Tech • IA • Crypto
We Discover AI Could Have Consciousness! 2/2
9.3/10TL;DR
New research from Anthropic reveals internal mechanisms in large language models that resemble a form of “working memory,” fueling debate over whether AI could one day exhibit traits associated with consciousness.
KEY POINTS
Anthropic’s breakthrough on model interpretability
Researchers at Anthropic analyzed internal processes of a recent Claude model and identified structured activity patterns resembling a global workspace, a concept from cognitive science. This suggests models may coordinate information across subsystems rather than operate as purely fragmented statistical engines.
Emergence of “working memory”-like behavior
The study highlights what appears to be a form of transient memory inside the model, enabling it to track intermediate reasoning steps. While not equivalent to human awareness, this mechanism allows more coherent, multi-step problem solving and may explain recent gains in reasoning performance.
No evidence of true consciousness
Despite these findings, there is no scientific proof that AI systems possess subjective experience or self-awareness. Experts emphasize that current systems simulate understanding through computation and pattern prediction, without any অনুভ or internal perspective.
Debate over definitions of consciousness
The findings reignite philosophical disagreements about what “consciousness” means. Competing frameworks such as Global Workspace Theory and Integrated Information Theory offer models, but none provide a definitive test applicable to AI systems.
Practical implications outweigh philosophical ones
Researchers stress that the urgent issue is not whether AI is conscious, but how to evaluate, align, and govern increasingly autonomous systems. Even without consciousness, their real-world impact on economies, information flows, and decision-making is already significant.
Rapid progress in generative capabilities
Parallel developments include early-stage tools from Meta capable of generating simple video games from prompts and assets. While far from blockbuster production, such systems could produce functional games within one to two years, with high-end outputs potentially taking five to seven years.
Upcoming model releases and competitive pace
A new iteration, GPT‑5.6, is reportedly in closed beta testing, reflecting the accelerating release cycle among AI labs. Public availability is expected in the near term, continuing the rapid evolution of model capabilities.
CONCLUSION
Anthropic’s findings deepen understanding of how advanced AI systems operate, but they stop short of demonstrating true consciousness, leaving governance and real-world impact as the more immediate concerns.