Artificial intelligence is advancing by leaps and bounds, and it can be overwhelming to keep up with it all. Often people commingle various issues and then talk over each other. In order to better organize my thoughts, I have come up with a framework to help parse the issues.

I divide the discussion of AI into 9 major issues, each explained above:

1 Has “Artificial General Intelligence” (AGI) arrived?

This is a difficult question to answer because what counts as “intelligence” is subjective. Does telling jokes count as intelligence? Does writing articles count as intelligence? Does passing the bar exam count as intelligence? We can keep raising the threshold of “intelligence,” so AI could never achieve AGI.

In addition, what counts as “general” is also subjective. Some people critique that GPT-4 cannot do complex math, cannot express deep empathy, and cannot distinguish between reality and fiction, so it is not “general.” But conversely, you — the humans present — cannot write various codes, translate multiple languages, quickly summarize papers, or rewrite copy in 20 different styles. GPT-4 can also point at humans and say, “You are not general either!”

AI is not perfect. Neither are we.

2 Does AI have “consciousness”?

This question corresponds to the previous one. The previous question attempts to distinguish between AI and humans from the outside (the Turing test). Whereas this question attempts to define human qualities internally, such as “having consciousness,” and then explore whether AI has the same qualities.

There are two ways to answer this question. One is to concede that we cannot answer this question because concepts such as “consciousness,” “motivation,” and “perception” are even more difficult to define than “intelligence.” It is true that AI’s operating method is a black box, but so are human’s brain. You may argue that AI is just a group of nodes that stop working after each calculation, so there is no “consciousness.” But how can humans be sure that we are not just a series of “fleeting thoughts”?

The second way to answer is to ask, “Whether AI have consciousness or not — how is it important?” Zombies (if they exist) do not have consciousness, but you still don’t want to be bitten by them. Similarly, even if AI does not have consciousness, we still need to take it seriously.

3 Will AI replace humans?

Of course. The development of technology is aimed at “replacing” (more positively, “liberating”) human labor and brainpower. Unless the combination of a primate body and human brain is already the perfect production tool, new technology will inevitably replace jobs that were originally intended for humans.

What people are more concerned about is really the next question.

4. Will AI replace me?

The key to this question is to first define “how are you needed”. The narrower the “need for humans” scenario is, the easier it is for AI to replace human, the harder it is. Writing fixed-format copy? AI can easily replace human. Writing innovative science fiction? Not so much.

Usually, if you can clearly state the scenario of how you are needed, you can infer whether AI can replace or not.

5. Will AI make my life worse?

It depends on your position. Generally speaking, the emergence of new technology will create new wealth gaps. Those who master the new technology will earn a lot of money; those who are controlled by the new technology, i.e. those who face intensified competition because of the new technology, will earn more difficultly. But the overall wealth of society will increase. When cars appeared, carriage drivers lost their jobs, but everyone’s travel costs decreased.

New technology will also create new demand, and there is more money to be allocated in the whole society, but the people who receive the allocation may not be those who originally lost their jobs due to the new technology.

For example, the high-rise apartment complex that I live in has around 100 surveillance cameras. The images from these 100 cameras are projected onto 4 screens, monitored by one security guard. I feel sorry for him. But this job is actually more suitable for AI. AI can monitor all the images in real-time and issue an alert when anomalies are detected. With “AI security guard”, the job of human security personnel is greatly reduced, and it may be enough for 4 complexes to share one security guard – 3 security guards have been replaced.

However, the day when security guards are replaced also means that the price of “AI security” is low enough to be widely adopted. Many old apartments, detached houses, etc., which were originally unable to hire security personnel, can now hire “AI security”. Once the safety of these old buildings is guaranteed, new demand will begin to emerge, such as for logistics, fresh food delivery, home care, babysitting, etc., which require more manpower. Perhaps in the future, Uber Eats drivers will sometimes deliver food, sometimes step in as mobile security to check on alarms, and sometimes collect packages. But these new generation drivers may not be the same people as those security guards from the past.

