Researchers report models suddenly solving their niche problems, suspect training on user inputs
matthew_d_green · x · 2026-09-09
Cryptographer Matthew Green relays that multiple researchers in his field have noticed models getting much better at solving their specific problems over short periods, even when asked in new contexts — several wonder if the models are being trained on their inputs. Green cautions that such anecdotes are hard to pin down: it could be normal model improvements, slightly different prompts, or synthetic data generated from researchers' inputs — but whatever it is, he has noticed it too.
Related event: Researchers Suspect OpenAI Secretly Trains on Their Inputs(13 posts)→
More from Models
- VDiff-Bench: 1,756-question benchmark shows frontier models fail at spot-the-difference — yixin_wan_ · 2026-09-10
- OpenAI team points users to official usage limit update details — athyuttamre · 2026-09-10
- MiniCPM5-2B: 2B-parameter open model runs agents offline on 2GB RAM, tops sub-4B open models — solyarisoftware · 2026-09-10
- OpenAI raises Plus and Pro $100-tier usage limits in new update — athyuttamre · 2026-09-10
- ChatGPT Voice gains GPT-5.6 Sol and GPT-6 Astra, Plus/Pro usage limits raised — athyuttamre · 2026-09-10
- Rumor: DeepSeek V4.1 Flash to launch ~Sept 10, V4 Pro requests rerouted at Flash pricing — solyarisoftware · 2026-09-10