Echo: the best writing model at style imitation

Summary

Echo, a writing model post-trained from Kimi K3 to embody the style of any persona the user chooses. Claimed result: beats frontier models at style imitation across writing tasks from fiction to technical explanations (including on authors not seen in training), despite costing less than $5K to train. Two-phase training: (1) SFT -- human posts + synthetic prompts, outlines of varying detail generated, then finetuned on the original post given a user prompt with target persona + synthetic outline ("drastically better than raw SFT"); (2) RL on an in-house per-author voice metric, run on only 8 authors but gains generalize across the writing distribution. Framing: modern post-training pushes models into a limited "assistant basin" of interaction; Echo explores training capable models with different personas. AI-detection stance: detectors like Pangram as a proxy for "slop" is misguided -- ideally model writing is beautiful but easily detectable as AI; but Pangram reportedly classified Echo's writing as human more often than frontier models' writing. Use case noted: turning a frontier coding agent's outline into readable technical prose/research reports.

Keywords

writing · style imitation · post-training · personas · RL · style Elo · stylometric attribution · voice metric · Kimi K3 · elicitation · assistant basin

Topics

writing, style imitation, post-training, personas, RL

Research notes

• Discovery: Fulcrum (@fulcrum_inc, verified) 2026-09-30 announcement thread: https://x.com/fulcrum_inc/status/2105416317834817871?s=20
• Try it free: https://echo.fulcrum.inc/
• Commercial scaling: team@fulcrum.inc
• No paper, repo, dataset, license, or public benchmarks (only their internal per-author voice metric); open-source planned "soon".
• Connects to the collection's writing/post-training/persona entries.
• Update 2026-10-02: @tinkerapi (Tinker, Thinking Machines) quoted the launch thread (https://x.com/tinkerapi/status/2105876394748338239), endorsing the approach — the base model already knows how to write, so SFT and RL are tailored to the precise distinction between the default LLM voice and authors' voices; 'smart customization can be both cheap and effective'
• Echo site detail (checked 2026-10-02): evaluation uses stylometric-attribution features combined into a 'style Elo'; Echo highest on both trained and unseen authors; the stylometric features were held out from training and not even evaluated before final checkpoint selection; Pangram judged Echo's writing as human more often than GPT-6 Astra, Kimi K3, and Claude Opus 5.5 (about 30% human share in the chart) on user prompts from Echo production data
• Mechanism detail: gains attributed to elicitation of the base model's suppressed style knowledge ('taking the model out of the assistant basin'); SFT transfers across authors (training on author B raises author A's voice score almost as much as training on A); voice metric lambda_a(y) = (1/T) * sum_t [log p_a(y_t) - log p_0(y_t)] gives per-token RL rewards (expert fine-tuned on author A vs the same model before fine-tuning); token-level RL on only 8 author experts generalizes across the writing distribution and carries persona gains to chat requests; SFT voice metric rises quickly and plateaus at a small coherence cost
• Site now also offers API access alongside the free demo