gpt mbr ext3_以gpt 3为讲故事者的双头类人形机器人-CSDN博客

gpt mbr ext3

GPT-3 with its huge text corpus and hundreds of billions parameters achieved to surprise us in various forms. Although it is primarily built to achieve NLP tasks such as; text generation, translation, chatbots… We have seen it even generate code, or database queries.

GPT-3具有巨大的文本语料库和数千亿个参数，以各种形式给我们带来惊喜。尽管它主要是为实现NLP任务而构建的；例如，文本生成，翻译，聊天机器人…我们甚至可以看到它生成代码或数据库查询。

These radical applications of GPT-3 shows it immense capability in generating output beyond just text. One could say, language is what differentiates us in the animal kingdom. Language is the mathematics of thought. You can do a lot without it, but you can go well beyond once you have the advantage of language; a carrier of thoughts through stories, arithmetics and gymnastics in thought and many other ways you can traverse and explore ideas merely by skillful use of language as humans do.

GPT-3的这些激进应用显示了它在生成不仅仅是文本的输出方面的巨大能力。可以说，语言是动物界与众不同的地方。语言是思想的数学。没有它，您可以做很多事情，但是一旦有了语言优势，您就可以超越。通过故事，算术和体操中的思想以及许多其他方式来作为思想的载体，您只需像人类一样熟练地使用语言就可以遍历和探索思想。

Let’s leave aggrandizing language aside for now, as I believe most of these are pretty much common sense.

现在让我们暂时不使用强化语言，因为我相信其中大多数都是常识。

If you have been following the news about GPT-3 and its applications, you might have come across a few cases where the GPT-3 generated content was so convincing that even people in tech were fooled (*1, *2). In fact, it was so convincing when one person suggested the article might be generated by GPT-3, they got a warning from moderators to keep their criticism constructive.

如果您一直在关注有关GPT-3及其应用程序的新闻，则可能遇到过几次GPT-3生成的内容令人信服，甚至使技术人员都被蒙骗的情况(* 1，* 2)。实际上，当有人暗示这篇文章可能是由GPT-3撰写时，这真令人信服，主持人警告他们要保持批评性建设性。

What if we applied this convincing ability in storytelling in an humanoid robot?

如果我们将这种令人信服的功能应用于类人机器人的讲故事中怎么办？

According to some scientists, this, bicameral mind, is the very origin of consciousness. We have even seen this theory mentioned in the popular TV series The WestWorld:

根据一些科学家的说法，这种双头思想是意识的起源。我们甚至已经看到了受欢迎的电视连续剧《西方世界》中提到的这一理论：

Bicameralism[Note 1] (the condition of being divided into “two-chambers”) is a hypothesis in psychology that argues that the human mind once operated in a state in which cognitive functions were divided between one part of the brain which appears to be “speaking”, and a second part which listens and obeys — a bicameral mind. The term was coined by Julian Jaynes, who presented the idea in his 1976 book The Origin of Consciousness in the Breakdown of the Bicameral Mind,[1] wherein he made the case that a bicameral mentality was the normal and ubiquitous state of the human mind as recently as 3,000 years ago, near the end of the Mediterranean bronze age.

双房事论 [注1] (被划分为“两个腔室”的条件)是 心理学 中的 一种假设 ，认为人的 大脑 曾经处于一种状态，在该状态下，认知功能被划分为大脑的一部分，这似乎是“说话”，以及听和听的第二部分-双重意识。 这个名词是 朱利安·杰恩斯 ( Julian Jaynes )创造的 ，他在1976年出版的《 两院制精神 的崩溃中的意识起源》一书中提出了这个想法， [1] 他提出了两院制的思想是人类思想的正常和普遍存在的情况。距今大约3000年前， 地中海青铜器时代 即将结束 。

The humanoid robots of our day achieve great tactility, mobility and convincing mimicry of human actions mechanically. Yet we have not seen any robot act in a convincing coherent manner in terms of the set of actions it take. They may respond to individual actions. Some with symbolic AI ingrained into their system can follow conversation to a certain point. But none yet to have created a line of interaction which would convince you to believe it might be the work of a “human”, as we have seen from the reactions to GPT-3 generated content.

