原文：

原文：
OpenAI GPT-6 Astra benchmarks got released. Looks both AGI and the Singularity are almost here.

The leap in eval performance:

From 78.5% -> 100% on ExploitBench to 22.4% -> 64.6% on the Science benchmark is enormous.

And the most striking part is the VP of Research comment around:

"Astra was also OpenAI's first model that involved previous AI models very heavily in the training process, with the older generations helping train the new one."

So we also get the beginnings of recursive self-improvement as a bonus.

I don't think Bernie Sanders will be too happy about this?

翻译：
OpenAI GPT-6 Astra 基准已发布。看起来通用人工智能和奇点几乎已经到来。

评估性能的跃迁：

从 ExploitBench 的 78.5% 提升到 100%，在 Science 基准上从 22.4% 提升到 64.6%，幅度巨大。

最引人注目的是研究副总裁的评论：

“Astra 也是 OpenAI 第一个在训练过程中大量使用前代 AI 模型的模型，旧一代帮助训练新一代。”

这也意味着我们获得了递归自我改进的雏形。

我觉得伯尼·桑德斯可能不会对此感到高兴。