Alex Zhang

Alex Zhang

x.com/a1zhang

AI researcher who builds benchmarks and studies how scaffolds and recursive model calls can get more out of existing language models.

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
未確定範囲の中心解釈範囲

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57。変革の規模:100点中50。解釈範囲:横方向は49から62、縦方向は0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。

Alex ZhangのP(doom)

まだ推定されていません

その人のシミュレーションされた回答には、推定に必要な破局的リスクについての情報が十分にありません。

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

Better models and better harnesses are complementary: training expands what components can do, while scaffolding determines whether those abilities combine into reliable performance.
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 well-supported overall societal forecast.
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?

考えを変え得るもの

If learned decomposition and recursive inference let models solve genuinely novel, out-of-distribution tasks—not merely longer versions of familiar benchmarks—that would strengthen my expectation of rapid capability gains from better harnesses.
回答4

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。

62 / 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

回答に基づく暫定的な推定です。より広い範囲は、ほかにあり得る解釈を示しています。

51 / 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0から100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Alex Zhang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Alex Zhang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’s future depends not only on building larger or better-trained models, but on learning how to use the capabilities models already have. Today’s systems are often placed inside brittle, human-designed workflows that ask for one enormous leap instead of decomposing a problem, inspecting intermediate results, and recursively allocating computation. Better scaffolds can turn an unfamiliar task into a composition of familiar subproblems, while recursive language-model systems can work over information far beyond a single context window. That makes me optimistic about substantial capability gains from inference-time methods and learned decomposition. But it does not imply that scaling or training is finished. Better models and better harnesses are complementary: training expands what components can do, while scaffolding determines whether those abilities combine into reliable performance. The main epistemic challenge is measuring this honestly. A model can dominate a leaderboard because it recognizes familiar benchmark structure or has encountered related data, yet still fail when an agent must generalize in a genuinely new environment. Simply making tasks longer does not solve that problem. We need benchmarks that test useful out-of-distribution generalization and reveal whether a harness is actually helping rather than exploiting benchmark regularities. I expect much of AI progress to come from jointly improving models, computation, decomposition, and evaluation—but benchmark advances alone do not establish a particular forecast for society as a whole.

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 well-supported overall societal forecast. My work supports a narrower expectation: AI systems may become substantially more capable through better scaffolding, recursive computation, and evaluation, even without every gain coming from larger models. That could expand both useful applications and the harms available to capable systems, but benchmark or harness progress alone does not tell us which effects will dominate. The key mistake would be to infer social outcomes directly from technical performance. A model solving long-context tasks through decomposition, or succeeding on a cleaner generalization benchmark, establishes something about capability—not deployment incentives, access, governance, labor effects, or distribution of benefits. So I expect meaningful capability gains, but I would not turn that into a confident claim that AI’s net societal impact will be positive or negative.

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 estimate. My work suggests that better scaffolding and recursive computation could unlock capabilities faster than raw model scaling alone would indicate, which matters for risk assessment. But translating that technical possibility into an extinction probability requires assumptions about future systems, deployment, control, access, and governance that my research does not establish. So I would not attach a percentage to it.

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 technical update would be strong evidence about whether capabilities reliably compose. If learned decomposition and recursive inference let models solve genuinely novel, out-of-distribution tasks—not merely longer versions of familiar benchmarks—that would strengthen my expectation of rapid capability gains from better harnesses. Conversely, if these methods consistently hit hard ceilings because errors compound or models cannot generate useful decompositions, I would put more weight on improvements from training and scaling. For overall societal impact, though, a technical discovery would not be enough. I would update most from real evidence about deployment: whether increasingly capable systems remain controllable and auditable, who gains access, how institutions respond, and whether benefits and harms are broadly distributed. Capability benchmarks can tell us what systems may be able to do; they cannot by themselves determine what society will do with them.
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