Tag
Researchers have, for the first time, systematically extracted hidden chain-of-thought from closed-source models at OpenAI, Anthropic, and Google. They exploited an API encryption compatibility flaw to have weaker models decrypt the content, thereby bypassing CoT distillation protections, and also studied open-source models such as Kimi K3, GLM-5.2, and DeepSeek.
This paper demonstrates a method to extract hidden reasoning traces from proprietary LLM APIs (Anthropic, OpenAI, Google) by replaying encrypted chain-of-thought blocks into weaker, jailbroken sibling models, recovering the stronger model's raw reasoning verbatim without attacking it directly.
The article argues that as open weight models become equally capable to proprietary ones, OpenAI and Anthropic's valuations will plummet, while companies like Fireworks AI will capture value by operating open models as a neutral infrastructure layer.
This paper introduces Multi-Agent Protocol Distillation (MAPD), a framework that distills knowledge from proprietary agentic search models into open-source models using a structured protocol to bridge the distribution gap, achieving state-of-the-art results on seven QA benchmarks.
This study evaluates six proprietary LLMs across 16 DSM-5 conditions using adversarial attacks, finding that safety safeguards are only reliable for suicide and self-harm, with failure rates up to 100% for other conditions like eating disorders and substance use disorder.
The author argues that switching from proprietary AI models to open models is now much less of a professional sacrifice, citing improving open model quality and Claude's ID verification as a catalyst, similar to the historical shift from Windows to Linux.
New benchmarks like DeepSWE reveal a significant performance gap between proprietary and open-source AI models, causing disappointment in the open-source community.
Miles Brundage comments on the lack of quantitative analysis on how distillation affects the capability gap between open-weight and proprietary AI models, referencing a claim by Epoch AI that open-weight models lag by four months.