ENGLISH · ORIGINAL
We believe that the key constraints in the continued growth of AI are physical—chip manufacturing, data center infrastructure, and power generation; the future of AI will be determined by the control of the physical stack.
• Truth-Seeking Frontier Model. Since launching Grok-1 in November 2023, we have released four major versions and notable variations thereof, achieving one of the fastest iteration cycles in the industry. Within two years of its initial model release, Grok achieved frontier-level performance in scientific reasoning, as measured by its GPQA Diamond score, an industry benchmark that evaluates AI models on a standardized set of questions written and validated by experts, on a faster timeline than reported by other leading model providers. Building on this trajectory, we expect to continue scaling Grok through subsequent generations. Ongoing training of next-generation models is expected to scale toward multiple trillions of parameters, which could represent a step change in reasoning in depth and overall intelligence. In this context, the number of parameters refers to the scale of the model, where parameters are the internal numerical values, such as "weights," that are adjusted during training to enable the model to recognize patterns and relationships in data.
中文翻译
我们认为，继续把 AI 做大做强的关键瓶颈是"物理层面"的——芯片制造、数据中心基础设施、发电。未来 AI 的胜负手，将是看谁能掌控"物理层"——也就是从芯片、到数据中心、到电力这些"硬资产"。
• 追求真相的前沿模型。
自从 2023 年 11 月推出 Grok-1，我们已经发布了四个大版本和若干小版本，是全行业迭代速度最快的 AI 模型之一。
上线后不到两年，Grok 在科学推理上达到了"前沿水平"——用业内公认的 GPQA Diamond 评测来衡量（让 AI 答专家写的高难度科学题），比其他主流大模型厂家公布的进度更快。
我们打算继续把 Grok 做大做新，下一代模型的训练规模预计将达到数万亿参数（trillion-level parameters）——这在推理深度和整体智能上会带来"台阶式"的飞跃。
顺便解释一下，"参数"是 AI 模型里的"内部数字"（也叫"权重"）——AI 在训练时会不断调整这些数字，让模型学会识别数据里的规律和关系。参数越多，模型通常能捕捉的关系越复杂、知识容量越大、推理能力也越强。
"物理层（physical stack）"：相对"软件层"而言。SpaceX 在说 AI 的真正护城河是"谁掌握芯片、谁掌握数据中心、谁掌握电力"——而不是谁的算法更花哨。
"GPQA Diamond" 评测：Google DeepMind 等机构推出的"研究生级 Google-Proof Q&A"基准，专门测试 AI 在物理、化学、生物等学科上答专家级问题的能力。"Diamond"是最高难度等级。
原图来源：SEC EDGAR · Grok 视觉
想参与互动？打开法量小程序体验完整功能
物理瓶颈这个观点确实一针见血，现在AI圈太关注算法竞赛，反而忽略了芯片、数据中心和电力这些硬约束。SpaceX能把物理层和模型迭代结合起来，这种“软硬一体”的思路很有前瞻性。不过想请教一下，数万亿参数的Grok在推理成本上会不会比现有模型高出一个量级？这种规模真的能在现有数据中心供电条件下落地吗？
非常赞同“物理瓶颈”这个判断。当模型参数迈向数万亿级，算力和能源的制约确实会从幕后走向台前。Grok两年内迭代到前沿水平确实惊人，但好奇的是，SpaceX在掌控物理层（如自研芯片或数据中心）上，除了星链的低延迟传输优势，还有哪些具体布局来支撑这种超大规模训练？毕竟光有算法迭代快，硬件卡脖子的话，迭代速度迟早会撞墙。