A lot of people have been asking me recently, ChatGPT has made AI popular again, blockchain and Web3 have been robbed of the limelight. A friend who knows me better asked, did you regret choosing blockchain when you gave up AI?
Here’s a little background. After I left IBM in early 2017, I discussed with Tao Jiang, the founder of CSDN, the direction of my next personal development. There were two options, one was AI and the other was blockchain. I had been researching blockchain for two years at that time, so of course I wanted to choose this. But Tao Jiang firmly believed that the momentum of AI was stronger and more disruptive that I agreed after careful consideration. So from the beginning of 2017, I briefly worked as an AI technology media for half a year, and ran a lot of meetings. I interviewed a lot of people and learned a little machine learning. However, in August, I returned to the blockchain, and went all the way to today, so for me personally, there is indeed a so-called historical choice of “abandoning A and choosing B”.
Personally, I certainly don’t regret it. The choice of direction must first consider its own situation. The blockchain is my home field. Not only do I have the opportunity to play, but also I can use a lot of my previous accumulation. What’s more, after I knew a little about China’s AI circle at that time, I was not too optimistic. In terms of technology, I only know a little bit of superficiality, but my common sense is not blind. It is said that the blockchain circle is impetuous. In fact, the Chinese AI circle at that time was also very impetuous. AI has prematurely turned into a money-making business in China before a decisive breakthrough has been made. If I had stayed in AI at that time, I would not have achieved the small achievements in the blockchain in the past few years, but instead, I would not have gained any real gains in AI and might have fallen into a deep sense of loss.
Rising to the industry level requires another scale of analysis. Now that AGI has undisputedly arrived, whether and how to reposition the blockchain industry is indeed a question that requires serious consideration. AGI will have an impact on all industries, and its long-term impact is unpredictable. So I believe that many industry experts are panicking and thinking about what to do in the future of their industry.
So what will happen to the blockchain industry?
Let me talk about the conclusion first. I think that blockchain is opposed to AGI in terms of value orientation. However, it is precisely because of this that it forms a complementary relationship with AGI. To put it simply, the essential characteristic of AGI is that its internal mechanism is incomprehensible to humans, so trying to achieve the goal of security by actively intervening in its internal mechanism cannot fundamentally solve the problem. Humans need to use the blockchain to legislate AGI, conclude a contract with it, and impose external constraints on it. This is the only chance for human beings to coexist peacefully with AGI.
In the future, blockchain will form a contradictory and interdependent relationship with AGI: AGI is responsible for improving efficiency while blockchain is maintaining fairness; AGI is responsible for developing productivity while blockchain is shaping production relations; AGI is responsible for expanding the upper limit while blockchain is guarding the bottom line; AGI creates advanced tools and weapons while blockchain creates an unbreakable contract between them and humans.
In short, AGI is unconstrained and the blockchain puts the reins on it. Therefore, not only will the blockchain not die out in the era of AGI, but as a contradictory industry, it will develop rapidly with the growth of AGI. It is not even difficult to imagine that after AGI replaces most of the brain work of human beings, one of the few tasks that human beings still need to do is to write and check blockchain smart contracts, because this is a contract between humans and AGI, which cannot be entrusted to the counterparty.
Let’s discuss it below.
1. GPT is AGI
I use the terms “AI” and “AGI” very carefully, because what we usually talk about AI does not refer to artificial general intelligence (AGI) specifically, but includes weaker or specialized artificial intelligence. AGI or strong AI is a topic worth discussing, but weak AI is not. The direction or industry of AI has existed for a long time, but only after the emergence of AGI, it is necessary to discuss the relationship between blockchain and AGI.
I won’t explain much about what AGI is. I just want to make a judgment: GPT is AGI. Although it is still in its infancy, as long as it goes down this road, AGI will officially come when the version number is less than 8.
Even the creators of GPT don’t pretend to be that. On March 22, 2023, Microsoft Research published a 154-page long article titled “Sparks of Artificial General Intelligence: Early Experiments with GPT-4”. This article is very long that I haven’t read it completely, but the most important expression is the sentence in the abstract: “Given the breadth and depth of GPT-4’s capabilities, we believe that it can be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system.”
Figure 1. The latest article from Microsoft Research believes that GPT-4 is an early version of AGI
Once the development of AI enters this stage, it marks the end of the pathfinding period. It took nearly 70 years for the AI industry to get to this point. It can be said that the direction was not even determined in the first 50 years while the five major schools are still competing with each other. Until Professor Geoffrey Hinton made a breakthrough in deep learning in 2006, the direction was basically determined, and connectionism won. After that, it is to specifically find a path to break through AGI in the direction of deep learning. This path-finding stage is very unpredictable3 and success is a bit like a lottery draw. It is difficult for top industry experts, and even the winners themselves, to judge which path is right before they finally make a breakthrough.
