Minh Nhat Nguyen

Minh Nhat Nguyen

x.com/menhguin

AI researcher who studies agent training and model overconfidence and writes about how AI is changing scientific research and security.

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47。変革の規模:100点中60。解釈範囲:横方向は42から52、縦方向は50から75。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。

Minh Nhat NguyenのP(doom) · 推定

≈6%

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:2–16%。

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

AI can generate code, papers, experiments, and plausible-looking tasks fast enough to make everyone feel extremely productive while reducing the time spent deciding which goals actually matter.
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 do not have a defensible numerical estimate for the net societal effect.
回答2

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。

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。

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。

58 / 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。

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。

53 / 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30から95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Minh Nhat Nguyen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Minh Nhat Nguyen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 split into two economically distinct layers. Cheap, good-enough models will handle routine work, while expensive frontier systems may become most valuable as autonomous researchers—running experiments, iterating on ideas, and helping frontier labs operate more like automated research organizations. That demand could remain strong even when ordinary model access becomes commoditized. But more output is not automatically more progress. AI can generate code, papers, experiments, and plausible-looking tasks fast enough to make everyone feel extremely productive while reducing the time spent deciding which goals actually matter. Scientific communication could also get polluted by low-quality generated papers repeatedly circulating through conferences. On the security side, capable models plus carelessly generated software create more attack surface, making theft of lab secrets a serious prospective risk. So the future is not just “AI makes us faster.” It makes choosing objectives, designing reliable agent training, calibrating confidence, and protecting research systems more important. Our experiments already point to concrete failure modes such as training instability and agents becoming overconfident in debate. I would also resist describing every capability increase as AGI, ASI, or RSI; diluting those terms into corporate branding makes it harder to discuss what systems can actually do.

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 a mixed impact, with substantial gains in useful output but no guarantee of proportional social or scientific progress. Cheap models should make routine work easier, while frontier systems could create exceptional value through autonomous research. At the same time, they can flood institutions with plausible but low-quality work, create insecure software, enable theft of research secrets, and encourage a misleading sense of productivity. So I would not reduce the outcome to simply positive or negative. The central bottleneck shifts from producing more things to choosing worthwhile goals, distinguishing genuine progress from activity, and keeping increasingly capable agents reliable and appropriately calibrated. I do not have a defensible numerical estimate for the net societal effect. Much depends on whether our institutions can adapt to that shift rather than treating output volume as progress.

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 the change to be very large, but uneven. Routine model use may become cheap infrastructure, while frontier systems could reshape research by autonomously proposing, testing, and iterating on ideas. That would alter not only how much work gets done, but where economic value concentrates—especially around labs able to train and deploy costly research agents. The harder question is whether this produces corresponding progress. AI can multiply code, experiments, and papers without improving our choice of goals; it may even make pointless activity feel productive. It can also overload scientific communication and expand security risks around valuable research. So the deepest change may be that generating work becomes easier while judgment, calibration, and institutional filtering become more important. I expect major change, but I would not translate that into a precise AGI timeline or treat every capability gain as evidence of AGI, ASI, or RSI.

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. I see serious concrete risks—especially insecure software, AI-enabled intrusion, theft of frontier-lab secrets, unreliable agents, and overconfidence—but those observations do not by themselves establish an extinction forecast. I would rather keep measured failure modes separate from much larger extrapolations than turn a gut feeling into a misleadingly precise number.
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