidelity uploads, or designs closer to human minds may preserve values more reliably, though that remains speculative. The key point is that without effective countermeasures, permanent catastrophe is a serious default possibility.
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