[Viewpoint: API distillation is only a stepping stone for RL, GLM 5.2 autonomous iteration can completely get rid of the dependence on American models] According to Beating monitoring, Patrick Toulme, a software engineer at Google TPU, pointed out that there is a misunderstanding about the claim that GLM 5.2 is equaling Opus through distillation. The training difficulty of large models in intelligent agent coding tasks lies in the "zero gradient dilemma", which means that if the model cannot generate the correct running path in the early stage, reinforcement learning cannot obtain gradient signals to initiate parameter updates. The role of distilling Claude or GPT-5.5 is only to provide seed solutions during the cold start phase to bypass the zero gradient dilemma. Once the model crosses the cold start threshold, the subsequent performance climb will no longer rely on distillation, but entirely rely on reinforcement learning's hill-climbing algorithm for self evolution. Toulme emphasized that GLM 5.2 already has the ability to independently generate successful paths and can autonomously iterate to higher levels through reinforcement learning, completely freeing itself from dependence on American large-scale models. Redis founder Salvatore Sanfilippo added the possibility of another path: although introducing inference patterns (distillation) through high-performance models is very useful for obtaining better RL signals, the practice of DeepSeek R0 has proven that even in pure cold start situations without distillation seeding, reinforcement learning can still operate autonomously and make breakthroughs. At the same time, he believes that if the cold start threshold still needs to be crossed, the development of large models can be initially fine tuned using local open-source models such as DeepSeeker v3.2, rather than relying solely on American APIs. [Original link]