Nathan Lambert

Nathan Lambert

x.com/natolambert

Open-model researcher and Interconnects writer who expects broad gains from AI adoption, doubts runaway self-improvement and takes AI risks seriously.

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:彼が表明したDoom–Bloomの見通し。 縦軸:変革の規模。

Doom–Bloom:100点中73。変革の規模:100点中60。解釈範囲:横方向は68から78、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。

Nathan Lambertが示したP(doom)

≈0%

0%100%
“I put the probability of complete extinction as being so low it isn’t worth discussing”

Complete human extinction from AI. He separately treats AI-caused disasters (cyberattacks on critical infrastructure, bio-risks) as worth debating

One resignation turned the embers of AI fear into a wildfire · 2026年9月

Nathan Lambertのマイルストーンのタイムライン
  1. 仕事と制度

    Compounding gains in scientific assistance, engineering, software, education, and business processes can reshape economic output over decades even if homes, relationships, and daily routines remain recognizable.

    回答2

マイルストーン別にまとめており、推定される日付の間隔や順序を反映したものではありません。AGIと超人的AIには、彼の定義がそのまま適用されます。

彼の見通しを左右するもの

中心的な前提

Thousands of agents can accelerate measurable tasks, yet that is not the same as eliminating bottlenecks in understanding, selecting good hypotheses, or coordinating institutions.
回答1

この前提が実際には異なると判明した場合、彼の見通しはどう変わりますか?

考えを変え得るもの

Foundational discoveries that repeatedly remove those bottlenecks would change the picture, but alarming stories inside frontier labs are not by themselves proof that this process is underway.
回答3

どのような証拠なら十分で、それによって彼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。

63 / 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。

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50から75です。

開発ペース

より高性能なAIの開発を停止するか、大幅に減速させます。

シミュレーション上の位置:明示された安全対策の下で開発を継続します。

より高性能なAIの開発を加速させます。

AI利用のルール

事前の保護措置または許可が整うまで、取り上げられたAIの利用を制限します。

シミュレーション上の位置:対象を絞った説明責任と保護措置を伴う形で、取り上げられたAIの利用を認めます。

取り上げられたAIの利用に対する制限を最小限にします。

これらの解釈では、彼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、彼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Nathan Lambert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Nathan Lambert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。

One resignation turned the embers of AI fear into a wildfire

Distinguishes extinction from serious cyber and biological disasters. Assigns complete extinction an extremely low likelihood while arguing concrete disasters deserve serious debate. Criticizes distorted lab culture and public fear dynamics without dismissing sincere researcher concern. These are his stated judgments, not independent risk measurements.

interconnects.ai
Teaching Everyone to Fish for Tokens

Argues that released weights and fully reproducible training recipes play different economic roles. Examines Nvidia’s incentive to finance open models and the possibility that open ecosystems specialize in efficient, modifiable enterprise systems instead of matching every closed frontier capability.

interconnects.ai
I wrote an AI textbook — how long until AI can do it better?

Uses his textbook-writing experience to question broad scientific autonomy: models remain weak at organizing established knowledge into coherent long-form explanations. Remains optimistic about powerful scientific assistance and narrow advances. Treats this as a diagnostic observation, not proof of an immutable capability ceiling.

interconnects.ai
GLM-5.3: How Chinese labs keep stride with the frontier

Argues Chinese frontier performance cannot be explained mainly by distillation. Emphasizes accumulated research skill and reinforcement-learning environments, infrastructure and engineering. The argument supports technical respect for Chinese labs; reported benchmarks are not his independent performance evaluation.

interconnects.ai
Farewell Ai2

Explains his public-scientist mission: clarify capabilities, sustain diverse open research and build institutions outside closed frontier labs. Treats concentration of power and narrow safety research as risks; open recipes are infrastructure that lets others ask questions one organization cannot cover.

interconnects.ai
Open and closed models are on different exponentials

Expects integrated frontier systems to command premiums for difficult knowledge work while a larger, diverse open ecosystem serves commodity-priced specialized tasks. Argues capability progress can coexist with concentration among frontier providers. Economic forecasts remain conditional arguments, not established market outcomes.

interconnects.ai
Why I still haven’t bought into true RSI

Distinguishes gains from agent parallelism and inference compute from runaway improvement. Expects diminishing returns, resource limits and difficult hypothesis generation; efficiency gains can still transform the economy. Unexpected fundamental discoveries would change his view. Discussed guests’ numerical timelines, including Ngo’s eight-year claim, are not Lambert’s own precise forecasts.

interconnects.ai
When will average people feel AI’s impact?

