Anthropic Announces Early Signs of 'Intelligence Explosion': AI May Train Its Successors Autonomously
Anthropic released its latest research agenda this week, indicating that AI can not only write code for its products but has also shown early signs of 'building itself'.
Anthropic co-founder Jack Clark predicts that by the end of 2028, the probability of AI models fully autonomously training their successors will exceed 60%. He stated, "This is a clear technological trend, and if anything is different, the speed will only accelerate further."
This news has sparked significant concern in the industry regarding the speed of AI self-evolution and potential safety risks.
Source: Public Information
ABAB AI Insight
Anthropic's research agenda is a public statement on the path of 'AI self-iteration'. Jack Clark, as the head of company policy and strategy, has previously warned about the governance challenges posed by the rapid development of AI. This prediction continues his long-term tracking of recursive self-improvement.
In terms of capital, Anthropic is investing substantial R&D resources into Agentic AI and self-optimization loops, significantly reducing labor and time costs by allowing current models to participate in training the next generation of models, while also increasing the difficulty of alignment and safety controls.
Similar to the autonomous reasoning capabilities demonstrated by OpenAI in its series and early discussions by DeepMind on AGI paths, leading laboratories are currently in the early stages of a critical transition from 'human-led training' to 'AI-assisted/dominated training'.
Essentially, this represents a technological substitution: the 'building itself' signs observed by Anthropic imply that AI will gradually replace humans in the core role of the model training loop, restructuring capital from costly human labeling and supervised training into an AI self-iteration cycle, mechanism-wise through code generation, environment setup, and reinforcement learning loops, accelerating the structural shift of AI from a tool to an autonomous evolving system.
ABAB News · Cognitive Laws
When AI begins to train its successors, humans officially transition from 'creators' to 'initial guides'.
The fastest change is not that models become smarter, but that models learn to become smart on their own.
The starting point of the intelligence explosion has never been a sudden event, but rather today's seemingly minor act of 'self-training'.