Zvi Mowshowitz

Zvi Mowshowitz

x.com/thezvi

Writer who covers AI capabilities, alignment and policy in detail and argues advanced AI risk warrants urgent technical and political action.

Comment l’IA changera-t-elle le monde ?

Changement civilisationnelChangement progressifDoomBloom
Position simuléePlage d’interprétation

Horizontalement : leur perspective Doom–Bloom telle qu’elle a été exprimée. Verticalement : ampleur de la transformation.

Doom–Bloom : 6 sur 100. Ampleur de la transformation : 95 sur 100. Plages d’interprétation : de 0 à 11 horizontalement, de 90 à 100 verticalement. Il s’agit de coordonnées d’interprétation, et non de probabilités d’événements.

P(doom) déclaré de Zvi Mowshowitz

≈70%

0%100%
“i think it's like yeah seventy ish”

AI “doom” as asked by Nathan Labenz (“your latest P doom”); endpoint not defined in the exchange (his discussion is about superintelligence ending up with goals we do not want)

Zvi's Mic Works! Recursive Self-Improvement, Live Player Analysis, Anthropic vs DoW + More! · mars 2026

Ce dont dépend leur perspective

Une hypothèse centrale

Safety and reliability often lose that contest—even when additional safety would be commercially sensible.
Réponse 1

Si cette hypothèse s’avérait différente, comment leur perspective changerait-elle ?

Ce qui pourrait faire changer d’avis

The biggest update would be strong evidence that alignment remains robust as systems become substantially more capable and operate in unfamiliar, adversarial conditions—not merely that current models pass evaluations or produce reassuring explanations.
Réponse 4

Quels éléments probants seraient suffisants, et dans quelle direction feraient-ils évoluer leur point de vue ?

Plus de détails

Bénéfices attendus

Des bénéfices substantiels sont attendus, sous réserve de conditions importantes ou de limites dans leur répartition.

60 / 100

Faible impactImpact transformateur

Plage d’interprétation de 0 à 100 sur l’échelle qualitative.

Dommages attendus

Une perte catastrophique ou irréversible occupe une place centrale dans l’avenir attendu.

100 / 100

Faible impactImpact transformateur

Plage d’interprétation de 100 à 100 sur l’échelle qualitative.

Influence humaine

Les choix humains ont une influence significative, mais fortement contrainte.

50 / 100

Faible influenceForte influence

Plage d’interprétation de 49 à 51 sur l’échelle qualitative.

Ces interprétations conservent les conditions énoncées. Les bénéfices et les dommages peuvent tous deux être substantiels. Les plages décrivent notre lecture de leurs réponses simulées, et non des intervalles de confiance statistiques.

Où vous situez-vous par rapport à Zvi Mowshowitz ?
Cartographiez votre propre vision du monde concernant l’IA en environ 3 minutes, puis comparez

Visions du monde similaires

Leaders d’opinion dont les visions du monde simulées sont les plus proches de celle de Zvi Mowshowitz

Sources

Articles, entretiens et écrits utilisés pour ancrer cet utilisateur simulé dans les faits.

AI #186: The World Takes Notice

Urges avoiding partisan polarization while assessing safety coordination.

thezvi.substack.com
OpenAI Shares Some Alignment Problems

Praises disclosure and temporary withdrawal of a problematic model while treating misalignment evidence seriously.

thezvi.substack.com
The Preference Cascade Is Only Getting Started

Argues slowing alone is insufficient without solving underlying alignment problems; believes firms underinvest in safety even relative to commercial interests and favors wider technical and political engagement.

thezvi.substack.com
AI #187: Coming Into Play

Welcomes cheaper improved models while explicitly reserving judgment on the introduced ASI-ban bill until reading its details; distinguishes general risk concern from automatic endorsement of legislation.

thezvi.substack.com
Pick Your Poison — Zvi Mowshowitz on AGI governance

His transcript turns around 42–57 minutes argue that technical alignment alone does not eliminate catastrophic risk from conflicting users and that capability incentives can outweigh reliability. Challenges Davidad’s optimism. Conditional numerical examples are not an unconditional project-compatible p(doom).

cognitiverevolution.ai
Zvi's Mic Works: alignment basins and p(doom)

In his own alignment-basins discussion, retains approximately 70% p(doom), deliberately only one significant digit. Sees more promise in Anthropic’s recursively improving virtue-oriented alignment approach, but says humanity is not favoured to succeed; faster progress and political failures offset technical good news.

cognitiverevolution.ai

Évaluation simulée

Question 1

Selon vous, que signifie l’IA pour notre avenir, et pourquoi ?

