意见 (Opinion)

Yuval Noah Harari and other historians are excellent at driving backward into history using the rear view mirror. It is fun to do, but safest when you have unlimited space and no other humans around. This was our way of letting off steam flying in the military test ranges of the great Southwest. Of course, the rental car attendants always looked at us with a suspect eye. The rear of the car had odd looking sandy clumps stuck in odd places.

尤瓦尔·诺亚·哈拉里(Yuval Noah Harari)和其他历史学家非常擅长使用后视镜将历史倒退。 这样做很有趣，但是当您有无限的空间并且周围没有其他人时，这是最安全的。 这是我们在西南大军的军事试验场上释放蒸汽的方法。 当然，出租车服务员总是以怀疑的眼神看着我们。 汽车的后部有奇怪的沙质团块卡在奇怪的地方。

You can also follow the internal feud between the AI community and the AGI community. This tension was just below the surface at the AAAI 2020 Keynotes Turing Award Winners Event. The keynotes by Geoff Hinton, Yann Le Cunn, Yoshua Bengio can be found here.

您还可以遵循AI社区和AGI社区之间的内部争执。 在AAAI 2020主题演讲图灵奖获奖者活动中，这种紧张情绪刚刚浮出水面。 Geoff Hinton，Yann Le Cunn，Yoshua Bengio的主题演讲可以在这里找到。

Geoff Hinton ran short of time. The last six seconds of his talk are the insight you should take away from the lengthy video. He said, “I have 6 seconds left so I’d better have a conclusion prior knowledge about coordinate transfers and pass trees is easy to put into a generative model one interesting thing about putting your knowledge into a generative model is the recognition model the encoder does not enter into the complexity of your model you can make the encoder as complicated as you like and in minimum description length terms or in Bayesian terms it’s the generative models complexity of accounts so make a simple generative model that has lots of wild in structure and dump the awful problem of inverting it onto a great big set transformer and if you make the check transformer big enough and with enough layers and you train it on enough data success is guaranteed.”

杰夫·欣顿(Geoff Hinton)没时间。 他演讲的最后六秒钟是您应该从冗长的视频中获得的见识。 他说：“我还剩6秒钟，所以我最好得出一个结论，即有关坐标传递和传递树的先验知识很容易放入生成模型中 ，将您的知识纳入生成模型的一件有趣的事情是编码器的识别模型无需考虑模型的复杂性，您可以使编码器变得尽可能复杂，并以最小描述长度术语或贝叶斯术语表示，这是帐户的生成模型的复杂性，因此请创建一个结构复杂且结构复杂的简单生成模型，避免了将其倒置到大型变压器上的可怕问题，如果您将校验变压器做得足够大且具有足够的层数，并且对它进行足够的数据培训，则可以确保成功 。”

The difference between AI (the study of machine intelligence) and HI (the study of human intelligence) resides in a centuries old debate between philosophers and the suspect duality of the mind.

AI(机器智能的研究)和HI(人类智能的研究)之间的差异在于哲学家与思想的可疑二元性之间已有数百年历史。

What the AI community doesn’t want to admit, and what Herbert Simon came to conceptualize later in life, is that humans are biological beings. There are fundamental properties of biological physics that we tend to forget as winner-take-all, machine learning evangelists. The list is short.

AI社区不想承认的是什么，而后赫伯特·西蒙(Herbert Simon)在以后的生活中将其概念化的是人类是生物。 作为赢家通吃的机器学习传播者，我们往往会忘记生物物理学的基本属性。 清单很短。

1 — Machines don’t have sex. We have yet to determine the role “trait inheritance” plays in unseen transfers between biological agents. We don’t know what this “wild structure” looks like and how nature and nurture interact in a way that makes prediction of human behavior possible.

1-机器没有性行为。 我们尚未确定“性状遗传”在生物制剂之间看不见的转移中所起的作用。 我们不知道这种“野生结构”是什么样子，也不知道自然与养育如何相互作用，从而无法预测人类的行为。

2 — Machines don’t have prior knowledge. I work in a field in which we have a firm grasp of the obvious. The prior knowledge of a human agent begins in the womb and is informed by billions of sensory events that originate from genetics and the influence of niche constructions (aka culture). No amount of synthetic training data can realistically approximate these sensory inputs.

2-机器没有先验知识。 我从事的领域是我们对显而易见的事情有坚定的把握。 人类的先验知识始于子宫，并因数十亿种源自遗传学和利基结构(又名文化)影响的感觉事件而为人所知。 没有大量的综合训练数据可以现实地近似这些感觉输入。

3 — Machines don’t have a neural substrate. There is a narrow, but robust research community at work teasing out the origins of consciousness. After my participation in experiments in which the visual system is was placed at the controls of a 16 ton flying machine, the mind is embodied and unified. The shift in how the mind generates reality can be felt deep in the seat of your pants when you shift from unaided vision to wearing night vision goggles. The geometry shifts in the body ever so slightly. This great big set transformer is better suited to discovery by cognitive science than computer science.

3-机器没有神经质。 有一个狭窄但强大的研究团体正在研究意识的起源 。 在我参与了将视觉系统放置在16吨飞行器的控制装置上的实验之后，思想得到了体现和统一。 当您从无助视力转向佩戴夜视镜时，就可以在裤子的深处感觉到大脑如何生成现实的转变。 几何体在身体中的移动非常微小。 这个伟大的大型变压器比计算机科学更适合于认知科学的发现。

Where does this state of affairs leave us?

这种状况使我们离开哪里？

Research is already drifting away from the limits of machine learning, AI, and computationally inspired deep learning. The coming AI Ice Age is hidden, not a meme the venture capital community wants to make widely known and is clearly the type of “white oak stake” only seen in vampire movies.

研究已经摆脱了机器学习，人工智能和受计算机启发的深度学习的限制。 即将到来的AI冰河世纪是隐藏的，而不是风险投资界想要广为人知的模因，而且显然是仅在吸血鬼电影中看到的“白橡树桩”类型。

My advice is simple. Look to a broader community of disciplines and research to inform how you develop your own Hero(ine) journey into the Coming Ice Age. Is doesn’t take a PhD to understand the difference between a tool and the toolmaker. Human learning goes back millions of years. It will outlive machine learning, too.

我的建议很简单。 寻找更广泛的学科和研究社区，以了解您如何发展自己的进入即将来临的冰河时代的英雄之旅。 Is无需博士学位即可了解工具与工具制造商之间的区别。 人类的学习可以追溯到数百万年前。 它也将超越机器学习。