ABSTRACT

Not many people seem to understand what it is to mimic human intelligence successfully irrespective – that is, irrespective of internal states such as intentions. Successful mimicking will involve embedding in human societies. AI practitioners have concerned themselves with reproducing something that can mimic an individual human brain, failing to notice that if such a brain is to mimic human intelligence it will also have to mimic the process of becoming socialised in human society because crucial features of human intelligence are located in human societies. One notable example of this is natural language, which is continually changing and enormously flexible in ways that cannot be predicted or controlled by individuals. The fluent natural language speaker draws on society's language in the same way that a thermometer draws on the liquid in which it is embedded. The temperature that registers on the thermometer is a property of the liquid, not the thermometer, changing when the liquid warms or cools. In the same way the language spoken by a fluent individual is a property of the society in which he or she is embedded, not a property of the individual; like the thermometer, it changes when the society changes. This means that a Turing Test, based purely on linguistic performance, can be an excellent test of human-like intelligence in a machine. The nature of natural language explains some of the recent stunning successes of deep learning and explains its failures. I also try to explain why the science of AI is so poor at presenting and testing its claims. It aims at convincing the outside world of its success rather than engaging in the kind of assiduous self-criticism engaged in by the physical sciences.

摘要

似乎没有多少人知道成功模仿人类智力的含义是什么，也就是说，不管意图之类的内部状态如何。成功的模仿将涉及嵌入人类社会。人工智能从业者一直致力于复制可以模仿人类大脑的东西，却没有注意到，如果这样的大脑模仿人类智力，那么它也必须模仿在人类社会中社会化的过程，因为人类智力的关键特征是位于人类社会。一个明显的例子是自然语言，它不断变化并且以个人无法预测或控制的方式极大地灵活。会说流利的自然语言的人会吸引社会的注意力 的语言，就像温度计在嵌入温度计的液体上画图一样。温度计上记录的温度是液体的属性，而不是温度计的属性，它在液体加热或冷却时会发生变化。同样，一个流利的人说的语言是他或她所处的社会的财产，而不是该人的财产；像温度计一样，它会随着社会的变化而变化。这意味着仅基于语言性能的图灵测试可以成为对机器中类人智力的出色测试。自然语言的本质解释了深度学习最近取得的一些惊人成就，并解释了其失败之处。我还试图解释为什么AI的科学在陈述和检验其主张方面如此差劲。