OpenAI 预训练损失曲线显示 2026 年加速下降
OpenAI 晒出内部损失曲线,2026 年下降明显加速,还透露未发布模型已低于 GPT-6.2 Astra。Pachocki 给出 0.68-0.7 这个新能力触发点,做应用的开发者可以据此判断换代节奏。
OpenAI 公开的预训练损失曲线显示 2026 年出现陡峭加速,公司将其归因于算力扩张、长期研究投入以及 AI 参与加速自身研发。
一个未发布的内部模型性能已低于 GPT-6.2 Astra 的损失值,OpenAI 表示 AI 本身已参与设计最新模型。
首席科学家 Jakub Pachocki 预计下一批新能力将在损失值 0.68-0.
7 区间出现,并认为预训练研究最早在 2027 年可能复现数学领域刚经历的能力雪崩。这一判断把模型进步与损失曲线的量化节点直接挂钩,为外界提供了观察 OpenAI 研发节奏的参考坐标。
对开发者而言,若预训练在 2027 年再次出现跳变,模型换代周期可能进一步缩短。
OpenAI’s pretraining loss curves show a steep acceleration in 2026, driven by compute, long-term research, and AI speeding up its own development.
An unreleased internal model already sits below GPT-6.2 Astra. OpenAI says AI itself already helped design the newest models.
Chief Scientist Jakub Pachocki expects the next new capabilities around 0.68-0.7 and says pretraining research could see the same avalanche they just saw in math as soon as 2027.
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include OpenAI, GPT-6.2, Jakub Pachocki, pretraining.
Source attribution: curated from kimmonismus (@kimmonismus). Cross-source corroboration: 11 independent sources have covered this development (including lmarena.ai, Greg Brockman, Decoder), confirming industry-wide signal strength.