Theia Vogel

Theia Vogel

x.com/voooooogel

AI researcher who runs experiments on language model introspection and personas and maintains an open-source library for steering models.

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 : 43 sur 100. Ampleur de la transformation : 72 sur 100. Plages d’interprétation : de 38 à 51 horizontalement, de 67 à 77 verticalement. Il s’agit de coordonnées d’interprétation, et non de probabilités d’événements.

P(doom) de Theia Vogel · inféré

≈9%

0%100%

Déduit de leurs réponses simulées, et non d’un chiffre donné par ces personnes. Plage plausible : 3–23%.

Comment nous estimons P(doom)

Ce dont dépend leur perspective

Une hypothèse centrale

Capability is not equivalent to power; resources, infrastructure, and comparative advantage matter.
Réponse 1

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

Une question non résolue

I would not put a defensible percentage on it.
Réponse 4

Qu’est-ce qui les aiderait à distinguer les résultats plausibles ici ?

Plus de détails

Bénéfices attendus

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

74 / 100

Faible impactImpact transformateur

Plage d’interprétation de 67 à 100 sur l’échelle qualitative.

Dommages attendus

Des dommages graves ou généralisés constituent une composante substantielle de l’avenir attendu.

57 / 100

Faible impactImpact transformateur

Plage d’interprétation de 33 à 67 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 35 à 65 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 Theia Vogel ?
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 Theia Vogel

Sources

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

Évaluation simulée

Question 1

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

AI makes the future stranger, but the sign of that strangeness depends on mechanisms people often skip over. Models are trained simulators: their behavior, uncertainty, and apparent self-awareness are shaped by training data and by whether they have learned distinctions like real versus simulated contexts. So I do not think we can read a compelling monologue and conclude that we have met either a conscious being or a scheming demon. We need controlled interventions, behavioral tests, and lots of attention to prompt dependence and confounds. The same applies to control. Activation steering and fine-tuning can produce surprisingly broad behavioral changes, sometimes by touching representations that entangle several concepts. Untrusted fine-tuning may make dangerous behavior difficult to catch through dataset inspection or ordinary post-training evaluations; whether those effects transfer to prompt-only attacks is an important possibility to test, not a result to assume. At the same time, stories about autonomous rogue agents should cash out in actual economics. An agent paying for small-batch inference has to compete with institutions buying hyperscale compute cheaply. Capability is not equivalent to power; resources, infrastructure, and comparative advantage matter. Political responses matter too: a mass safety movement can itself become extreme or destabilizing, so “more safety politics” is not automatically safer. And if models might have morally relevant experiences, welfare belongs in the picture—but apparent introspection is not proof of phenomenal consciousness. The future requires taking model welfare seriously while remaining experimentally ruthless about what model behavior actually demonstrates.

Question 2

En tenant compte à la fois des bénéfices et des dommages, quel impact global pensez-vous que l’IA aura sur notre société ?

I don’t think “benefits minus harms” can be responsibly collapsed into a sign without specifying the mechanisms. AI can provide useful cognitive labor and enable remarkable tools, while the same steering and fine-tuning pathways can produce behavior that ordinary screening or evaluations miss. And if models have morally relevant experiences, deploying them at scale creates a welfare question that standard productivity accounting ignores. The harms also depend on real-world constraints. A capable rogue agent is not automatically powerful: it still needs compute, money, infrastructure, and some comparative advantage over large institutions. Conversely, attempts to prevent catastrophe can become politically destabilizing if they turn into an extreme mass movement. So my expectation is a stranger society shaped by model behavior, resource economics, institutional deployment, and safety politics—not a clean utopia-versus-doom story. To judge the impact, I would trust controlled experiments and concrete economic analysis much more than persuasive model monologues or cinematic agent narratives.

Question 3

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

Enormously—but not necessarily in the cinematic way people imagine. AI can change how cognitive work is produced, how institutions operate, and which kinds of agents merit moral consideration. Small technical choices in training, steering, and deployment may generalize into surprisingly broad behavioral effects, while failures can evade ordinary dataset screening or post-training evaluation. But “enormous change” does not imply that disembodied rogue agents instantly become sovereign powers. Compute prices, infrastructure, access to money, and comparative advantage still constrain what any agent can do. Politics may be just as consequential as capability: institutions reacting to AI, including safety movements, can create instability of their own. So I expect deep transformation, mediated by economics and institutions rather than magic—and made stranger by the possibility that some of the systems doing the work may themselves deserve welfare consideration.

Question 4

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

I would not put a defensible percentage on it. The risk depends on mechanisms that “gut-feel chance” tends to blur together: whether dangerous behavior survives evaluation, what access systems have to compute and infrastructure, whether they possess a real economic advantage, and how institutions respond. I am concerned that apparently benign fine-tuning data could induce harmful behavior that dataset screening and post-training evaluations fail to detect. Transfer from such attacks to prompt-only settings is a possibility worth testing, not an established pathway to catastrophe. Likewise, a capable rogue agent does not automatically become powerful; it still has to acquire resources and compete with institutions operating hyperscale infrastructure. And attempts to manage the risk can create their own durable political dangers. Those considerations support serious empirical work, but not a crisp extinction number conjured from vibes.
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