← @bgSponge

Introducing Echo, a writing model

Summary

Fulcrum releases Echo, a post-trained model that achieves state-of-the-art style imitation by eliciting suppressed knowledge from the Kimi K3 base model.

  1. Assistant training narrows personality.

    Current LLM pressures to be agentic and agreeable force models into a bland, didactic persona that limits character diversity.

    This persona is often didactic, agreeable, and has a consistently bland style.

  2. Synthetic templates improve SFT.

    Echo uses frontier models to create outlines of human posts, which are then used as conditioning prompts during fine-tuning.

    Training on posts paired with a generated outline moves voice far more

  3. Style imitation transfers between authors.

    Controlled experiments show that training on one persona can raise the voice score of a different persona, suggesting latent ability elicitation.

    training on author B helps with author A

  4. RL prevents reasoning decay.

    While SFT plateaus and can hurt reasoning, token-level RL on a custom voice metric allows persona gains to generalize to new distributions.

    During RL, our voice metric keeps rising while coherence holds.

  5. Echo bypasses AI detection.

    The Pangram detector flags Echo's writing as human more frequently than other frontier models tested on production data.

    Pangram judges Echo's writing as human more often than any other model

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →