About Deft
Better AI Writing
Deft is an new AI lab that makes LLMs better at writing. We developed a new training algorithm (Distribution Fine Tuning) that makes the LLM outputs match human style much better than existing models. We offer a variety of solutions to help improve your writing, including the Deft models, custom models trained on your style, and slop analysis to improve your existing model outputs.
System
What makes it different
Most AI models are trained to be good for problem solving, where there is one correct answer. This makes the outputs repetitive, as if it is trying to follow a formula of certain phrases/concepts that will give "correct" writing. Real writing doesn't have a single correct form or style, variety is necessary for it to be good. Deft better captures the style of human writing, using a custom algorithm designed to make model outputs more like human writing.
This algorithm is called Distribution Fine Tuning, you can read more about it in our research article, where we rigorously show that it is better at producing outputs that are more human than other training methods.
- 01
Bring your ideas
Put a prompt, draft, outline, fragment, or generic model output into the editor.
- 02
Generate with Deft
The model usually takes .5-1.5 minutes to generate a response, depending on the length and thinking level (faster or smarter)
- 03
Edit the result
The console returns an editable draft with lightweight metrics and optional checking tools, so you can make any changes you want.
FAQ
Frequently asked questions
What is distribution fine-tuning (DFT)?
Standard training algorithms make the model outputs very formulaic, they overuse the same phrases (like emdashes or "It's not X, it's Y") that make the output very boring and painful to read. DFT makes the model outputs better match human writing samples, so you get much more natural sounding text without those overused slop patterns.
How is Deft different from ChatGPT or Claude?
General assistants are designed to be good at problem solving, where there is a single right answer. This makes their outputs formulaic, as if there is only one way to make good writing. Deft models are designed to be more creative, and have more varied phrasing and content.
What kind of writing is Deft designed to produce?
Deft is designed to produce clear, engaging, varied, human-quality prose. The practical test is whether the writing communicates the idea, preserves what matters, and is enjoyable to read.
Can I try Deft without paying?
Yes. You get three free generations before signing in. After you create an account, the free plan includes two requests per day. Paid website plans offer larger monthly allowances, and usage-based billing applies separately to the server-to-server API.
How much does the Deft API cost?
API generation is billed by usage: $2.50 per million customer-supplied prompt, style, and rewrite-instruction tokens and $12 per million final DFT output, preprocessing thinking tokens, and unselected DFT candidate output, rounded up to the nearest cent with a one-cent minimum per request. Internal prompt-template and routing input, candidate scoring, judging, and post-processing tokens are not charged.
Am I charged for failed generations?
No. A failed generation releases its reservation and is never billed.
How do I create a Deft API key?
Sign in, open the API section of Account, and create a key. No access request or manual approval is required.