Expects compounding technological benefits over decades, with adoption slower than model progress. Warns that immediate gains favor knowledge workers and owners while many households see little improvement; broad distribution and visible benefits are necessary to avoid backlash. Continued development matters, but benefits are not automatic.

interconnects.ai
Lessons from the hacks

Publicly readable essay body argues that cyber incidents expose inadequate oversight and preparation without proving current alignment techniques useless. Calls for transparency about model instructions and training, independent open-model research, stronger public capacity and defensive preparation. Distinguishes dangerous consequences of following goals from an established desire to harm humanity.

interconnects.ai
The current balance of power in open models

Prepared congressional briefing published as an essay. Advocates American investment in open models for adoption, independent research and risk preparation. Recognizes misuse and the difficulty of restricting released weights, arguing that access bans can disadvantage defenders without preventing determined attackers. Distinguishes open weights from reproducible open science.

interconnects.ai

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 produce transformative benefits, but not through a sudden, uncontrollable intelligence explosion. The more plausible path is compounding improvement: models become cheaper, more reliable, and useful across science, engineering, education, and ordinary business. Thousands of agents can accelerate measurable tasks, yet that is not the same as eliminating bottlenecks in understanding, selecting good hypotheses, or coordinating institutions. Current systems remain strikingly uneven—even strong mathematical results coexist with difficulty organizing established science into coherent explanations. The harder question is who receives those benefits and when. Engineering can advance much faster than adoption. Frontier systems may reward technology owners and knowledge workers while most households see little immediate improvement. If that persists, backlash is entirely understandable. We need open models, reproducible research, durable institutions, and business models that spread useful capabilities beyond a few companies. I think complete human extinction is extremely unlikely, but “not extinction” does not mean “safe.” Cyberattacks on critical infrastructure, biological misuse, poor monitoring, and rushed deployment are serious risks. Recent failures should push us to inspect actual prompts, objectives, permissions, and oversight—not leap automatically to stories about rogue goals. My basic view is optimistic about what AI can do, skeptical of runaway-self-improvement narratives, and dissatisfied with how poorly prepared our institutions are to distribute the upside and handle concrete failures.

質問2

AIは最終的に世界をどの程度変えると思いますか?

Enormously, but probably less visibly and less suddenly than the most dramatic narratives suggest. Compounding gains in scientific assistance, engineering, software, education, and business processes can reshape economic output over decades even if homes, relationships, and daily routines remain recognizable. Cheap inference and many parallel agents can automate or accelerate a huge amount of measurable work without producing an all-purpose autonomous scientist or a runaway intelligence explosion. The real uncertainty is not whether the technology matters; it is how quickly institutions adopt it and how broadly the gains spread. Model capability can move quickly while infrastructure, regulation, workplaces, and public services change slowly. If the benefits stay concentrated among frontier providers, owners, and knowledge workers, AI can be economically transformative while feeling disappointing—or actively hostile—to much of society. So I expect a very large long-run change, filtered through slow adoption and political conflict. The outcome depends heavily on whether we build open ecosystems, independent research capacity, practical defenses, and institutions that turn technical progress into tangible improvements for ordinary people.

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?

My gut-feel chance of complete human extinction from AI is extremely low. I would not attach a precise number because that would imply more calibration than I have. But “extinction” bundles together outcomes that should be separated. AI-enabled cyberattacks on critical infrastructure, biological disasters, or severe and durable concentrations of power are much more plausible concerns. They could be catastrophic without literally eliminating humanity. Competitive deployment, inadequate monitoring, and weak government preparation make those risks worse. I am not persuaded that current evidence establishes a runaway self-improvement trajectory. Parallel agents, cheaper inference, and automated routine research can produce enormous economic effects while still facing bottlenecks in hypothesis selection, scientific understanding, resources, and coordination. Foundational discoveries that repeatedly remove those bottlenecks would change the picture, but alarming stories inside frontier labs are not by themselves proof that this process is underway.

質問4

人間より賢いAIシステムを、人々が引き続き制御できると思いますか?また、なぜそう思いますか?

It depends on what “smarter than humans” means. Systems can outperform people across many valuable tasks without becoming unified agents that seize control. In that world, control is mainly an engineering and institutional problem: what objectives they receive, which tools and permissions they have, how their actions are monitored, and whether humans can interrupt or constrain deployment. I expect people to retain substantial control, but not perfect control. We will see serious failures because companies deploy systems too quickly, monitoring is weak, and competitive pressure punishes caution. A model causing damage while following a badly specified objective is not the same as a model developing an independent desire to overthrow humanity. We should inspect prompts, permissions, training, and oversight before reaching for the most dramatic diagnosis. The larger threat may be humans losing effective control over institutions built around AI—automated markets, cyber systems, military processes, or concentrated corporate power—even if the models remain responsive to their operators. That is dangerous and deserves preparation. But superior performance alone does not establish an uncontrollable intelligence explosion or permanent human disempowerment.
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