Show HN: A minimal implementation of LLM output watermarking (github.com)

A new project introduces a minimal implementation of output watermarking for language models, building on Andrej Karpathy's microgpt. The approach employs a secret-keyed sampling method that allows for statistically detectable watermarks without altering the model's original token probabilities or introducing visible markers in the text. This is achieved using Python's standard library, minimizing dependencies and keeping the code concise. This development is significant for the AI/ML community as it addresses concerns about the authenticity and provenance of generated text. The watermark emerges from the choice of tokens during generation, and can be effectively detected using a known key, demonstrating a notable p-value of ~6.8×10⁻⁵³, indicating strong statistical confidence in distinguishing watermarked text from normal output. This technique not only enhances transparency in AI-generated content but also provides a framework for future advancements in responsible AI usage.