Management Summary:

The report under consideration analyzes the technological evolution towards Artificial Superintelligence (ASI), defined as intelligence that exceeds human cognitive abilities in all domains. It describes a gradual progression from the currently dominant Artificial Narrow Intelligence (ANI) to Artificial General Intelligence (AGI) and then to ASI, based on concrete advances key technologies such as:

• Diverse reinforcement learning techniques

• Transformer and Mixture-of-Experts architectures

• Multimodal systems and external memory

• Methods of self-improvement and metacognitive reasoning

Critical limitations are identified, such as the absence of embodiment, the lack of causal understanding, and symbolic disconnection. The report identifies cognitive principles such as structured memory, symbolic anchoring, and sensory interaction as being essential to exceed current capabilities.

The report highlights the major risks associated with AGI/ASI: loss of control, misalignment of objectives, and malicious use. Furthermore, it is emphasized that current regulatory frameworks, most notably the EU AI Act, are inadequate when confronted with the potential of self-improving autonomous systems.

Finally, a summarized technology roadmap is proposed, divided into three phases (current AI, transition to AGI, speculative horizon of ASI), and based on four cross-cutting priorities: sustainability, security, agile regulation„ and international cooperation.

Research Paper:

article

Source: CYD Campus

BibTex:

@article{vallotton2025superintelligenceAn,
  title = {Superintelligence: An Initial Assessment and Trend Analysis},
  author={Vallotton, Alexandre},
  journal = {CYD Campus},
}