Simo Ryu highlights how the actual internal reasoning steps generated by models like GLM 5.2 stretch into lengthy sequences packed with odd terminology that humans would struggle to follow, while the tidy step-by-step displays people actually read come from later summarization applied on top.
@cloneofsimo > unlike non-bootstrapped CoT which tends to be completely unreadable
where have you seen this, even GPT 5.5 is pretty readable
Its so funny because their COT is so very much verbose (unlike non-bootstrapped CoT which tends to be completely unreadable) and look *EXACTLY* like claude's interface COT, which is probably summarized ones.
Bootstrapped outputs can mimic the clean style seen in certain interfaces, whereas non-bootstrapped traces stay dense and unfamiliar, prompting side-by-side comparisons with systems such as GPT-5.5.
The discussion leaves open whether summarization happens inside training, during inference, or only at the final UI stage, an uncertainty that affects how much weight to give visible reasoning traces.
No Digg Deeper questions have been answered for this story yet.
@teortaxesTex I dont mean the summarized CoTs, i mean the raw ones. Raw ones are so long and full of weird words, unlike the summarized ones we see on interfaces
@cloneofsimo > unlike non-bootstrapped CoT which tends to be completely unreadable
where have you seen this, even GPT 5.5 is pretty readable
@cloneofsimo I understand you and I have seen raw CoTs they're terse, that's about it. Words are just high-signal
@teortaxesTex I dont mean the summarized CoTs, i mean the raw ones. Raw ones are so long and full of weird words, unlike the summarized ones we see on interfaces