GPT-4 once gave top-5 token probs with no hidden CoT — researcher says it could've been cloned for under $10M
carrigmat · x · 2026-10-23
Researcher carrigmat points out an irony given today's stringent anti-distillation measures: GPT-4 used to expose top-5 token probabilities with no chain-of-thought to hide, and he estimates you could probably have recovered the whole model for under $10M.
A striking observation on how quickly the model-weight protection landscape has shifted, from freely available logprobs to aggressive distillation defenses.
More from Models
- Indie dev open-sources an 11M sparse MoE SLM pretrained on 5.7B tokens for fine-tuning — Nameis19letterslong · 2026-10-23
- Gemini 4's animated SVG generation praised as exceptionally strong — DynamicWebPaige · 2026-10-23
- Prediction: OpenAI likely to ship a new model within its 28-day window — haider1 · 2026-10-23
- GPT Image Stealth Models Slammed for Low-Taste Output; Users Want Finetuning — flowersslop · 2026-10-23
- Recursive open-sources Flint, a fast MoE coding model auto-built by its Eureka Machine in 4 months — Sam Witteveen · 2026-10-23
- ChatGPT suddenly 'misunderstands' simple prompts — distillation guardrails suspected — TerminalNoop · 2026-10-23