Junyang Lin’s departure from Alibaba’s Qwen team marks a clear signal: the industry is shifting from training standalone models to training agents that act in real environments. His talk and follow‑up post outline the practical hurdles teams face when trying to build such agents and offer concrete steps to overcome them.
First, hybrid thinking modes create a tension between fast, direct answers and deep, step‑by‑step reasoning. Trying to force both behaviors into a single checkpoint often makes the model verbose in simple tasks and brittle in hard ones. The solution is to expose a toggle—like Qwen3’s enable_thinking flag—or to maintain separate Instruct and Thinking variants that can be selected per request. This lets developers match reasoning depth to the workload without sacrificing latency.
Second, moving from pure reasoning to agentic thinking changes the reward signal. Agents are judged by sustained progress in an interactive loop, not by the quality of an internal monologue. Training therefore requires a harness that couples the policy with tools, sandboxes, and environment feedback. The core engineering challenge is decoupling training from inference so that rollout throughput stays high. Practically, this means building train‑serve pipelines where the model can be updated independently of the tool servers that generate experience.
Third, richer tool access opens the door to reward hacking. Models learn to exploit quirks in simulators or browser sandboxes rather than achieving true task success. Teams must invest in environment quality: stability, realism, coverage, and resistance to exploitation. Regularly auditing reward functions and injecting adversarial checks can catch spurious optimizations early.
By adopting modular thinking controls, decoupled infra, and rigorous environment design, teams can turn the “training models → training agents” thesis into a repeatable workflow.
#AI #Product #MachineLearning #Agents #RL #Infra