当今的类人机器人实现了出色的触感，机动性，并令人信服地模仿了人类的机械行为。然而，我们还没有看到任何机器人就其所采取的一系列行动以令人信服的连贯方式行事。他们可能会对个人行为做出回应。有些将符号AI根深蒂固地融入其系统中的人可以将对话进行到一定程度。但是，正如我们从对GPT-3生成的内容的React中所看到的那样，还没有一种互动方式能够说服您相信这可能是“人类”的工作。

In its simplest form, I believe it could work by taking all the inputs from sensors, cameras. Break it down into a text interpretation of the surrounding world. Once the surroundings and the world is interpreted enough, place the “I”. “I”, as the subject, the storyteller, the robot itself.

以最简单的形式，我相信它可以通过获取传感器，摄像机的所有输入来工作。将其分解为对周围世界的文字解释。一旦对周围环境和世界有了足够的了解，请放置“ I”。 “我”作为主题，讲故事的人，机器人本身。

Let’s observe simple example on how it would work;

让我们观察一下它如何工作的简单示例。

The robot sits by a desk, observes the surroundings.
机器人坐在桌子旁，观察周围的环境。
There is a TV
有电视
There is a basket of fruits
一篮子水果
There are 2 humans sitting across.
有2个人坐在对面。

In this scenario, the input into GPT-3 would be:

在这种情况下，GPT-3的输入为：

“I was sitting by the desk, looking at TV and a basket of fruits. The other 2 people were watching the TV”.

“我坐在办公桌旁，看着电视和一篮水果。其他2个人正在看电视”。

Let’s assume, GPT-3 generates a text such as:

假设，GPT-3生成如下文本：

I picked up a fruit from the basket and offered to those two people.

我从篮子里摘了一个水果，献给了那两个人。

Here, already we achieved something I have not observed in any other AI bot. All chatbots, or humanoid robots I have seen usually expect a conversation starter from a human. Using sensory input & GPT-3 as a story builder, we can already start seeing robots acting on their own “will”.

在这里，我们已经实现了我在任何其他AI机器人中都没有观察到的东西。我见过的所有聊天机器人或类人机器人通常都希望人们进行对话。使用感官输入和GPT-3作为故事的构建者，我们已经可以开始看到机器人按照自己的“意愿”行事了。

Besides the immediate interaction benefits we may achieve from such an application, I believe there will be great benefits in further collecting training data from such an humanoid robot. Once we reach to the point where humanoid robots are as convincing as the text content. We might be ready to surpass and reach to a point where we have GPT-like convincing general AI for multi-sensory input including audio, video, touch.

除了我们可以从这样的应用程序中获得的即时交互收益之外，我相信进一步从此类人形机器人中收集训练数据将带来巨大的收益。一旦到达人形机器人与文字内容一样令人信服的地步。我们可能已经准备好超越并达到一个类似于GPT的令人信服的通用AI，以进行包括音频，视频，触摸的多传感器输入。

In a way, language is might be the kickstarter to get the AI into the real world as a convincing agent. That will collect similar, if not more, data as the GPT-3.

在某种程度上，语言可能是使AI作为令人信服的代理人进入现实世界的启动器。这将收集与GPT-3类似的数据(如果不是更多的话)。

*1 : https://liamp.substack.com/p/my-gpt-3-blog-got-26-thousand-visitors*2 : https://maraoz.com/2020/07/18/openai-gpt3/

* 1： https : //liamp.substack.com/p/my-gpt-3-blog-got-26-thousand-visitors * 2： https : //maraoz.com/2020/07/18/openai-gpt3 /