In a recent interview, Jerry Tworek pointed out that with the advent of the "native agent" era, AI itself is rapidly taking over R&D work. He founded Core Automation with the aim of creating a "third-generation AI lab" where AI agents replace humans in performing complex experiments and coding, compressing R&D cycles from months to a single day. He believes that the current bottleneck for further AI development is precisely the Transformer architecture itself. The new generation of labs is utilizing programming agents to explore new architectures, with the goal of achieving new models capable of learning during testing. He also revealed that the breakthrough in OpenAI's o-series inference models was due to large-scale reinforcement learning and the leadership's decisive allocation of hashrate.