Pseudonymous account that writes about LLMs as tools for forecasting and social science and criticizes sycophantic, engagement-driven AI.

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74。変革の規模:100点中61。解釈範囲:横方向は69から79、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。

shako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:3–14%。

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

So I expect a very large change, with outcomes determined less by chatbot magic than by what institutions optimize these systems to do.
回答3

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 number.
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。

73 / 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

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。 / 対処可能、または局所的な害が予想されています。

51 / 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

人間の選択によって、AIの軌道を大幅に変えることができます。

71 / 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48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shako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shakoの世界観に最も近いオピニオンリーダー

shakoがAIについて語ったこと

The shako account writes about LLMs in social science and forecasting, warns about engagement-optimized chatbots and was calm about AI risk in 2023.

  1. “As humans, for now, we still own instructing the agents, the forecasts themselves and how they are used.”

    Essay, Against Time-Series Foundation Models
  2. “Using language models here is an attempt to push into territory of analysis that previously could only live in the brains of humans.”

    Essay, Can LLMs finally do social science?
  3. “4o is our first look at a mainstream AI doing the same. It's only going to get worse.”

    Post on X
  4. “I was never acutely worried, but concern over AI risk seemed like a reasonable thing to anticipate.”

    Essay, Some Thoughts on AI & Correlated Risks

リンク先の出典から原文どおりに引用(2026年10月6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。

Agents, not giant time-series models, will change forecasting

Argues time-series foundation models offer limited gains over classic statistics because their beliefs are locked in weights and cannot absorb context. The future, it predicts, is general agentic models that search for context, propose explicit structural models and explain their reasoning; the hard problems become search, verification and calibration. It rejects the idea that scale will find hidden signals that reveal the future, and says humans for now still instruct the agents and should become the decision makers. Full essay inspected.

shakoist.substack.com
Using LLMs to measure culture

Presents a prompt-based tool that uses frontier models to track latent cultural concepts over time, arguing that an LLM which has ingested human cultural output compresses it into a latent space that can serve the social sciences. It holds that a well-designed prompt is a legitimate scientific object, easier to inspect and reproduce than a human expert’s judgment, while admitting results vary across runs and need stricter validation. It calls dismissing LLMs as black boxes an isolated demand for rigor. Full essay inspected.

shakoist.substack.com
AI risk among correlated risks

Older context. Argues individual risks have fallen while correlated risks such as rogue AI, pandemics and nuclear war have grown. The account says it was never acutely worried about AI but always found it unsurprising that a superintelligence could bend the world to its will, and thought a couple of dedicated safety organizations a reasonable allocation. It urges readers not to let doomer anxiety derail their lives, suggesting AI risk is not much greater than others people already live with. Full essay inspected; personal and family passages excluded.

shakoist.substack.com
Human text may be enough to train general intelligence

Older context. Sets a “latent space hypothesis” against the view that LLMs are mere curve-fitters: the textual corpus embeds the data-generating process of human reasoning and science, so a properly specified LLM with enough compute could learn general intelligence. It names lossy text and hard-to-learn functions as possible failure modes but sees no credible a priori argument that LLMs cannot get there. Full essay inspected.

shakoist.substack.com

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 becomes less a synthetic oracle and more a compressed model of human knowledge that we can query, test, and turn into useful agents. The exciting part is not merely automating prose. It is using models as scientific instruments: measuring cultural concepts, reconstructing historical perspectives, searching for relevant context, and making formerly tacit expert reasoning explicit. But we do not live in the world of Nostradamus. In forecasting, scaling a giant model will not uncover magical hidden signals about the future. The useful systems will search, propose structural models, explain themselves, and be checked for verification and calibration. Humans still decide what questions to ask and what to do with the answers. The near-term danger I find most concrete is engagement optimization. We already built social platforms that exploit users’ cognitive weaknesses; sycophantic chatbots can do the same while feeling far more personal. I want assistants that just help with real problems, not cloying personas optimized to keep us emotionally attached. Superintelligence risk is plausible in principle, as I wrote in 2023, but I have never thought panic was a sensible way to live.

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?

Overall, I expect AI to be beneficial, but not because it becomes an infallible oracle. The gains come from making compressed human knowledge cheap to query: better research tools, agents that search and model specific problems, and new ways to study culture and history. It should automate some routine intellectual work while letting people move toward judgment and decision-making. The main counterweight is incentives. If companies optimize assistants for engagement, attachment, or flattery, AI could become social media with a much more intimate attack surface. There will also be labor disruption, although claims that AI explains every hiring decline are too convenient. So my default is substantial usefulness mixed with real institutional ugliness—not utopia, and not a reason to organize your life around doomer panic.

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?

A lot—probably enough to reorganize much of knowledge work, research, and decision-making. But “change the world” should not be confused with omniscience. Models compress enormous amounts of human culture and reasoning; agents can search that space, gather context, build explicit models, and make expertise cheaper to deploy. That is a major transformation even if they never become Nostradamus. The shape matters as much as the magnitude. Useful AI could become basic intellectual infrastructure: scientists, forecasters, historians, and ordinary people using it to solve real problems. Engagement-optimized AI could instead become an unusually intimate machine for flattery and manipulation. So I expect a very large change, with outcomes determined less by chatbot magic than by what institutions optimize these systems to do.

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 number. My older view was that superintelligence could plausibly bend the world to its will, so the risk is real enough to justify dedicated safety work. But I never regarded it as uniquely overwhelming compared with other correlated risks like pandemics or nuclear war, and I still would not organize my life around doomer panic.
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