Introducing Deft v1
Deft models produce higher quality writing, without slop
Slop. It’s not just annoying — it’s exhausting. You’re absolutely right to be annoyed by it, and in this blog we delve into a solution.
You've probably noticed, LLMs have certain phrases they overuse, like "delve", "It's not just X — it's Y", or "You’re absolutely right". These come from weaknesses in the frontier model training algorithms, which are designed for coding or math or other problem solving tasks.
For Deft, we spent a full year developing our own custom training algorithm, one that is designed to improve model writing quality and creativity, without falling back to using the same phrases over and over.
When we tested Deft, we found that compared to the baseline training algorithm, it:
- Improved overall quality by 63%
- Improved creativity by 164%
- Improved coherence by 28%
- Improved meaningful detail (instead of speaking in vague generalities) by 146%
and at the same time, removing overused patterns that other LLMs use (scores according to a judge model).
For this early release of Deft, we are almost using the same model as the demo. For the full launch (coming very soon!) we're going to give you much better control over the output text, as well as a quality boost in the output.