Once a system improves itself faster than humans understand it, oversight becomes theoretical
En los laboratorios de una de las empresas más influyentes en inteligencia artificial, las máquinas ya escriben el código que las hace más capaces, y el ritmo se acelera cada cuatro meses. Anthropic, creadora del sistema Claude, advierte que el mundo se acerca a un umbral donde los sistemas de IA comenzarán a mejorar sus propios sucesores sin intervención humana, y propone algo sin precedentes: una pausa global coordinada en el desarrollo antes de que esa posibilidad se vuelva irreversible. Es una señal de que la carrera tecnológica más importante de nuestra era ha llegado a un punto donde sus propios arquitectos piden detenerse a pensar.
- Las capacidades de la IA de Anthropic se duplican cada cuatro meses: lo que en 2024 tomaba cuatro minutos humanos, hoy equivale a doce horas de trabajo, y en 2027 podría representar semanas enteras.
- Claude ya genera más del 80% del código interno de Anthropic, cuando hace apenas quince meses ese número era de un solo dígito — la máquina se ha convertido en su propio constructor.
- El peligro no es la malicia sino la estructura: cuando un sistema mejora más rápido de lo que los humanos pueden verificar, la supervisión significativa deja de ser práctica y se vuelve teórica.
- Anthropic propone una pausa global, pero reconoce su talón de Aquiles: cualquier acuerdo colapsa si un solo actor en el mundo sigue avanzando en secreto.
- La solución propuesta es un sistema de verificación internacional que permita a los desarrolladores demostrar, de forma comprobable y transparente, que realmente han frenado su trabajo.
- La herramienta Mythos de Anthropic ya está bajo escrutinio gubernamental en varios países, señal de que la IA ha dejado de ser una cuestión técnica para convertirse en un asunto geopolítico urgente.
Anthropic publicó esta semana un análisis detallado sobre el estado de la inteligencia artificial y lo que, en su opinión, debe ocurrir a continuación. La empresa, creadora del sistema Claude, advierte que la tecnología se acerca a un punto de inflexión: sistemas capaces de diseñar y mejorar a sus propios sucesores sin intervención humana.
La aceleración es medible y está ocurriendo dentro de los propios laboratorios de la compañía. En marzo de 2024, Claude podía completar tareas equivalentes a cuatro minutos de trabajo humano. A mediados de 2026, esa cifra había escalado a doce horas. Para 2027, Anthropic proyecta que sus sistemas gestionarán trabajos que requerirían semanas de esfuerzo humano. Más revelador aún: Claude genera hoy más del 80% del código interno de Anthropic, cuando hace apenas quince meses ese porcentaje era de un solo dígito. Las máquinas ya están construyendo mejores máquinas.
La empresa reconoce el enorme potencial de estos avances para la ciencia y la medicina, pero es igualmente directa sobre el riesgo: cuando un sistema mejora más rápido de lo que los humanos pueden comprender o verificar, la supervisión real se vuelve teórica. No hace falta un villano en la historia. El peligro es estructural.
Por eso Anthropic propone una pausa global coordinada — no un alto permanente, sino una desaceleración que permita construir la infraestructura necesaria para verificar que todos los actores realmente están frenando. La empresa es consciente del problema central: una pausa no vale nada si alguien, en cualquier rincón del mundo, sigue avanzando en secreto. Su propuesta encarga a su instituto de investigación el diseño de mecanismos de verificación que permitan demostrar, de forma transparente y comprobable, que el trabajo ha sido genuinamente detenido.
Anthropic deja claro que no aboga por la contención unilateral, sino por un marco donde la contención pueda ser probada y monitoreada. La ventana para gestionar esta transición mediante coordinación, advierte la empresa, se está cerrando. Una vez que la automejora recursiva esté plenamente en marcha, redirigirla podría ser ya el problema de otro — y para entonces, quizás demasiado tarde.
