Andrej Karpathy：AI将取代哪些工作岗位？

分享一个最近关于AI对经济影响的有趣讨论。

Sharing an interesting recent conversation on AI's impact on the economy.

AI常被比作各种历史先例，如电力、工业革命等等。

AI has been compared to various historical precedents, electricity, industrial revolution, etcetera.

我认为最贴切的类比是将AI视为一种新的计算范式，软件亦是如此，因为两者本质上都是关于数字信息处理的自动化。

I think the strongest analogy is that of AI as a new computing paradigm, software too, because both are fundamentally about the automation of digital information processing.

如果要预测计算技术对1980年代就业市场的影响，最具预测性的任务特征就是看其算法固定程度。

If you were to forecast the impact of computing on the job market into one thousand nine hundred and eighty seconds, the most predictive feature of a task job you'd look at is to what extent the algorithm of it is fixed.

也就是说，你是否只是机械地按照既定规则转换信息？

That is to say, are you just mechanically transforming information according to rote, easy to specify rules?

例如打字、簿记、人工计算等等。

For example, typing, bookkeeping, human calculators, etcetera.

在当时，这类程序是那个时代的计算能力允许我们手工编写的。

Back then, this was the class of programs that the computing capability of that era allowed us to write by hand manually.

而现在有了AI，我们能够编写以前根本不敢想象的手工程序。

With AI now, we are able to write new programs that we could never hope to write by hand before.

我们通过设定目标来实现这一点。

We do it by specifying objectives.

例如分类准确率、奖励函数，我们通过梯度下降搜索程序空间，找到能很好实现该目标的神经网络。

For example, classification accuracy, reward functions, and we search the program space via gradient descent to find neural networks that work well against that objective.

这是我早些时候写的软件2.0博文。

This is my software two blog post from a while ago.

在这种新的编程范式中，最具预测性的新特征是可验证性。

In this new programming paradigm then, the new most predictive feature to look at is verifiability.

如果一个任务作业是可验证的，那么它可以直接或通过强化学习进行优化，神经网络可以被训练得极其出色。

If a task job is verifiable, then it is optimizable directly or via reinforcement learning, and a neural net can be trained to work extremely well.

这关乎人工智能能在多大程度上实践某事。

It's about to what extent an AI can practice something.

环境必须是可重置的。

The environment has to be resettable.

你可以开始新的尝试。

You can start a new attempt.

高效。

Efficient.

可以进行大量尝试。

A lot attempts can be made.

并且可奖励。

And rewardable.

存在某种自动化流程来奖励任何已做出的特定尝试。

There is some automated process to reward any specific attempt that was made.

任务作业的可验证性越高，在新编程范式中就越适合自动化。

The more a task job is verifiable, the more amenable it is to automation in the new programming paradigm.

如果不可验证，它就只能依赖神经网络泛化的魔力（祝好运），或通过模仿等较弱的手段实现。

If it is not verifiable, it has to fall out from neural net magic of generalization, fingers crossed, or via weaker means like imitation.

这正是推动大语言模型发展锯齿前沿的动力。

This is what's driving the jagged frontier of progress in LLMs.

可验证的任务进展迅速，甚至可能超越顶级专家的能力，例如数学、代码、观看视频的时间长度，任何看起来像有正确答案的谜题。

Tasks that are verifiable progress rapidly, including possibly beyond the ability of top experts, for example, math, code, amount of time spent watching videos, anything that looks like puzzles with correct answers.

相比之下，许多其他领域在创造性、战略性任务上显得滞后，这些任务需要结合现实世界的知识、状态、情境和常识。

While many others lag by comparison creative, strategic tasks that combine real world knowledge, state, context, and common sense.

软件一能轻松自动化任何你可以明确指定的内容。

Software one easily automates what you can specify.

软件二能轻松自动化任何你可以验证的内容。

Software two easily automates what you can verify.