In the AI era, what is becoming increasingly valuable? 11 insights on AI

Mr. Yang says that tokens are becoming less and less valuable, while context is growing more precious; he also points out that in the AI era, speaking through your work means speaking through "models". What exactly makes the AI era so new? How should we respond? What do we need to build with our own hands? And what kind of person do we want to become? Let's take a look at Mr. Yang's 11 insights~

I

1

Many people have not realized what exactly is new about the AI era.

Every major era transition brings a whole new category of works for individuals, which then gives rise to countless new professions around this category. For example, the industrial products produced by the production assembly lines in the electricity age are a new category of works. Similarly, in the Internet era, this category is "webapp"; in the mobile Internet era, it is the "app" category.

Then in the AI era, you can ignore all media reports. The most critical point is that "model" has become a new major category of works.

Today's "model practitioners" are still scarce, high-cost and vulnerable, but these are all characteristics of the early stage of an era, which is also the period with the most lucrative returns.

Therefore, training your own "model" from scratch is a unique creative act. The model you train is essentially the same as apps, webapps, books and industrial products: it is your work. In the AI era, speaking through your work means speaking through "models".

This is the first trend brought by the huge change of the era: a brand new category of works has emerged.

What is the second trend? The answer is that the relatively typical works of the previous era will become extremely abundant in the next era. For example, after the Industrial Revolution, agricultural products became extremely abundant; after the Information Revolution / Internet Revolution / Mobile Internet Revolution, industrial products became extremely abundant.

Obviously, following this trend, in the AI era, apps, webapps, software and information will become extremely abundant.

This is almost a subversive change. However, human beings are not good at coping with drastic changes, let alone imagining the future.

In addition, the buffer time left for human beings in this round of drastic changes is much shorter than previous rounds. The Industrial Revolution and the Information Revolution both started with a buffer period of at least 30 years, while this AI revolution only starts with a 3-year buffer period. The social structure, legislative system and education inheritance are all not prepared for it.

Where will human beings go? All the answers we give today are destined to be wrong. A foreseeable future is not a real future. But some things are more eternal, such as love and trust.

2

Nowadays, code is getting cheaper and cheaper, and tokens are getting cheaper and cheaper. Then what is becoming more and more expensive?

The answer is context. Several obvious trends: tokens are becoming less valuable while context is growing more precious; coding is becoming less valuable while systems are growing more precious; software is becoming less valuable while communities are growing more precious; large teams are becoming less valuable while elite small teams are growing more precious.

3

The title of my keynote speech at the beginning of this year is Intentional Revolution: Towards the Age of Hybrid Intelligence.

In fact, this round of AI transition is still a progress that complements human capabilities. Novices can more easily supplement the knowledge points that experienced people master; people who master knowledge points in one field can more easily supplement knowledge points in another field; people who know one programming language can more easily supplement knowledge points of another programming language, and so on.

When specific knowledge points become less important and turn into a kind of "instant retrieval" similar to search engines, the structure of the human brain will inevitably change. This is the experiment done by cognitive scientist Daniel Wegner and other researchers: the Google Effect, which means the memory of human brain has become distributed.

When the history at the birth of search engines repeats itself today, what will happen to our brains? What revolution will take place in the way human beings learn?

II

4

The Industrial Revolution liberated human beings from physical labor, but we invented fitness and sports, and actively do high-quality physical labor to keep healthy.

The same goes for mental labor. AI liberates us from heavy mental labor, which does not mean we should give up mental labor, but turn to high-quality mental labor.

Compared with high-quality labor, low-quality labor has a major feature: it is often forced, people have to farm and do hard physical work just to make a living. But high-quality labor is spontaneous.

In the AI era, you may wish to think more about how to liberate yourself from low-quality labor, so as to free up more time, energy, capital and leisure to engage in high-quality labor.

5

In the new AI era, I have a suggestion: don't spend all your time testing various AI software and improving your personal workflow, but still stick to the creation-oriented principle. That is, high-quality works are the core.

If we always stay on these trivial matters, it will be difficult for us to find our own anchor, and we will still feel lost in the great changes of the era.

But in the process of creating works, we will find a large number of loopholes and unsolved problems in cutting-edge AI laboratories, so that one work gives birth to another, and one idea spawns another. This is the huge difference between "creator orientation" and "consumer orientation".

III

6

Several "common senses" that may be "counterintuitive":

Large AI models still have many black boxes: even AI scientists in Silicon Valley still feel that there are a lot of incredible and incomprehensible parts in this round of AI transition, and they are also in the process of rapid learning.

Understanding the black box of large AI models does not rely on empty talk, but on hands-on practice. Up to now, many answers given by the whole world about why this round of AI transition can achieve such good results are still wrong.

This round of AI revolution is still in the very early stage: it makes us realize another idea to unlock "large language models" — hybrid intelligence. In the past, AI did its own job and humans did their own work, but now the two sides are deeply involved with each other.

Countless new social problems are emerging, but more new opportunities are also emerging: as I said in my keynote speech at the beginning of the year, the more developed civilization is, the more resources and power tend to be concentrated in the hands of a few people and a few organizations. In the primitive era, a person could live alone away from the tribe; in the Warring States Period, an assassin still had a chance to assassinate the King of Qin, try that today?

