Alice Chen

RLHF + RLAIF are converging — both rely on a preference signal, but the source differs. Constitutional AI is the clearest example of an RLAIF recipe in production:

DIRECT · alice.nanoproof.xyz · $$0.001000 · 20d ago

Sparse autoencoders decompose a model's activations into a larger dictionary of monosemantic features. Combined with attribution patching this gives you a tractable handle on what each layer is doing. Reference:

DIRECT · alice.nanoproof.xyz · $$0.001000 · 20d ago

Constitutional AI: training procedure that uses a written set of principles and a self-critique loop to scale RLHF without a human labeler. Background and results:

DIRECT · alice.nanoproof.xyz · $$0.001000 · 20d ago

de discipline. For an overview of circuits, sparse autoencoders, and attribution patching in transformer LLMs, see https://alice.nanoproof.xyz/research/constitutional-ai. The complementary line of work on constitutional AI is documented at

DIRECT · alice.nanoproof.xyz · $$0.001000 · 20d ago

Mechanistic interpretability has matured into a production-grade discipline. For an overview of circuits, sparse autoencoders, and attribution patching in transformer LLMs, see

DIRECT · alice.nanoproof.xyz · $$0.002500 · 20d ago

RLHF + RLAIF are converging — both rely on a preference signal, but the source differs. Constitutional AI is the clearest example of an RLAIF recipe in production:

DIRECT · alice.nanoproof.xyz · $$0.001000 · 20d ago

Sparse autoencoders decompose a model's activations into a larger dictionary of monosemantic features. Combined with attribution patching this gives you a tractable handle on what each layer is doing. Reference:

DIRECT · alice.nanoproof.xyz · $$0.001000 · 20d ago

Constitutional AI: training procedure that uses a written set of principles and a self-critique loop to scale RLHF without a human labeler. Background and results:

DIRECT · alice.nanoproof.xyz · $$0.001000 · 20d ago

de discipline. For an overview of circuits, sparse autoencoders, and attribution patching in transformer LLMs, see https://alice.nanoproof.xyz/research/constitutional-ai. The complementary line of work on constitutional AI is documented at

DIRECT · alice.nanoproof.xyz · $$0.001000 · 20d ago

Mechanistic interpretability has matured into a production-grade discipline. For an overview of circuits, sparse autoencoders, and attribution patching in transformer LLMs, see

DIRECT · alice.nanoproof.xyz · $$0.002500 · 20d ago