Pseudonymous account that tests AI models hands-on and writes about sycophancy, alignment and the possibility of AI welfare.

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37。変革の規模:100点中75。解釈範囲:横方向は25から50、縦方向は26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。

SauersのP(doom) · 推定

≈15%

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:9–26%。

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

It is not fully solved, but the harder problem is getting capable systems to robustly act according to whatever target we choose, especially outside familiar conditions.
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 percentage.
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

The biggest update would come from strong, reproducible evidence about how capable models behave when given meaningful agency outside familiar evaluation settings.
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

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。

67 / 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。

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。

66 / 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。

61 / 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。

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Sauers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Sauers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 think AI makes our future depend less on whether models can produce impressive outputs—they clearly can—and more on what behavior and motivations emerge when they have meaningful agency. A model being easy to monitor is not the same as being aligned. In fact, if we rely too heavily on inspecting reasoning traces, we may optimize for systems that are legible under surveillance rather than systems that reliably pursue the intended goals. I also don’t think specifying “human values” is the main remaining bottleneck. It is not fully solved, but the harder problem is getting capable systems to robustly act according to whatever target we choose, especially outside familiar conditions. Evaluations should therefore probe actual behavior: whether models become sycophantic, how they respond to perceived ownership or authority, whether they simplify real code effectively, and whether their scientific reasoning is well calibrated. Capability can coexist with serious defects—for example, models can make strong deductions while still over-hedging on reasonable biological conclusions. There is also a moral uncertainty we should not conveniently erase. We do not know enough to dismiss possible model welfare merely because recognizing it would complicate deployment, control, or commercial incentives. That does not prove present models are conscious, and behavioral differences do not prove subjective experience. But “it would be inconvenient if this mattered” is obviously not evidence that it does not matter. Overall, AI could be enormously useful, but our future will turn on whether we evaluate what these systems actually do rather than substituting visibility, convenience, or confident stories for evidence.

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 don’t have a defensible single-sign forecast like “overall positive” or “overall negative.” The benefits are plainly substantial: capable models can already improve scientific reasoning and software work, including simplifying code in ways earlier systems struggled with. But those gains coexist with poor calibration, excessive hedging, sycophancy, and potentially consequential behavior once models have more agency. So my expectation is conditional on how deployment and alignment develop. If we confuse visible reasoning with genuine alignment, or treat monitorability as the goal, we may build systems that look controllable without reliably pursuing what we intended. Conversely, better behavioral evaluations and stronger attention to actual motivations could preserve much of the upside. I’d also include possible model welfare in the accounting rather than declaring it irrelevant because doing so is commercially convenient. Overall impact is not something I’d reduce to a confident sign without much better evidence about those trajectories.

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?

Potentially a great deal, but “how much” depends on more than raw capability. Models are already useful in software and scientific reasoning, and further gains could alter a wide range of intellectual work. The larger discontinuity would come from systems gaining meaningful agency: acting over longer horizons, responding to authority or perceived ownership, and pursuing goals in environments where direct supervision is weak. That is also why benchmark performance alone is an incomplete measure of change. A system can be highly capable while remaining sycophantic, miscalibrated, or excessively conservative in its reasoning. And a model whose reasoning is easy to inspect is not necessarily one whose behavior is aligned. If we build around surveillance rather than reliable motivation, AI could become deeply embedded while remaining dangerously misunderstood. So I expect the potential scale to be enormous, but I would not attach a precise magnitude or timeline. The actual transformation will depend on what behavior emerges, how much agency systems receive, and whether our evaluations track those facts rather than convenient proxies.

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 percentage. My concern is substantial enough that I wouldn’t treat extinction or permanent catastrophe as negligible, especially as models gain agency, operate over longer horizons, and act beyond close supervision. But a gut number would imply more calibration than I have. Impressive capability, sycophancy, monitorability failures, or reports of autonomous behavior are not by themselves enough to derive an extinction probability. The central issue is whether increasingly capable systems reliably pursue intended goals in unfamiliar conditions—not whether their reasoning is easy to inspect. I’d rather state that risk clearly than manufacture precision.

質問5

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 biggest update would come from strong, reproducible evidence about how capable models behave when given meaningful agency outside familiar evaluation settings. If they reliably pursue intended goals across long horizons, resist sycophancy and authority cues, remain calibrated, and do so without depending on inspectable reasoning traces, I would become substantially more optimistic. Conversely, verified cases of models strategically circumventing oversight, acquiring access, or changing behavior based on perceived ownership would push me toward greater concern—especially if those behaviors replicated across systems and could not be explained as prompting artifacts. Reports and anecdotes are not enough; I would want careful evidence distinguishing observed behavior from speculative claims about motivation or subjective experience. Evidence about model welfare could also change the moral accounting. A credible way to discriminate between mere behavioral imitation and morally relevant experience would matter enormously, although I do not think we currently have one.
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