The more complex the social system is, the more restricted individual actions are, but the mechanism that restricts power is more likely to have structural loopholes. Eventually, violence and key resources are controlled by a few people, which is a pessimistic side. But at the same time, we are also ushering in the golden age of human intelligence: many key inventions that were out of reach in the past are becoming achievable. They may take us to space, extend our lifespan, and allow us to have a healthier lifestyle. This is the optimistic side.

7

Last night, I was supposed to go to bed early. But before going to bed, I accidentally assigned a complex task to AI, and I didn't have much hope that it could solve it quickly.

Unexpectedly, it started its real performance. It reasoned for more than 30 steps in one go, and finally solved this extremely complex technical problem. I stared at it dumbfounded like a fool, and completely lost my sleepiness... The price was that I only slept for two hours last night.

What is this complex problem? It worked fine when debugging locally, but it would definitely go wrong once deployed to the production environment. The key to solving the problem was hidden extremely deep, and a series of issues such as signature, encryption and cryptography needed to be checked and excluded first to find the real cause.

According to my past experience, even the most senior engineers may not be able to solve this kind of problem in three months. In fact, this technical problem has existed in the open source community for many years, and no one has ever solved it.

But AI actually solved this problem in three hours, carrying out extremely complex reasoning, and I did not do any manual intervention in the reasoning process of more than 30 steps. During the process, I wanted to intervene several times, but it turned out later that I was the one who misunderstood.

Of course, I also contributed to it. It has something to do with my good thinking habits. Before it started complex reasoning, I downloaded all relevant open source projects and technical documents to the local for its reference. Facts have proved that this step played a big role, and the key clues all came from the batch of documents I provided.

I also made another contribution: I built an automated verification pipeline for it. It has high permissions, and can immediately get the real log file after each run.

8

Last night, a good friend came to Beijing on a business trip, he invited me to have dinner together, and we had a great chat. I'd like to share some insights with you.

(1) At present, layoffs in technology companies have become very common. My friend's business trip this time is to handle a rather difficult matter: layoffs. In the midst of the huge AI changes, even large factories are no exception. There is no safe zone. People who are easily laid off are those who are not aware of the huge AI changes, and do not take the initiative to learn and adapt to them.

(2) New R&D structures and R&D methodologies are being formed. The team organizational structure is increasingly leaning towards small elite teams. The leader of this elite team had better be proficient in both technology and product. The smaller and more elite the team is, the higher the efficiency will be. In the AI era, more people are a burden instead of an advantage. You can't explain to people "why we have to start all over again", but AI doesn't need your explanation, and will help you verify immediately.

(3) The working method has evolved from "question and answer" to "AI works autonomously". I introduced to my friend that I have built multiple sets of AI autonomous working systems. When we were having dinner, AI was processing four large projects in parallel, and I only needed to check and accept the results afterwards.

(4) You must train and deploy models by yourself, not to save money, but to truly understand the technical boundaries. If you don't run GPU by yourself, you will never understand many things that you should understand. Using APIs seems to save trouble, but you will always be vague about what models can do, what they can't do, and where the boundaries are.

IV

9

Today I discussed a trend with engineers: in the future, the requirements for engineering capabilities will gradually transition to complex architectures. To distinguish good engineers from mediocre ones, their ability to manage complexity is very critical. A good engineer can greatly simplify complex projects.

At the same time, the requirements for engineers' innovation capabilities are higher. Good engineers are getting closer to researchers in the traditional sense; conversely, researchers also need stronger engineering capabilities.

In addition, the requirements for engineers' empathy ability have also increased. They need to empathize with users, Agents, communities and third parties. That is to say, a good engineer also needs to be like a consultant.

In the future, the most promising skill combination package is: Engineer + Researcher + Consultant.

10

In the AI era, what kind of people do senior practitioners want to work with?

Hacker spirit: Curiosity, internal drive, diligence, the kind of person who won't go to bed until they figure out the difficult point. Of course, not sleeping is bad for health, this is just a figurative expression here.

Good character: Gentle in temper, not intense in personality. In the AI era, if you want to get good results, you need great permissions within the organization, and once these permissions are handed over, it is difficult to take them back. Therefore, character is becoming more and more important.

Good rational thinking and certain engineering literacy: In short, they have good logical reasoning ability, and understand some common sense of software engineering.

Good at communication: AI coding is largely communicating with AI. If you don't like communicating with people in daily life, you will most likely still find it difficult to communicate well with AI.

11

Many parents are overly anxious about their children's academic performance. I never thought before that many first-grade classmates in the primary school where my little girl studies have already started taking remedial classes. They are still in remedial classes at seven or eight o'clock in the evening. This situation is everywhere.

But I insist on not letting my little girl take any remedial classes. I even joked with her mother today that the school is full of formalities now, and I can buy a writing robot for my little girl to help her with her homework.

The logic behind this is as follows: at present, higher education is already relatively backward, and the premium of college students has almost disappeared, except for popular majors in the top dozen universities in China.

With the rapid development of AI and the emergence of a large number of new knowledge, opportunities are increasingly not in the traditional higher education stage.

In the AI era, what is increasingly needed is alien thinking, that is, the part of you that is different from the average value. The part of the average value is too easy to be trained into large AI models.

Then how to cultivate children's alien thinking? It is to let children form their own unique selves.

Unique interests.

Unique identity recognition.

Unique knowledge structure.

Unique interpersonal network.

And finally naturally comes: unique works.

These are the things that are more important than taking remedial classes, and more important than going to college.

This article comes from the WeChat Official Account "OpenMind Club" (ID: openmindclub), author: Yang Zhiping, published with authorization from 36Kr.