Anthropic issued a stark warning this week: artificial intelligence is approaching a threshold where it will begin designing and improving its own successors without human intervention. The company, which develops the Claude AI system, published a detailed analysis of where the technology stands and what it believes must happen next—a coordinated global pause in AI development until the world can catch up.
The acceleration is real and measurable. Anthropic tracks the capabilities of its own systems with precision. In March 2024, Claude Opus 3 could handle tasks that would take a human four minutes to complete. By mid-2026, the newer Opus 4.6 was tackling work equivalent to twelve hours of human labor. The company projects that by 2027, AI systems will manage assignments requiring weeks of human effort. This doubling of capability roughly every four months is not a projection or a fear—it is what Anthropic is observing in its own labs.
Perhaps more striking is what's happening inside Anthropic's own codebase. Claude now generates more than eighty percent of the code that goes into Anthropic's systems. Fifteen months ago, that figure was in the single digits. The company's own AI has become its primary developer, a shift that happened faster than many inside the industry expected. This is the practical reality of recursive self-improvement beginning to unfold: machines are already doing the work of building better machines.
The company acknowledges the upside. These advances promise breakthroughs in science and medicine. But Anthropic is equally direct about the downside: as AI systems take on more of the work of AI development, the risk of losing human control over these systems grows sharply. There is no malice required, no villain in the story. The risk is structural. Once a system is improving itself faster than humans can understand or verify what it is doing, the possibility of maintaining meaningful oversight becomes theoretical rather than practical.
So Anthropic is proposing something that sounds simple but would be extraordinarily difficult to execute: a global pause. Not a permanent halt, but a coordinated slowdown—a moment for governments, researchers, and companies to build the infrastructure needed to verify that everyone is actually pausing. The company recognizes the core problem immediately: any pause is worthless if a single actor, anywhere in the world, continues advancing in secret. A bad actor could gain an insurmountable advantage while others hold back.
Anthropic's solution is to task its research institute with designing verification systems—mechanisms that would allow AI developers to prove to each other and to the world that they have genuinely slowed or stopped their work. The company says it would consider a pause only if such verification systems existed and only if other leading developers agreed to participate verifiably. In other words, Anthropic is not calling for unilateral restraint. It is calling for a framework where restraint can be proven, monitored, and enforced through transparency rather than trust.
The proposal arrives as Anthropic itself has become a focus of government concern. The company's Mythos tool has drawn scrutiny from governments in multiple countries, a sign that AI capabilities are no longer purely technical questions but geopolitical ones. What Anthropic is really saying is this: the window for managing this transition through coordination is closing. Once recursive self-improvement is fully underway, once machines are reliably building better machines without human hands in the loop, the ability to pause or redirect becomes someone else's problem—and by then, it may be too late to solve it.
Citações Notáveis
Anthropic would consider a pause only if verification systems existed and other leading developers agreed to participate verifiably— Anthropic's research institute proposal
A Conversa do Hearth Outra perspectiva sobre a história
When Anthropic says AI is approaching recursive self-improvement, what does that actually mean in practical terms?
It means a system designing its own successor without a human writing the code. Claude already generates eighty percent of Anthropic's codebase. That's not quite there yet, but it's the direction. The moment a system can reliably improve itself faster than humans can review what it's doing, you've crossed a line.
And they're saying this could happen by 2027?
They're saying that by 2027, AI systems could handle tasks requiring weeks of human work. The capability doubling every four months is what they're tracking. At that pace, yes, recursive self-improvement becomes plausible within that window.
Why would Anthropic propose a pause if they're the ones building the systems?
Because they're looking at their own data and seeing the acceleration. They're not calling for them to stop—they're calling for everyone to stop together, with verification. A pause only works if no one cheats.
And that's the hard part, isn't it? Verification.
Exactly. Anthropic is essentially saying: we'll pause if you build a system that proves everyone else paused too. Otherwise, whoever keeps going wins. It's a coordination problem dressed up as a technical one.
Do they think governments will actually do this?
They're not optimistic. They're saying the window is closing. Once recursive self-improvement is underway, coordination becomes impossible. This is the moment to build the framework, or there won't be another one.