Applied Compute CEO：后训练如何赢得推理，以及为什么企业必须拥有自己的智能Applied Compute CEO: Post-Training Wins Inference & Why Enterprises Must Own Their Intelligence
The Takeaway：真正的护城河不是调用最强 frontier model，而是用专有数据把 Pareto 曲线整体外推，让同一规模模型在成本、延迟和领域能力上全面优于通用模型。
Applied Compute CEO Josh 从 OpenAI 内部走出来后，把公司定位为新一代 AI hyperscaler。他认为“拥有自己的智能”不是因为担心 lab 恶意抽数据，而是为了部署位置、优化目标、成本结构和领域适配的完全控制权。OpenAI 与 Base10 的合作本身就是多模型未来的明确信号：客户要的是选择权和灵活性。
Josh 的核心观察是：后训练在“分布外”数据上能真正拉开能力差距，而大多数企业其实先用后训练优化价格性能。更远的未来是在线 RL——生产推理数据回流成持续学习信号，把组织内部的判断痕迹固化成模型策略。他直言：“我们现在拥有的其实是一台爬山机，最难的部分是定义要爬的山，所以 eval 既要认真建，也要小心保护。”
公司筛选客户的标准很清晰：要么能力提升极高价值（制药、芯片、网络安全），要么推理规模极大，能把效率收益摊到海量 token 上。训练与推理被一起 co-optimize，因为模型训练方式会直接影响大规模推理部署的拓扑选择。最终愿景是在 GPU 之上构建完整软件栈——训练、推理、路由、安全、沙箱——从最难、最高杠杆的模型层开始向上扩展。
Applied Compute CEO Josh 从 OpenAI 内部走出来后，把公司定位为新一代 AI hyperscaler。他认为“拥有自己的智能”不是因为担心 lab 恶意抽数据，而是为了部署位置、优化目标、成本结构和领域适配的完全控制权。OpenAI 与 Base10 的合作本身就是多模型未来的明确信号：客户要的是选择权和灵活性。
Josh 的核心观察是：后训练在“分布外”数据上能真正拉开能力差距，而大多数企业其实先用后训练优化价格性能。更远的未来是在线 RL——生产推理数据回流成持续学习信号，把组织内部的判断痕迹固化成模型策略。他直言：“我们现在拥有的其实是一台爬山机，最难的部分是定义要爬的山，所以 eval 既要认真建，也要小心保护。”
公司筛选客户的标准很清晰：要么能力提升极高价值（制药、芯片、网络安全），要么推理规模极大，能把效率收益摊到海量 token 上。训练与推理被一起 co-optimize，因为模型训练方式会直接影响大规模推理部署的拓扑选择。最终愿景是在 GPU 之上构建完整软件栈——训练、推理、路由、安全、沙箱——从最难、最高杠杆的模型层开始向上扩展。
The Takeaway: The real moat is not calling the strongest frontier model, but using proprietary data to push the entire Pareto curve outward so the same-size model beats general models on cost, latency and domain capability.
Applied Compute CEO Josh, after years inside OpenAI, positions the company as a new AI hyperscaler. “Owning your intelligence” is less about fearing malicious data theft and more about full control over where models run, what they optimize for, cost structure and domain fit. OpenAI’s Base10 partnership is itself a clear signal of the multi-model future: customers want choice and flexibility.
Josh’s core observation: post-training delivers real capability gains precisely when data is out-of-distribution; most enterprises first use it to optimize price-performance. Looking further, online RL will turn production inference traces into continuous learning signals that codify organizational judgment into model policy. He puts it bluntly: “What we have with RL is we essentially have a hill climbing machine. The hardest part is actually defining the hill to climb, which is why evals… focus on building and… safeguard pretty carefully.”
Customer qualification is strict: either extreme capability value (pharma, chips, cyber) or massive inference volume where efficiency gains compound. Training and inference are co-optimized because the way a model is trained directly shapes the topology of large-scale serving. The long-term vision is a full software stack on top of GPUs—training, inference, routing, security, sandboxing—starting from the hardest, highest-leverage layer: the model itself.
查看原文 →
Applied Compute CEO Josh, after years inside OpenAI, positions the company as a new AI hyperscaler. “Owning your intelligence” is less about fearing malicious data theft and more about full control over where models run, what they optimize for, cost structure and domain fit. OpenAI’s Base10 partnership is itself a clear signal of the multi-model future: customers want choice and flexibility.
Josh’s core observation: post-training delivers real capability gains precisely when data is out-of-distribution; most enterprises first use it to optimize price-performance. Looking further, online RL will turn production inference traces into continuous learning signals that codify organizational judgment into model policy. He puts it bluntly: “What we have with RL is we essentially have a hill climbing machine. The hardest part is actually defining the hill to climb, which is why evals… focus on building and… safeguard pretty carefully.”
Customer qualification is strict: either extreme capability value (pharma, chips, cyber) or massive inference volume where efficiency gains compound. Training and inference are co-optimized because the way a model is trained directly shapes the topology of large-scale serving. The long-term vision is a full software stack on top of GPUs—training, inference, routing, security, sandboxing—starting from the hardest, highest-leverage layer: the model itself.