Uploaded June 2026 | Updated September 2026, 2 weeks ago
Harper Carroll fed a thousand examples of her own writing to Claude. The output still registered 100% AI-created according to a detector. After less than half an hour fine-tuning an open source Llama model on the same data, the detector read 100% human. In this clip from her recent discussion with Tim O'Reilly, Harper explains why prompting can't shift the output token distribution the way fine-tuning can and walks through a practical shortcut for building the training dataset: Run your own writing through the model to get the AI-ified version, then fine-tune with that as the input and your original as the target output.
Follow O'Reilly on:
LinkedIn: linkedin.com/company/oreilly
Facebook: facebook.com/OReilly
Instagram: instagram.com/oreillymedia
BlueSky: https://bsky.app/profile/oreilly.bsky.social
TikTok: tiktok.com/@oreillymedia
Harper Carroll fed a thousand examples of her own writing to Claude. The output still registered 100% AI-created according to a detector. After less than half an hour fine-tuning an open source Llama model on the same data, the detector read 100% human. In this clip from her recent discussion with Tim O'Reilly, Harper explains why prompting can't shift the output token distribution the way fine-tuning can and walks through a practical shortcut for building the training dataset: Run your own writing through the model to get the AI-ified version, then fine-tune with that as the input and your original as the target output.
Follow O'Reilly on:
LinkedIn: linkedin.com/company/oreilly
Facebook: facebook.com/OReilly
Instagram: instagram.com/oreillymedia
BlueSky: https://bsky.app/profile/oreilly.bsky.social
TikTok: tiktok.com/@oreillymedia
