Introducing Echo, a writing model

gist

Fulcrum releases Echo, a Kimi K3 writing model trained for under $5K that leads held-out style Elo and is flagged human more often than frontier models. The lab argues assistant post-training squeezes models into a bland persona, then describes supervised fine-tuning on outline-conditioned human posts and token-level reinforcement learning on a voice likelihood ratio that generalizes past the eight author experts in training.

ideas

• Assistant post-training narrows model character. Agentic competence plus a didactic, agreeable assistant persona push models into a bland region of mind space that Echo aims to leave.
• Style Elo ranks imitation without those features in training. Held-out stylometric features put Echo ahead of frontier models on authors seen in training and authors never seen.
• Pangram is a poor slop proxy. Fulcrum wants writing that sounds strong yet stays easy to detect as AI; Pangram still judges Echo human more often than the other models tested.
• Outline-conditioned supervised fine-tuning moves voice farther than raw posts. A frontier model writes a brief from a human post, Echo trains on the original text given that prompt, and training on persona B raises persona A's voice score almost as much as training on A.
• Token-level reinforcement learning carries voice onto new tasks. After supervised fine-tuning plateaus, reinforcement learning on an author-expert minus base-model likelihood ratio extends persona gains into chat and varied writing prompts using only eight author experts.

quotes