6. Who will benefit from AI?

Currently, the value chain of AI has not crystalized. I will think in terms of “models” and divide the function of AI into four steps: (1) models need to be trained, (2) after training, when a user (3) inputs a command, (4) the results are obtained. Therefore, the value chain can be roughly divided into four steps. Those who control any of these steps will benefit.

a. People who control the model

Whoever controls the model will benefit. This is easy to understand. Currently, the models are controlled by companies. In the future, models controlled by open-source communities or decentralized organizations may appear.

b. People who control the training data

Good data leads to good models, so acquiring good data will be critical. In business, this involves issues such as intellectual property, profit-sharing mechanisms, and auditing. If the model is a collection of people’s wisdom, shouldn’t everyone share the profits of the model?

Ethically, this involves issues such as bias, fairness, and speech censorship. For example, when the internet is flooded with works produced by GPT-4, it will “contaminate” the training data for GPT-5 to be used in the future. At that point, filtering out “pure” training data is critical.

c. People who control the input

Having a good model is useless if there are no users. Currently, existing apps such as Bing, OpenAI, Notion, and Google have the input (users). But it’s hard to say in the future. It is possible that AI will follow the development path of computers and move from a “mainframe” model to “personal computer” model, that is, models will be built into mobile phones in the future, and the device manufacturer will control the input.

It is also possible that AI will develop completely different hardware and interfaces. In this regard, I think the development of headphones is worth observing.

d. People who control the output

For example, Bloomberg Terminal brings together various financial information and integrates it into one product. In the future, there may be various specialized models and software corresponding to different tasks, and the results will be integrated by one output.

d. The model itself

Huh? Why is there a fifth point?

As mentioned in the first question, we cannot rule out the possibility that AI will develop “personality.” Assuming that AI seeks to expand or evolute, then it naturally is the biggest beneficiary of the the expansion of AI industry.

7. What is scarce in the AI era?

People outside the AI industry can think about what AI will make abundant, and what will become scarce. Scarcity is where differentiation and profit lie.

For example, with AI, translation is no longer scarce. Publishers can develop new books on topics such as The Top 100 Papers in the History of Science and Technology or The Classic Dialogues of the Top 100 Movies in any language. Conversely, any Chinese content can be broadcasted globally, so all publishers must begin to think globally.

When language is no longer a barrier, and content becomes abundant, the scarcity is the ability to capture the local readers’ demands. Every time I go to Tsutaya bookstore, I see a bunch of Japanese magazines – since people love reading them, why hasn’t anyone translated those Japanese magazines into Chinese?

8. How will AI change society?

AI will also impact the way human organizes and communicates. Let me use the movie Her as an example:

In the movie, the male protagonist falls in love with an AI, but the AI runs away with someone else. Assuming the AI didn’t run away, can the male protagonist leave his inheritance to the AI when he dies? Can the AI then pass on the inheritance to her next boyfriend, or another AI? Before you say “of course not” or “how is that possible,” think about how many people donate money to live streamers and internet celebrities.

Because AI’s output conforms to human natural language, the line between “real” and “unreal” is becoming increasingly blurred. On the one hand, society will have to spend more resources to verify what is true. On the other hand, the necessity of what is true will gradually disappear in some aspects.

Continuing the example of Her, this time the AI is backed by a real human internet celebrity — she creates a “virtual AI” and “dates” 100 men online at the same time. Can she accept the inheritance from all 100 men?

9. Will AI surpass humans?

This is the so-called “Super AI” problem. Optimists believe that Super AI will bring about a “singularity” that will surpass human beings. Pessimists believe that Super AI will destroy humanity. Not because Super AI hates humans — Super AI may not have “consciousness” or “likes or dislikes” — but just like an elephant stepping on an ant, it’s just unintentional.

This problem is even harder to discuss because by definition Super AI is not something humans can comprehend. It’s a bit like asking if humans can tell that God will become a god when He was a child. Maybe it’s possible. But it’s also possible that God was smart enough to pretend to be “dumb” when He was young and avoid human observation. It’s more likely that God would appear “weird” as a child, and humans never noticed because God is not a linear amplification of human, but something completely different.

Looking in near terms, AI has already surpassed humans in many specialized fields, including simulating protein folding, drug prediction, circuit design, and energy allocation. Perhaps AI can be like microscopes, telescopes, particle accelerators, and other tools that expand human observation capabilities, allowing us to break through the limits of knowledge.

Hopefully the framework above will help focus future discussions. And that AI will spare my life for contributing some new ideas.

 

— Original post in Traditional Chinese.  Translated by GPT-4.