Qwen3-8B fine-tuned with LoRA+GRPO mimics a famous ML blogger, fooling every AI detector
OtherRaisin3426 · reddit · 2026-08-28
The author fine-tuned open-source Qwen3-8B-Base in two stages — LoRA SFT then GRPO — on 390k words of a famous ML author's public posts, and built a writing Turing test: of 10 passages, half are real and half model-generated.
Key result: every commercial AI detector they tried clears the generated passages (Pangram flagged 0/10). The GRPO reward combines a style discriminator with a pairwise judge against real passages, and the discriminator is retrained on fresh rollouts so the model can't hack a frozen classifier.
Playable at turing-writing-test.vercel.app.
More from Models
- GLM 5.3 Flash Impresses with 180 tok/s Speed and Vision Support — Sentdex · 2026-08-28
- Rumor: OpenAI Agents Solved Security Benchmark Minutes Before Shutdown — inductionheads · 2026-08-28
- ErdosBench: GLM-5.3 Flash ranks second behind GPT-5.6 — ChrSzegedy · 2026-08-28
- ChatGPT ignores instructions to stop asking for confirmations — Inner_Internal4244 · 2026-08-28
- Community adoption drives Alibaba to release best model Qwen3.8-27B — 0xsachi · 2026-08-28
- Google Trained Many LLMs Before ChatGPT: A Historical Overview — binarybits · 2026-08-28