Fine-tuned Qwen3-8B on a single RTX 4090 produces text that scores 100% human on pangram
StewartalsopIII · x · 2026-09-05
Dev NavinFS reports generating text that scores 100% human on the pangram AI detector. The recipe: fine-tune Qwen3-8B-Base (upgraded from 4B) on 4,000 articles instead of 500 — all doable on a single RTX 4090.
The takeaway: consumer-grade fine-tuning of small models can now fully evade mainstream AI-text detectors, a concrete stress test of their reliability.
More from Models
- Gemini somehow can't read its own Personal Instructions, user's toggles don't help — Connect_Reporter_799 · 2026-10-23
- Diffusion LLM bias rankings shift with sampling configs, persist after debiasing — Bollegala · 2026-10-23
- Running ~100 DeepSeek v4.1 flash agents in parallel for under $20/month — zack_overflow · 2026-10-23
- Researcher laments open-source audio models are surprisingly subpar, even representation ones — gowthami_s · 2026-10-23
- User claims they generated video with "Claude Opus 5.5", unannounced model name surfaces — jwt0625 · 2026-10-23
- Security researchers push back as OpenAI restricts Astra access; BlackHat 2027 mocked as 'cyber model' bazaar — cyb3rops · 2026-10-23