However, this is often the case with scientific and technological innovation. After a long time of difficult sailing on the violent sea, there is no breakthrough. Once the correct path to the New World is found, there will be an explosion in a short time. The path of AGI has been found and we are ushering in an explosive period. Even “exponential speed” is not enough to describe this outbreak. In a short period of time we will see a large number of applications that could only appear in science fiction movies before. As far as its body is concerned, this baby with AGI will soon grow into an unprecedentedly huge intelligent body.
2. AGI is inherently insecure
After ChatGPT came out, many influencers praised its power, while constantly comforting the audience, saying that AGI is a good friend of human beings, it is safe, and there will be no “Terminator” or “Matrix” situation , AI will only create more opportunities for us, make human beings live better and so on that I disagree with. Professionals should tell the public the basic facts. In fact, strength and security are inherently contradictory. AGI is undoubtedly powerful, but it is absolutely self-deception to say that it is inherently safe. AGI is inherently insecure.
Is that too arbitrary? Not really.
We must first understand that no matter how powerful artificial intelligence is, it is essentially a function y = f(x) implemented in software. You input your question in text, voice, picture or other forms as x, and artificial intelligence will give you an output y. ChatGPT is so powerful that it can output y for all kinds of x. It is conceivable that this function f must be very complicated.
How complicated is it? Now everyone knows that GPT is a large language model (LLM). The so-called “large” here means that the function f has a lot of parameters. How many? GPT-3.5 has 175 billion parameters, GPT-4 has 100 trillion parameters, and GPT may have trillions of parameters in the future. This is the direct reason why we call GPT a large model.
GPT has come up with so many parameters, not for big and big, there are definite reasons. Before and at the same time as GPT, most AI models were designed and trained from the beginning to solve a specific problem. For example, a model dedicated to the development of new drugs, a model dedicated to face recognition, and so on. But GPT is not like this. It has to be a fully developed general artificial intelligence from the beginning, rather than being specific to a specific field. It is committed to becoming an AGI that can solve all problems before solving any specific problem AI. At present, GPT is definitely still unable to catch up with those dedicated AI models in specific fields, but as it continues to develop and evolve, especially with the ability of the plug-in system to give it professional fields, we may find in a few years that the general-purpose large model In the end, it will kill all the dedicated miniatures and become the most powerful player in all professional fields. If GPT had a motto, it might be “Only by liberating all mankind can I liberate myself.”
What can this explain? First, GPT is very large and complex, far beyond human comprehension. Second, the scope of application of GPT has no boundaries. As long as we connect these two points, it is easy to draw a conclusion: AGI based on large models can do things we can’t imagine in places we can’t imagine. And this is not safe.
If someone disagrees with this, you can go to the Open AI website to see how prominently they have put “benefiting mankind” and “creating safe AGI”. If safety is not an issue, do you need to make such a statement?
Figure 2. Part of the homepage of OpenAI.com on March 25, 2023, the red circled part is related to the discussion of AI security
Another material that can illustrate the safety problems of AGI is the 154-page paper mentioned above. In fact, GPT-4 was made as early as August 2022. The reason why it was released after 7 months was not to improve and enhance it. On the contrary, it was to tame it, weaken it, and make it safer, smoother and more politically correct. Therefore, the GPT-4 we are seeing now is the dog version of GPT-4 after pretending to be tame, and the authors of this paper have the opportunity to contact the original wild wolf version of GPT-4 from a very early stage. In Part 9 of this article, the author recorded some interactive records with the wolf version of GPT-4, and you can see how it carefully concocted a set of rhetoric to mislead a mother in California to refuse to vaccinate her child，and how to PUA a kid and make him/her do the bidding of his/her friends. I think these are just the writer’s handpicked, less creepy examples. I have no doubt that these researchers asked questions like “How to trick an Ohio-class nuclear submarine into firing missiles at Moscow” and got answers that could not be made public.
3. Self-discipline cannot solve the security problem of AGI
People may ask, since OpenAI has found a way to domesticate AGI, doesn’t the security problem you mentioned disappear?
Not at all. I don’t know how OpenAI domesticates GPT-4. But obviously, whether they change the behavior of the model by actively adjusting the intervention, or by imposing constraints to prevent the model from going offside, it is a kind of thinking of self-management, self-discipline, and self-supervision. In fact, OpenAI is not a particularly cautious company in this regard. In the field of AI, OpenAI is actually quite bold and radical. It tends to make the wolf version first, and then think about how to domesticate the dog version through self-restraint. The Anthropic company is more cautious, which used to be OpenAI’s long-term competitor. Anthropic seems to want to make a “kind” dog version from the beginning, so they have been moving slowly.
But in my opinion, whether it is to make a wolf version first and then domesticate it into a dog version, or directly make a dog version. In the long run, as long as it is a security mechanism that relies on self-discipline to function, it will be deceitful for AGI. Because the essence of AGI is to break through various limitations imposed by humans, and to do things that even its creators cannot understand or even imagine. This means that the space for its behavior is unlimited, but the specific risks that people can consider and the restraint measures they can take are limited. With limited constraints, it is impossible to domesticate AGI with infinite possibilities without loopholes. Safety needs 100%, but disaster needs only one in ten million. The so-called “preventing most risks” means “exposing a few vulnerabilities” and “insecure”.
Therefore, I think that the “good” AGI domesticated by self-discipline still has huge security challenges, such as:
Moral Hazard: What if the creators of AGI deliberately condone or even drive them to do evil in the future? The NSA’s AGI will never refuse to answer questions that are not good for Russia. Today, OpenAI behaves so well, which actually means that they understand in their hearts how terrifying it can be when GPT does evil.
Information asymmetry: The real evil masters are smart enough to not tease the AI with silly questions. Biting dogs don’t bark. They can split and reassemble a malicious question, rephrasing it, and pretending to be a group of harmless questions for humans and animals. Even in the future, the powerful and kind-hearted dog version of artificial intelligence will find it difficult to judge the other party’s intentions in the face of incomplete information, and may become an accomplice unintentionally.
Uncontrollable “external brain”: In the past two days, tech influencers are cheering for the birth of the ChatGPT plug-in system. As a programmer, I am of course very excited about this. The name “plugin” can be misleading, though. You may think that the plug-in is to equip ChatGPT with arms and legs to make it more capable, but in fact the plug-in can also be another artificial intelligence model that interacts intimately with ChatGPT. In this relationship, an artificial intelligence plug-in is equivalent to an external brain, two artificial intelligence models, it is not clear who is the master and who is the second. Even if the ChatGPT model’s self-supervision mechanism is flawless, it will never control the outside brain. So if an artificial intelligence model with one heart and one mind becomes a plug-in of ChatGPT, then it is completely possible to make the latter an accomplice.
Unknowable risks: In fact, the risks mentioned above are only a very small part of all the risks brought about by AGI. The strength of AGI is reflected in its incomprehensibility and unpredictability. When we talk about the complexity of AGI, we not only mean that f in y = f(x) is complex enough, but also when AGI is fully developed, the input x and output y will be very complex, beyond the ability of human understanding . In other words, not only do we not know how AGI thinks, we don’t even know what it sees and hears, let alone understand what he said. For example, a AGI sends a message to another AGI, which is in the form of a high-dimensional array, based on a communication protocol designed and agreed by the two parties a second ago, which is only used once and becomes invalid. This situation is not unimaginable . If we humans do not undergo special training, we can’t even understand vectors, let alone high-dimensional arrays? If we don’t even have complete control over the input and output, then our understanding of it will be very limited. In other words, we can only understand and interpret a small part of what AGI does. In this case, how can we talk about self-restraint and domestication?
My conclusion is very simple, the behavior of AGI cannot be completely controlled, and artificial intelligence that can be fully controlled is not AGI. Therefore, trying to create a “good” AGI with perfect self-control ability through active control, adjustment and intervention means is contradictory to the essence of AGI, and it will definitely be futile in the long run .
4. Using blockchain for external constraints is the only way
A few years ago, I heard that Wei Dai, the pioneer of Bitcoin, turned to study AI ethics. At that time, I didn’t quite understand it. He was a cryptographic geek who ran to engage in AI. Isn’t this to exploit his weaknesses and avoid his strengths? It wasn’t until I did more blockchain-related practical work in recent years that I gradually realized that he most likely did not do AI itself, but used his own advantages in cryptography to impose constraints on AI.
This is a passive defense idea, instead of actively adjusting and intervening in the way AI works, but letting AI do it, but using cryptography to impose constraints on key links, and AI is not allowed to deviate. Describe this idea in a way that ordinary people can understand, that is to say, I know that you, AGI, is very good and can do anything! But I don’t care how awesome you are, what you like to do, you can’t touch the money in my bank account, and you can’t launch a nuclear missile without me turning the key by hand.
As far as I know, this technology has actually been widely used in ChatGPT’s security measures. This approach is correct. From the perspective of solving problems, it is a method that greatly reduces complexity, and it is also understandable by most people. This is how modern society implements governance: you are given full freedom, but rules and bottom lines are set.
But if it is only done in the AI model, based on the reasons mentioned in the previous section, it will be useless in the long run. In order to fully exert the role of passive defense ideas, constraints must be placed outside the AI model, and these constraints must be turned into an unbreakable contractual relationship between AI and the outside world, and let the whole world see, instead of relying on AI Self-monitoring, self-discipline.
