Opinion: API distillation is merely a stepping stone for RL; GLM 5.2's autonomous iteration can completely eliminate dependence on American models.

According to Beating's monitoring, Google TPU software engineer Patrick Toulme pointed out that there is a misunderstanding regarding the claim that GLM 5.2 caught up with Opus through distillation. The training difficulty of large models in agent coding tasks lies in the "zero gradient dilemma," meaning that if the model cannot generate the correct running path in the early stages, reinforcement learning cannot obtain gradient signals to initiate parameter updates. The role of distillation Claude or GPT-5.5 is merely to provide seed solutions to bypass the zero gradient dilemma during the cold start phase. Once the model crosses the cold start threshold, subsequent performance improvements will no longer depend on distillation, but will rely entirely on the hill-climbing algorithm of reinforcement learning for self-evolution. Toulme emphasized that GLM 5.2 already has the ability to independently generate successful paths and can completely iterate to higher levels through reinforcement learning, completely eliminating its dependence on large US models. Redis founder Salvatore Sanfilippo added another possible path: while introducing inference patterns (distillation) through high-capacity models is very useful for obtaining better RL signals, the DeepSeek R0 practice has proven that reinforcement learning can still operate autonomously and achieve breakthroughs even in a pure cold start without distillation seeding. He also believes that if the cold start hurdle still needs to be overcome, large-scale model development can initially use local open-source models such as DeepSeek-v3.2 for fine-tuning, rather than necessarily relying on US APIs.