Artificial Intelligence

Artificial Intelligence

数据的种类：

• Features 特征

  输入的数据

• Labels 标签

  供学习使用，相关的输出

学习的种类

• 有监督学习 Supervised learning

  learn from a given sample data set of features (input) and labels (output) values

  从一堆Features和Labels里面学习得到结果

• 无监督学习 Unsupervised learning

  learn from a given sample data set of features (input) values only, often with the goal of finding structure in the data

  只有一堆Features，通常是为了学习得到数据的一些特征

• 增强学习 Reinforcement learning

  learn from trial and error by receiving a reward for taking an action

  从不断的尝试中获取奖励

任务的种类

• Estimation

  由浮点数表示的估计值

• Classification

  有限的类别，由整数表示

方法的种类：

• Artificial Intelligence

  AI 包含了所有种类的学习算法

• Machine Learning

  ML是AI的子集，是一种根据某一个成功标准来获取数据集的关系、结构等的方法，如MSE等。

• Deep Learning

  DL是ML的子集，包含所有使用神经网络的算法。Deep代表神经网络有超过一层隐藏层。

Super Intelligence

The fact that there are many paths that lead to superintelligence should increase our confidence that we will eventually get there. If one path turns out to be blocked, we can still progress.

Nick Bostrom (2014)

Within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended.

Vinge (1993)

硬件的重要性

Version	Year	Elo ratinga	Hardware	Power consumption [TDP]
AlphaGo Fan	2015	>3,000	176 GPUs	>40,000
AlphaGo Lee	2016	>3,500	48 TPUs	10,000+
AlphaGo Master	2016	>4,500	4 TPUs	<2,000
AlphaGo Zero	2017	>5,000	4 TPUs	<2,000

没有硬件的发展，AI也无法取得现有的成就。

硬件的几个维度：

• 性能
• 成本
• 功耗

智能的形态

• Artificial Narrow Intelligence (ANI)

  This specifies an AI agent that exceeds human-expert-level capabilities and skills in a narrow field.

  在某一个单独的领域超过人类专家的能力，比如AlphaZero

• Artificial General Intelligence (AGI)

  This specifies an AI agent that reaches human-level intelligence in any field

  在任何方面都打到人类的智能水平

• Superintelligence (SI)

  This specifies an intellect or AI agent that exceeds human-level intelligence in any respect.

  在任何方面都超过人类的智能水平

生命的版本 (Versions of life)

• Life 1.0 (biological):

  Life-forms with basically fixed hardware (biological bodies) and software (genes).

  固定的硬件和软件，进化速度很慢。

• Life 2.0 (cultural):

  Life-forms with basically fixed and slowly evolving hardware but mostly designed and learned software (genes plus language, knowledge, skills, etc.).

  硬件固定，但软件发展迅速

• Life 3.0 (technological)

  Life-forms with designed and adjustable hardware and fully learned and evolved software.

  由硬件、软件、AI算法组成的Superintelligence

脑机混合系统 Brain-Machine Hybrids

All in all, the brain-machine hybrid seems practically feasible and likely to surpass human intelligence significantly. However, whether it will lead to superintelligence is not obvious

脑机混合系统能增强人的智能，但不一定会通向超智能

全脑模拟 Whole Brain Emulation

当这个技术成熟后，人脑可以在更有效率的硬件上运行，而不再需要在人体中运行了。

The more efficiently it can be provided, the better. There is no need to mimic nature.

仿生学只是一种手段。如果能做到更有效率，就不需要仿生

智能爆炸 Intelligence Explosion

once the technological singularity is reached, there will be an explosion in intelligence

当技术奇点已经到达后，智能水平将爆炸式发展。

目标

• Self-preservation

  maybe harmful to humans, to ensure its survival.

  保持生存。为了达到这个目的，可能会伤害人类。

• Goal-content integrity

  Preserve its current main goal

  保留其当前的主要目标

• Cognitive Enhancement

  提升能力来更好地服务于目标。

• Technological perfection

  使用更好的硬件和软件

• Resource Acquisition

  可能会采取一些出格的方式获取资源。

如何限制AI？

Boxing

Without being connected to the outside world, such as to stock trading platforms, the AI agent has no chance of achieving its goal.

信息闭塞等

Incentives

激励想要得到的行为，惩罚不希望得到的行为

Stunting

功能限制，如硬件限制

Tripwires

报警系统、监控

超智能出现的情况

Singleton

某一种超智能垄断

Multipolar

多点开花，互相竞争、牵制

Atomic

从技术奇点之后，大量的超智能涌现