What's this about?
A new AI lab called Deft launched, cofounded by rosmine and jmrphy, focused entirely on writing quality. Two claims stand out. The launch announcement was written by the Deft model itself. And 86% of user queries are rated fully human by Pangram, the AI-detection service. It's in public beta, explicitly described as a small model that may make mistakes, launched early to gather feedback before scaling.
What exactly does "86% rated fully human by Pangram" mean?
Pangram is a detector that estimates whether text was AI-generated. So the claim is that most of what the model writes doesn't read as machine-written to a system built to spot exactly that. Two caveats matter. Detectors are imperfect in both directions, and optimizing to defeat one is a narrower target than writing well. Ethan Mollick made precisely this point the same day about a different company: the output of an AI-detector-fooling startup is bad writing. Evading detection and writing well overlap but are not the same objective.
Why is this worth noticing?
- A lab specializing in one capability is a notable structural bet in a field where everyone builds general models and treats writing as an emergent property.
- The self-referential launch is a real demonstration, since you can read the announcement and judge the product directly rather than trusting a benchmark.
- It lands in the AI;DR week. The complaint driving that coinage was unedited AI prose — a lab explicitly targeting the quality problem is a response to the same irritation, though whether "undetectable" is the right proxy for "good" is exactly the open question.