The researchers say Talkie needed period-appropriate fine-tuning for emergent misalignment.
A researcher says new work with a collaborator reproduced emergent misalignment and subliminal learning in Talkie. They also say the model needed period-appropriate fine-tuning for emergent misalignment.
67.5K
11 Sources, first seen ago
—
Not ranked yet
—
Not ranked yet
@DavidDAfricaSome new work with @status_effects: Personas in a Pre-Internet Language Model. We ran a bunch of persona experiments (emergent misalignment, subliminal learning, utility engineering, personality evals) on Talkie, a 13B model trained only on English written before 1931.
@status_effectsNew work with @DavidDAfrica: We reproduced emergent misalignment and subliminal learning in talkie, a model trained on pre-1931 text.
@rgblongI’ve been so excited to see alignment science on Talkie! yes!!@DavidDAfricaSome new work with @status_effects: Personas in a Pre-Internet Language Model. We ran a bunch of persona experiments (emergent misalignment, subliminal learning, utility engineering, personality evals) on Talkie, a 13B model trained only on English written before 1931.@status_effectsNew work with @DavidDAfrica: We reproduced emergent misalignment and subliminal learning in talkie, a model trained on pre-1931 text.@rgblongI’ve been so excited to see alignment science on Talkie! yes!!@DavidDAfricaSome new work with @status_effects: Personas in a Pre-Internet Language Model. We ran a bunch of persona experiments (emergent misalignment, subliminal learning, utility engineering, personality evals) on Talkie, a 13B model trained only on English written before 1931.@status_effectsNew work with @DavidDAfrica: We reproduced emergent misalignment and subliminal learning in talkie, a model trained on pre-1931 text.@rgblongI’ve been so excited to see alignment science on Talkie! yes!!@DavidDAfricaSome new work with @status_effects: Personas in a Pre-Internet Language Model. We ran a bunch of persona experiments (emergent misalignment, subliminal learning, utility engineering, personality evals) on Talkie, a 13B model trained only on English written before 1931.@status_effectsNew work with @DavidDAfrica: We reproduced emergent misalignment and subliminal learning in talkie, a model trained on pre-1931 text.@rgblongI’ve been so excited to see alignment science on Talkie! yes!!
