Yann LeCun

LeCun 被报道称批评 Amodei 妄想，认为 AI 实验室并不足够了解网络安全，不足以支撑其末日警告。

杨立昆在《财富》杂志采访中称 Anthropic CEO Dario Amodei“彻底妄想”且“疯狂”，认为 AI 实验室对网络安全的理解不足以支撑其末日式警告。

原文 (en)

Yann LeCun called Anthropic CEO Dario Amodei "completely deluded" and "crazy" in a Fortune interview, arguing AI labs don't understand cybersecurity well enough to justify their doomsday warnings.

LeCun 指出 AI 领域 70 年来一直误以为机器能完成复杂智力任务就比人类更聪明，而在许多任务上超越人类并不意味着机器已接近真正的智能。

AI 领域的人 70 年来一直犯同样的错误：只要机器能完成人类认为智力上复杂的任务，就认为它比人类更聪明。这完全是错的。我们的计算机在许多任务上已经超越人类，但距离真正的智能机器仍然很远。

原文 (en)

People in AI have made the same mistake for 70 years: if a machine can do a task humans see as intellectually complex, it must be smarter than humans. This is completely false We have computers that are superhuman at many tasks but we're still far from machines

LeCun 与 Balestriero 在论文中认为，构建柏拉图式表征的关键在于优质的正向视图，而非仅靠规模。

因此，构建柏拉图式表征的杠杆是正向视图，而不仅仅是规模。

原文 (en)

positive views, and not scale alone, are thus the lever to build platonic representations.

论文将下一词预测重新定义为视图较差的 JEPA，暗示其获得对齐表征的效率较低。

下一个词预测就是一种视图很差的 JEPA。

原文 (en)

next-token prediction is a JEPA with bad views.

LeCun 认为特斯拉 Rocket Roadster 飞跃展示只是营销噱头，估算冷气推进的有效推力仅能维持约 5 秒、不到 10 秒。

人工智能科学家 Yann LeCun 称这次展示是营销噱头，并对这种设计的实用性表示怀疑。他认为，即使压缩空气占车辆质量的一半，有效推力持续时间也不到 10 秒，更可能只有约 5 秒。

原文 (en)

Artificial intelligence scientist Yann LeCun called the display a marketing stunt and expressed doubt about the practicality of such a design. He suggested that even if compressed air accounted for half of the car's mass, useful thrust would last less than 10 seconds, and more likely about five seconds.

LeCun 通过计算指出，特斯拉飞行版 Roadster 的压缩空气推进器只能提供不到 10 秒推力，暗示演示车只是轻量外壳而非实用汽车。

他在 X 上发帖估算，即便压缩空气占车辆重量的一半，也只能提供不到 10 秒的有效推力，可能更接近 5 秒。他因此暗示，演示车辆可能只是装载大型压缩空气罐的轻量化外壳，而不是一辆实用的载客汽车。

原文 (en)

In a post on X, he estimated that even if compressed air accounted for half the vehicle's weight, it would provide less than 10 seconds of useful thrust, possibly closer to five seconds. He suggested the demonstration vehicle might therefore be a lightweight shell carrying large compressed-air tanks rather than a practical passenger car.

勒昆称超级智能尚远，并预计突破将来自理解物理世界的世界模型而非大语言模型。

据 French Tech Journal 报道，他在2月的印度 AI 峰会上表示超级智能还很遥远。勒昆预计决定性突破将来自理解物理世界的所谓世界模型。

原文 (en)

Superintelligence is not close, he said in February at the A.I. summit in India, according to French Tech Journal. Mr. LeCun expects the decisive breakthrough to come from so-called world models that understand the physical world.