Tau: The most powerful blockchain built with Logical AI. Interview with Smitha Kolan

Smitha:
Hi guys, I’m Smitha, and welcome back to my channel where I talk about all things AI and machine learning related. Today is a very special video because I’m going to be joined by Ohad Asor from IDNI and Dr. Enrico Franconi. IDNI, also known as the Intelligent Decentralized Network Initiatives, is a company hoping to solve many problems with today’s blockchain with something called Tau. Tau is a decentralized network that IDNI has created that makes a lot of decisions based on logical AI. Logical AI is something really interesting and really cool. In today’s video, we will be discussing with Ohad Asor and Dr. Enrico Franconi what exactly it is Tau, how does it solve a lot of the problems with today’s blockchain, and how does Logical AI play a huge role in this. Also, What is Logical AI, how does it differ from existing machine learning technologies. So I’m sure this video is going to be extremely helpful for you guys who are interested in machine learning, interested in logical AI, and it’s going to be a really interesting video interview to watch. So, let’s get into it. What led to the creation of IDNI?

Ohad:
Some fundamental problems within the blockchain world extend to more general issues outside the blockchain world. Specifically for blockchain, as you know, it’s all about decentralization. However, the most crucial part is not decentralized, which is the human and the social aspect. A group of developers defines all blockchain networks. If we treat blockchain seriously, as something fundamental to our economy, who guarantees that this group of developers will make the best choices that are good for the participants in the economy. Even with all good intentions, there is no reason to believe so. The developers themselves are aware of this and don’t want to make serious decisions. An economy controlled by a small group of developers cannot really work. How can we make a blockchain network controlled by its users in a way that can work in practice?

Smitha:
So, what you’re talking about refers to decentralization and centralization of blockchain. That’s a fascinating point you have mentioned, which led you to create IDNI to counteract this problem.

Ohad:
We are not the first to identify this problem, but the way other projects address this problem doesn’t solve the problem. Current projects handle this by allowing users to propose code patches and then vote on them, but this cannot work for three main reasons.

1. The inability to reason over the code.
2. The unscalability of coding.
3. The inability of code to speak about global requirements.

What do I mean by global requirements? Code is instructions for what the computer should do. Suppose you want to say that “private data should never be sent over the network.” This you cannot say in code because this is not something to do but something not to do. To ensure that your program does not send private data over the network, you need to look all over the code. That is why you call it a global statement. And make sure this requirement is satisfied. So if you propose a code patch, you still cannot say anything global about the system.

Smitha:
So security is a significant aspect that you are highlighting here. Since you can’t actually hard code these things, people would naturally assume that you are using machine learning or AI to identify what is actually secure data and, what is not, in order to protect it.

Ohad:
Yes. As you know very well, there are two main kinds of AI. One is machine learning which everyone has heard about, and there is another kind that is less known called Logical AI. Machine learning is probabilistic and is only about learning from examples. So if you want to say “do not send private data over the network,” are you going to show to your machine learning network many examples of programs that send private data over the network and many examples of programs that do not send private data over the network? Then maybe, in some probability, your software will be secure? Or you just say the thing, ”don’t send private data over the network” as it is without using examples. And that’s why Logical AI is much more suitable for this cause, unlike machine learning. We have Professor Franconi here, a leading researcher in the field of logical AI. He can explain the field and industry much better than I. I will just say that I believe the next wave of AI will be logical AI.

Smitha:
Thank you, Ohad. I will address the following question to Professor Franconi. You have specialized in the field of Logical AI and machine learning. On my Youtube channel, I talk a lot about machine learning, and many viewers are interested in what may be the next thing in machine learning. What is the following field they should focus on career-wise, etc.?. Could you give us a brief explanation about the difference between the machine learning AI that we have currently and Logical AI?

