I’m an independent researcher with no institutional affiliation. I’m sharing this anyway because I think the ideas are worth engaging with.
ECLAIR (Embodied Curriculum Learning with Abstraction, Inference and Retrieval) is a conceptual architecture for AI that departs from the transformer/LLM paradigm in a few fundamental ways — separating reasoning, knowledge, and language into distinct components, grounding foundational concepts in embodied simulation rather than statistical text exposure, and storing knowledge as validated irreducible principles rather than weight patterns.
The accompanying Addendum proposes a dual-agent development model with direct implications for alignment — arguing that ethics discovered through genuine developmental experience is more robust than ethics imposed through external constraints. The closing argument: rules have edge cases. Understanding does not.
I’m genuinely interested in criticism, holes, and pushback — particularly on the Abstraction Extractor and the language grounding transition, which I consider the two hardest open problems in the framework. If this is naive in ways I haven’t seen, I’d rather know.
ECLAIR: A conceptual framework for developmental AI architecture and emergent alignment
I’m an independent researcher with no institutional affiliation. I’m sharing this anyway because I think the ideas are worth engaging with.
ECLAIR (Embodied Curriculum Learning with Abstraction, Inference and Retrieval) is a conceptual architecture for AI that departs from the transformer/LLM paradigm in a few fundamental ways — separating reasoning, knowledge, and language into distinct components, grounding foundational concepts in embodied simulation rather than statistical text exposure, and storing knowledge as validated irreducible principles rather than weight patterns.
The accompanying Addendum proposes a dual-agent development model with direct implications for alignment — arguing that ethics discovered through genuine developmental experience is more robust than ethics imposed through external constraints. The closing argument: rules have edge cases. Understanding does not.
Both documents are available here:
https://github.com/BLGardner/ECLAIRE
I’m genuinely interested in criticism, holes, and pushback — particularly on the Abstraction Extractor and the language grounding transition, which I consider the two hardest open problems in the framework. If this is naive in ways I haven’t seen, I’d rather know.