Claude's anti distillation or ineffective: Small models can directly fit into the thinking chain of large models According to Beating monitoring, large model manufacturers hide the complete thought chain and only provide users with a summary to prevent their abilities from being distilled. But the latest research has found that this layer of protection may not be as strong as imagined: handing over Opus' encrypted thought chain to a weaker Haiku from the same factory, and then jailbreaking Haiku, can allow it to directly repeat Opus' original reasoning. The researchers have been successful on Claude, GPT, and Gemini. This matter has actually been laying the groundwork for several months. In May, cryptographer Matthew Green discovered that these encrypted thought chains could be replayed across sessions and accounts, but at the time, a stable way to read the content had not yet been found. The new paper has added the last step: you don't need to crack the encryption algorithm, just let the manufacturer's own model help you 'decrypt'. As early as March, John X. Morris, co-founder and research leader of Engram, and others demonstrated another path: even if the original thought chain is completely unavailable and only the answer and reasoning summary are viewed, a detailed set of synthetic reasoning can be deduced and used to train small models. This also makes the previously circulated statement that "domestic AI companies have already found a way to obtain Claude and Codex's original reasoning and then use it to train models" even more technically plausible. But this still only proves the feasibility of the method, and cannot prove that any domestic AI laboratory has actually done so. The paper also tested Kimi K3. Just add a few Opus inference tokens to it, and the inference behind K3 will clearly align with Opus; The original reasoning fragments of some Claude and GPT were extracted from Kimi K3, which is about 6 orders of magnitude easier than other models, but the author explicitly stated that distillation cannot be proven based on this. [Original link]