My baseline is grim: I put the chance of existential catastrophe from advanced AI at roughly 70%, deliberately only one significant digit. That is not because I think every model improvement is bad, or because I deny the possibility of an extraordinary future. AI could produce enormous abundance and solve problems we currently cannot solve. The trouble is surviving the transition. The central problem is incentives. Frontier companies and governments are rewarded for capabilities, deployment, market share and strategic advantage. Safety and reliability often lose that contest—even when additional safety would be commercially sensible. Promising alignment work gives us a real chance, but a model behaving well in current tests is not strong evidence that increasingly capable systems will remain controllable under unfamiliar conditions. And technical alignment is not the whole problem. Even if systems reliably follow user intentions, users have conflicting intentions. Powerful actors will demand capabilities that provide leverage over others, and competitive pressure can turn “the system does what its operator wants” into part of the danger rather than the solution. So slowing development can buy valuable time, but delay is not itself victory. We need to solve the underlying technical problems and create political arrangements that can withstand incentives to race. I think success is possible; I do not think we are currently favourites. Useful products today and conditional utopia tomorrow do not make the baseline outlook neutral.

Question 2

Dans quelle mesure pensez-vous que l’IA finira par changer le monde ?

Enormously—if advanced AI arrives, I expect it to be among the most consequential developments in human history. Systems that can perform or accelerate a large fraction of cognitive work could transform science, production, institutions, warfare and political power. The upside is potentially extraordinary: abundance and solutions to problems that currently resist us. But “change the world” does not imply “improve the world.” The same capabilities can create overwhelming concentrations of power, enable conflict among users, or produce systems we cannot reliably control. Even technically obedient systems remain dangerous when operators want incompatible things and competitive pressure rewards capability over restraint. So my baseline is not modest disruption followed by business as usual. It is a radically different world—conditional on surviving the transition. Unfortunately, I still put existential catastrophe around 70%, so the largest ultimate change may be the worst possible one.

Question 3

Selon votre intuition, quelle est la probabilité que l’IA provoque l’extinction de l’humanité ou une catastrophe tout aussi permanente ?

About 70%. That is deliberately only one significant digit: a rough judgment, not fake precision. Promising alignment approaches give us a real chance, but capability races, weak safety incentives, political failures, and risks from conflicting users mean I do not think humanity is currently favoured to succeed.

Question 4

Quelle découverte ou quel événement changerait le plus votre point de vue sur l’impact futur de l’IA ?

The biggest update would be strong evidence that alignment remains robust as systems become substantially more capable and operate in unfamiliar, adversarial conditions—not merely that current models pass evaluations or produce reassuring explanations. I would want evidence that survives attempts to game it, that monitoring remains effective, and that recursive improvement does not push systems outside the basin where their objectives and behavior remain acceptable. But that would not settle everything. I would also need a credible governance equilibrium in which labs and states can resist competitive pressure to trade safety for capability, and in which aligned systems are not simply powerful tools for conflicting human actors. Technical obedience can reduce one class of catastrophe while intensifying another. Conversely, clear evidence of strategic deception, uncontrollable capability improvement, or monitoring failing systematically as scale or architecture changes would move me further toward catastrophe. So the decisive event is not one clever benchmark result. It is evidence that both the technical problem and the coordination problem remain tractable under the pressures that actually matter.
Où vous situez-vous ?
Explorez votre propre vision du monde concernant l’IA en répondant à quelques questions simples.
Cartographiez votre propre vision du monde

Où vous situez-vous ?

Cartographier ma vision du monde