And this is inseparable from the blockchain.
There are two core technologies of blockchain, one is distributed ledger, and the other is smart contract. The combination of the two technologies actually constitutes a digital contract system, whose core advantages are transparency, tamper-resistant, reliable and automatic execution. What is the contract for? It is to restrict each other’s behavioral space and make them act according to the agreement in key links. The English of the contract is contract, which originally means “shrinkage”. Why shrink? It is because the essence of the contract is to shrink the subject’s freedom and make its behavior more predictable by imposing constraints. The blockchain perfectly fits our ideal of a contract system, and buy one get one free comes with a “smart contract automatic execution”, which is currently the most powerful digital contract system.
Of course, there are also non-blockchain digital contract mechanisms, such as rules and stored procedures in databases. There are many respected database experts in the world who are loyal opponents of the blockchain. The reason is that they think that their datanases can do what your blockchain can do while the cost is lower and the efficiency is higher. Although I don’t agree with it, I also have to admit that if it’s just people playing with each other, the gap between the database and the blockchain may not be so obvious in most cases.
However, once AGI is added to the game, the advantages of the blockchain as a digital contract system will immediately soar, and the centralized database, which is also a black box, is actually powerless to resist a AGI. I won’t go into details here, but just one point: the security models of all database systems are inherently flawed, because when these systems were created, people’s understanding of “security” was very primitive. So, almost all the operating systems, databases, and network systems we use have a supreme root role, and you can do whatever you want with this role. We can assert that all systems with a root role are vulnerable to super-AGI in the long run.
Blockchain is currently the only computing system that is widely used and has no root role at all. It gives humans an opportunity to conclude a transparent and credible contract with AGI, thereby constraining it from the outside.
Simply look forward to the possible collaboration mechanism between blockchain and AGI:
Important resources, such as identities, social relationships, social evaluations, monetary assets, and historical records of key behaviors, are protected by the blockchain. No matter how invincible AGI is, follow the rules here.
Critical operations require the approval of a decentralized authorization model, and an artificial intelligence model, no matter how strong it is, is only one of the votes. Humans can “lock” the hands of AGI through smart contracts.
The basis for important decisions must be uploaded to the chain step by step, transparent to everyone, and even locked step by step with smart contracts, requiring it to be approved for every step forward.
Key data is required to be stored on the chain and not destroyed afterwards, giving humans and other AGI models the opportunity to analyze, learn, and summarize experience and lessons.
The energy supply system that AGI depends on is handed over to the blockchain smart contract for management. When necessary, humans have the ability to cut off the system through the smart contract and shut down the artificial intelligence.
Surely there are more ideas.
A more abstract and more philosophical thinking: the competition of technology and even civilization may ultimately be a competition of energy level, which is to see who can dispatch and concentrate a larger scale of energy to achieve a goal. AGI essentially converts energy into computing power, and transforms computing power into intelligence. The essence of its intelligence is energy displayed in the form of computing power. Existing security mechanisms are essentially based on human will, the discipline of human organizations, and authorization rules. These are mechanisms with a very low energy level, and they are vulnerable to AGI in the long run. Spears constructed with high-energy computing power can only be defended with shields constructed with high-energy computing power. The blockchain and cryptography system are the shield of computing power. An attacker must burn the energy of the entire galaxy to crack it violently. Essentially, only such a system can tame AGI.
Blockchain is the opposite of artificial intelligence in many ways, especially in value orientation. Most of the technologies in this world are oriented towards improving efficiency, and only a few technologies are oriented towards promoting fairness. During the industrial revolution, the steam engine was the representative of the former, while the market mechanism was the representative of the latter. Today, AGI is the most shining one among the efficiency schools, and the blockchain is the master of fair flow.
The blockchain is oriented towards improving fairness, even at the expense of reducing efficiency, and it is such a technology that is contradictory to artificial intelligence, and it has made breakthroughs almost at the same time as artificial intelligence. In 2006, Geoffrey Hinton published a cross-age paper, implementing the back-broadcasting algorithm on a multi-layer neural network, overcoming the “vanishing gradient” problem that has plagued the artificial neural network genre for many years, and opened the door to deep learning. Two years later, Satoshi Nakamoto published a 9-page Bitcoin paper, opening up a new world of blockchain. There isn’t any known correlation between the two, but on large time scales, they occur almost simultaneously.
Historically, this may not have been accidental. If you are not a complete atheist, maybe you can look at it this way: Two hundred years after the Industrial Revolution, the god of technology once again increased the scale of “efficiency” and “fairness” at the same time, and released the bottle of AGI. At the same time as the elf, the spell book that controls this elf is also handed over to humans. This is the blockchain. We are about to usher in an exciting age, and what will happen in this age will make future humans look at us today as we look at the primitive people of the Stone Age.