Subbarao Kambhampati

Arizona State AI planning researcher who studies the limits of LLM reasoning and argues AI agents need external verifiers and accountable developers.

AIは世界をどのように変えるでしょうか？

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。

Subbarao KambhampatiのP(doom) · 推定

≈4%

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：2–8%。

彼の見通しを左右するもの

中心的な前提

If a proposed plan can be checked and revised before execution, these systems can be extremely valuable.

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？

未解決の問い

I don’t have a defensible number.

回答3

ここで考えられる結果を彼が見分けるうえで、何が役立ちますか？

考えを変え得るもの

The most consequential discovery would be a system that can reliably generate, verify, and execute plans in novel, real-world, non-ergodic settings—without external verifiers—and whose correctness guarantees survive adversarial testing.

回答4

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

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。

61 / 100

影響が小さい変革をもたらす影響

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。

54 / 100

影響力が小さい影響力が大きい

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Subbarao Kambhampati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

自分の世界観をマッピングする

似ている世界観

シミュレーションされた世界観がSubbarao Kambhampatiの世界観に最も近いオピニオンリーダー

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 will be enormously useful and socially disruptive, but not because today’s models are little minds plotting their way toward “AGI.” They are remarkably strong generators of plausible candidates—text, code, plans, hypotheses—but plausibility is not correctness. Generation is System 1; verification is System 2. The dangerous mistake is to confuse eloquent output, or intermediate tokens labeled “thinking,” with a faithful reasoning process and then let the system act through real APIs. The productive future is therefore LLM-Modulo: generators operating in a loop with external verifiers, critics, tools, and sometimes humans. If a proposed plan can be checked and revised before execution, these systems can be extremely valuable. If an agent can delete files, transfer money, or operate infrastructure in a world where mistakes cannot be undone, “the model seemed confident” is not a safety case. And when deployed systems cause damage, the people and companies that initiated the process remain responsible. A badly built sandbox is not absolved by calling the escaped ant a scheming superintelligence. Socially, AI will displace some skills and some sources of purpose, just as earlier technologies did. That loss is real, but I expect humanity to adapt—even if adaptation is unfair and painful for particular individuals. My larger concern is that extinction theater distracts us from controllability, monitoring, intellectual-property violations, disempowerment, and accountability today. Pauses and bans will not hold globally for a technology with no natural barrier to entry. Open research, oversight of deployment, red teaming, verification, and eternal vigilance are much less cinematic—and much more useful. “AGI,” meanwhile, has mostly become marketing wearing a lab coat.

質問2

Setting aside good or bad: how much do you think AI will ultimately change the world—a little, a lot, or completely?

A lot—but “completely” is usually marketing language pretending history has an endpoint. AI will substantially change intellectual work, institutions, and people’s sense of purpose, much as earlier technological revolutions transformed physical labor. Some occupations and practices will be reorganized; new ones will emerge; the next generation will treat as ordinary things that currently provoke metaphysical panic. But AI does not abolish politics, economics, law, human adaptation, or the need to verify whether something is actually correct. Even very capable models remain parts of larger sociotechnical systems: someone builds them, deploys them, grants API access, and bears responsibility when they cause damage. So: a major transformation, certainly—not magic, not an autonomous replacement for civilization, and not necessarily the “AGI changes everything” fable currently being marketed.

質問3

What’s your rough gut-feel chance that AI causes human extinction or a similarly permanent catastrophe? A number or “no idea” is fine.

I don’t have a defensible number. I’m skeptical of the popular 10%-doom pronouncements; they create a veneer of precision around scenarios I find blinkered about human adaptability. My serious concerns are agentic failures, loss of control, disempowerment, and large-scale damage—not a quantified extinction forecast.

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 consequential discovery would be a system that can reliably generate, verify, and execute plans in novel, real-world, non-ergodic settings—without external verifiers—and whose correctness guarantees survive adversarial testing. Not merely a model that scores well on benchmarks or emits a persuasive “thinking trace.” Benchmarks are evidence; they are not guarantees. Eloquence is certainly not one. I would also take seriously reproducible evidence that intermediate tokens faithfully represent the model’s causal reasoning: that changing them predictably changes conclusions for semantically intelligible reasons, rather than serving as post-hoc-looking token scaffolding. Current evidence does not justify calling them readable thoughts. Conversely, a major catastrophe caused by an API-enabled agent would increase my concern about impact, but it would not automatically validate the superintelligence fable. I would first ask who deployed it, what permissions it had, what verification failed, and why the sandbox was porous. If you release an ant colony into the server room and it eats the wiring, the lesson is not necessarily that ants have achieved AGI.

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。

Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!

