Sara Hooker 在 PyTorchCon 2026 谈规模化之外的路线：自适应智能
Sara Hooker 在 PyTorchCon 拆解了为什么堆预训练规模不是唯一出路，还用 AutoScientist 案例演示自适应计算怎么落地，做模型的都该看看。
Sara Hooker（adaption_ai）在 PyTorchCon 2026 主题演讲中讨论了单纯扩大预训练规模的局限。她提出的替代方向包括 post-training、test-time scaling、自适应计算与软硬件协同设计。演讲中展示了 AutoScientist 的三个案例：智能体框架优化、智能体记忆和数据集构建。她的核心主张是让 AI 更易适配和定制，使更多人能构建并拥有自己的前沿模型。
At #PyTorchCon 2026, Sara Hooker @sarahookr (@adaption_ai) examines the limits of scaling larger models and the shift toward adaptive intelligence.
Sara discusses post-training, test-time scaling, adaptive compute, and hardware co-design as alternatives to simply scaling pretraining. She highlights AutoScientist examples spanning agent harness optimization, agent memory, and dataset creation.
Her keynote explores how making AI easier to adapt and customize could enable more people to build and own frontier AI.
2026 prediction? "You will build and own frontier AI"