Princeton researcher Peter Henderson is weighing a short-lived reinforcement learning venture whose only mission would be scrubbing the repetitive, unnatural prose that frontier models keep generating, after which the company would close up shop.
Seriously thinking of starting an RL env company with the sole purpose of fixing the slop/neuralese writing style in frontier model outputs. Then shutting it down once its public benefit purpose is fulfilled. So tired of the slop.
No entity has been incorporated and no technical details or timelines have surfaced, leaving the concept as an open experiment rather than a ready product.
If pursued, the effort would test whether targeted RL environments can push models past their current stylistic ruts without creating a permanent new vendor.
Users are encouraging the researcher to launch an RL company fixing slop in frontier model outputs, offering direct support for pursuing the idea.
No Digg Deeper questions have been answered for this story yet.
@PeterHndrsn you can do it peter!!
Seriously thinking of starting an RL env company with the sole purpose of fixing the slop/neuralese writing style in frontier model outputs. Then shutting it down once its public benefit purpose is fulfilled. So tired of the slop.
@PeterHndrsn w/ recent GDM research, I'd say, we also need an SFT env company atp
@PeterHndrsn could u start with the obvious ai x slop
@PeterHndrsn AI slop being recognizable is a public benefit
@PeterHndrsn The fact that "AI generated" is becoming a recognizable writing style shows there's still a lot of room for improvement.