Conjecture / EleutherAI (cognitive emulation)
Does emulation-style controllability still imply corrigibility—that systems remain open to correction—as capability scales?
Introduction
Conjecture pursued Cognitive Emulation (CoEm)—controllable LLM stacks framed as emulation rather than opaque end-to-end agents—alongside the EleutherAI open-model lineage. The cluster asks whether emulation-style controllability implies corrigibility as capability scales. Correction-channel integrity and successor value transport remain separate load-bearing concerns beyond controllability claims.
Who carries it: Conjecture (Connor Leahy, Sid Black, Gabriel Alfour; winding down 2026); EleutherAI open-science institute; Leahy → ControlAI (pause advocacy cousin)
What they aim to do. Build controllable systems using an emulation framing rather than opaque end-to-end agents.
The hard question. Does emulation-style controllability still imply corrigibility—that systems remain open to correction—as capability scales?
What they produce. Cognitive Emulation (CoEm) and controllable LLM stacks, plus the EleutherAI open LLM lineage.
Key terms. Key terms include cognitive emulation, controllable LLMs, emulation versus alignment, EleutherAI, and open models.
Related field cruxes. Inner Alignment
What they contribute. An alternative controllability framing; EleutherAI spinoff lineage; a Redwood-adjacent prosaic alignment spirit.
How this project treats it. Correction-channel integrity and successor value-bundle-transport remain separate load-bearing concerns; Conjecture’s active bet is partly historical; ControlAI overlap is advocacy, not CoEm research.
Links
See the coverage matrix for evidence tagged to this agenda, and the glossary for shared terms.