If you’ve spent any time prompting Large Language Models, you’ve probably noticed they have a very specific voice.
They love the word tapestry. They’re obsessed with anything shimmering.
And characters are constantly speaking in a voice barely above a whisper.
This phenomenon is called slop.
Slop is the repetitive, statistically over-represented phrasing that makes AI-generated text feel obvious.
We just wrapped up a demo showcasing a new inference-technique pipeline designed to kill the slop and force the model to be creative.
Here’s the workflow we built:
How It Works (The “Secret Sauce”)
If you simply ban the word apple, the model breaks when you ask for a fruit-pie recipe.
The Antislop Sampler takes a different approach. As the model generates tokens, the sampler watches for thousands of overused clichés identified through forensic analysis of billions of tokens.
When the model tries to output a cliché like “a profound sense of…” or “neon-soaked streets,” the sampler pauses, rewinds the generation to the start of that phrase, down-weights the probability of the cliché, and forces the model to find a more original way to express the idea.
It also supports soft-banning, meaning that if the context absolutely requires the word, it allows it, but otherwise pushes for variety.
The Results
The difference is night and day.
Curious to see it in action?
Check out the full demo environment and see how the Antislop Sampler transforms predictable AI text into something genuinely creative, all on BUZZ HPC.
