Microsoft engineer floats overlooked theory: OpenAI may have used math user sessions as context, not training data
JFPuget · x · 2026-09-09
Amid speculation that OpenAI trained a model on users' math sessions in record time, Microsoft's JF Puget highlights less-discussed technical possibilities: the sessions may have been used as context rather than to modify weights, or humans may have read relevant sessions to guide their own prompting. He stresses he's only discussing what's technically possible, not what happened—pointing out that weight updates are not the only path to improving LLM performance.
More from Models
- Early Astra user: great model underneath, but zero proactivity that's 'a system prompt fix away' — oyacaro · 2026-09-09
- Hands-On With Hunyuan Hy4 Preview: Strong Game Generation and Document Auditing — TencentHunyuan · 2026-09-09
- Tencent Hunyuan Hy4 Preview: 770B-Parameter Open Model With 49B Active and Native 1M Context — TencentHunyuan · 2026-09-09
- Dev: models have stagnated since Opus 4.6, just harder to parse and more task-obsessed — mertdumenci · 2026-09-09
- Dev claims GLM 5.3 Flash runs better locally than via API, suspecting lower-quality quant — MaziyarPanahi · 2026-09-09
- Blogger teases Gemini 'monster model' codenamed Fable Killer — ChrisGPT · 2026-09-09