Tech • IA • Crypto
AI Could Have Consciousness, New Findings Suggest!
7/10TL;DR
Recent advances shed light on the internal workings of AI models while coinciding with notable product changes at Anthropic.
KEY POINTS
An internal “working memory” finally better understood
Recent work highlights a key mechanism of LLMs: a form of transient internal representation, distinct from visible outputs. This “working memory” allows the model to link intermediate concepts that were not explicitly requested, facilitating calculations, analogies, and logical chains. It may explain why some answers seem to go beyond simple token sequencing.
Origins of behaviors seen as “moral” or “strategic”
This mechanism sheds light on behaviors sometimes interpreted as cunning, ethical consistency, or “intentions.” Rather than consciousness, these are emergent effects from complex internal associations shaped by training. Models do not “want” anything, but activate latent structures that produce responses perceived as intentional.
Clearing the fog around AI “consciousness”
The debate over possible AI consciousness remains open but largely philosophical. The lack of a shared definition makes firm conclusions fragile. New explanations point toward a functional illusion: an ability to simulate rich reasoning without real subjectivity.
Major overhaul of the Claude interface
Anthropic is reorganizing Claude. The former chat / work / code trio is evolving into a structure where chat and work are merged into a main space, while code becomes a separate module. The goal is to simplify usage while clarifying advanced features.
“Work” arrives on web and mobile
The Claude Work feature is expanding beyond the local app. It will be available on web and mobile, enabling long-running tasks executed server-side. Users can start an action and retrieve the result later, without relying on their own machine.
Strategic positioning against Mistral
This overhaul comes as Mistral recently adopted a similar organization. Anthropic regains the edge with deeper integration and a unified experience, confirming intensifying competition in consumer AI interfaces.
Unexpected extension of Claude Fable
Initially set to end on July 7, the Claude Fable feature is extended until July 12. This extra time gives users a few more days to use the tool before it is removed or possibly transitioned to a paid offering.
CONCLUSION
These developments reflect both a better scientific understanding of AI models and accelerating product competition, where usability and advanced capabilities are becoming central stakes.