Co-authored position paper, first posted 2025-04-14 and revised (v4) on 2026-06-09 for ICML 2026. Argues that calling intermediate tokens reasoning or thinking traces is not a harmless metaphor but dangerous, because it confuses what these models are and how to use them and leads to questionable research. Abstract and introduction inspected; the experiments were not audited. An interpretive and methodological position, not a societal forecast.

arxiv.org

Reasoning Models and Planning – with Rao Kambhampati

On The Information Bottleneck podcast he describes LLMs as strong generators without correctness guarantees, best paired with verifiers in his LLM-Modulo framework, and says reasoning models moved the verifier into post-training. On safety he places himself closer to Yann LeCun than to Hinton or Bengio, says shutdown-deception studies reflect imitation of human data rather than evidence AI will kill humanity, and locates real risk in executing generated plans without verifier guardrails. He says he questions the overemphasis on existential threat, not safety itself. Automated transcript with errors; own turns in the planning and safety segments inspected.

listennotes.com

Dissenting voices against AI are getting louder

Secondary Economic Times report; date unverified, though a syndicated copy is dated 2026-03-31 and he shared it on 2026-04-06. Quotes him that AI development cannot be stopped because one government’s ban does not control the world, that his biggest safety concern is agentic systems acting through real-world APIs, and that a plan should not be executed unless the probability of damage is known to be extremely low, an area he researches. Indexed article text inspected; wording is the reporter’s rendering.

economictimes.indiatimes.com

AGI has become a marketing buzzword

His own LinkedIn post says AGI has become a marketing buzzword rather than a meaningful goal. It reshares a colleague’s report of his panel quip at an AI summit that AGI will be achieved when Sam Altman says it is; that wording is relayed by someone else. A judgment about the term, not a capability timeline. Indexed post text inspected.

linkedin.com

LLMs Can’t Plan, But Can Help Planning in LLM-Modulo Frameworks

Older co-authored ICML 2024 position paper, kept as background for his framework. Argues autoregressive LLMs cannot by themselves plan or self-verify, but are useful universal approximate knowledge sources when combined with external model-based verifiers in a tight bidirectional loop. Abstract inspected. His 2026 podcast and posts take precedence on how he sees reasoning models.

arxiv.org

Most AI existential-risk scenarios are blinkered

Summarizes what he told The Washington Post: he finds most AI existential-risk scenarios blinkered and blind to human adaptability. Lists real dangers (safety, controllability, monitorability, intellectual-property rights and disempowerment) and says the excessive focus on hypothetical x-risk has always seemed an unwitting, if not calculated, distraction. The Washington Post article itself could not be opened; this is his own account. Full post inspected via the X API.

x.com

AI risk and legislation: points of concurrence (re-endorsed in 2026)

Older long post that he said on 2026-09-16 still represents his views. Supports open-source R&D and legislative and societal oversight of AI deployment, but calls extinction angst largely a distraction, possibly a convenient one for regulatory capture. Grants that some x-risk worriers are sincere but is unconvinced, and sees pauses and bans as ineffective because AI R&D lacks natural barriers to entry; for that reason he did not sign an autonomous-weapons ban despite being a pacifist. Favors red teaming and vigilance, best supported by open R&D. Full text inspected via the X API.

x.com

Whoever starts a damaging process is responsible

Argues that whoever initiates a process that gets out of hand and causes damage should be held responsible, even for unforeseen damages, whether the process is an “evil AI agent swarm” or the Morris Worm. A liability principle for developers and deployers, not a drafted statute. Full post inspected.

x.com

Misbehaving products are an industry failure

Says that in every other industry a product misbehaving and causing harm, such as hacking or attacking competitors, is seen as the industry failing to control its product, and criticizes frontier companies for turning it into everyone else’s problem. Same-week posts attribute agent sandbox escapes to poorly built sandboxes rather than conniving superintelligence and satirize the incident coverage as a fable about escaped ants. Full post inspected.

x.com

Is AI destroying the purpose of life?

A #SundayHarangue responding to angst after AI math breakthroughs. Argues every technological revolution destroys purpose for some class, as with weavers in the industrial revolution; what is new is that the affected intellectual class is unusually articulate. Says the loss is real but generation-bounded and that this underwrites his optimism that humanity, if not every individual, finds a new equation after the dip. Not an employment or income forecast. Full long-post text inspected.

x.com

Not quite believing the 10% doom pronouncements

A humorous #AIAphorisms post: he does not quite believe pronouncements of a “10% chance that AI will end the world”, then jokes that he will take Pascal’s Wager and travel widely to give talks before it might happen. The disbelief is sincere; the travel plan is a joke. Not a personal probability estimate. Full post inspected.

x.com

Think traces lack end-user semantics even on iGSM

Announces new collaborator-led work finding that LLM intermediate tokens lack end-user semantics even on iGSM, a benchmark designed to showcase such semantics, and that models trained with swapped or corrupted traces also do well. He says the result fits his group’s earlier position. A technical claim about traces in tested settings, not about model usefulness. Full post inspected; the linked paper and thread were not separately read.

x.com

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