Prof. Enrico Franconi:
I would say that machine learning is based on similarities. What do we have to do to have an answer to your question? If you want to get information from your data, from your system. What machine learning does is say “well, what you’re asking me is very similar to several things I’ve already seen, and therefore I can derive this solution for you.” It’s also about approximation, so sometimes or most of the time, if you’ve done an excellent job, if you get the correct answer, sometimes you get a very approximate answer, and many more times you get the wrong answers. Because it may be biased by the learning set and your similarity function may not be so good in that particular domain. We’re talking about software verification, for example. Logic-based AI is a very old field, to be fair, and I would say there is a significant example of success. Probably 80% of the money in IT revolves around Logic-based AI, which is a database relation technology. Relation technology is all about logic.

Suppose I have a bank and want to get some money out of it. I want to be sure all the computations I’m doing are correct. They’ll probably use some relationship-based technology that tells me the exact answer. Hence, there is no approximation, no similarity. I can’t lose money because of that. So I’m using logic to get the exact answer from the data that I have because the data has a very specific meaning. We call it “semantics”, and so for this reason, logic, going back to Aristotle 2000 years ago, is all about trust, all about the fact that what I get is undisputable. We can discuss it as much as you want; I will always be right because I’m using logic. Maybe with machine learning, you could argue about the results. In logic, you cannot argue, that’s the whole point. With Ohad’s example of “do not send private data over the network,” we can prove that this will never happen. The second major success field of Logic-based AI is software verification and verification in general. Maybe you don’t want to fly a plane whose software has been programmed by a Microsoft programmer that gives you the blue screen of death. So, you want to be sure that no matter which kind of RAM is used or which type of side effect this software will end up with, it will never switch off the engine, never get into a problem. Then what you do is formal verification that these bad things will never happen; something like “you never send private data over the network.” Because the software or program is in itself based on logic, because its execution has exactly one meaning.

There’s this thinking that machine learning has become popular now because it's effective. Because there is more data and better GPUs. Machine learning started in 1958, the exact moment when AI was born, by John Mccartney. Both disciplines are very old. They’ve developed in different ways. Machine learning was lagging because of its lack of resources in terms of data of computational resources. Now it’s successful, and now what can we do with both of these things. So, the next step if you want to advise your viewers what to do with machine learning and logical AI is explainable AI. In Europe, we know that GDPR is a significant law about privacy in Europe that says that we cannot make a decision which can not be explained to private citizens. For example, you cannot be refused a loan from a bank because a neural network said, “well, you are very similar to people that do not make repayments.” After all, it is the right of the citizen to ask “why?” “Well, I don’t know why, a neural network told me that.” Maybe this neural network was biased because I was living in the wrong neighborhood or I am a person of color or whatever, right? In Europe, these neural networks are a forbidden technology because they cannot give you an explanation. Logic would give you an explanation but would not be as effective with big data as machine learning. So an interesting development, which is not about IDNI, would be the idea of explainable AI. So I would use a neural network that gives me an explanation in terms of a logical statement of rules of why it made a decision after all.

Smitha:
That’s a very good explanation of Logic AI and a significant difference between Logic AI and machine learning. Would you say Logic AI is relatively new, or has it been around for a long time?

Prof. Enrico Franconi:

Yes, it’s very old, and database technology was created from first-order logic in the early 70s. It was the only AI as of 10 years ago. Machine learning is very recent in terms of the AI level. AI has been hijacked by machine learning, I would say.

Smitha:
You could say that Logic AI is a branch within AI, with machine learning being another branch then?

Prof. Enrico Franconi:

Yes, that is exactly the case.

Smitha:
How does Tau use Logical AI?

Prof. Enrico Franconi:

The idea is that all the descriptions of software and specifications, whatever you need to describe as the system’s behavior, all the specifications will be done using logic. This means that whatever decision Tau will ever take or impose on its users will be undisputable and will collect information based on logic. That it will not gather information by looking at similarities of the goal. It will take the exact information that the users provide. With all this information, it will answer the question in an undisputable way. In the context of blockchain, cryptocurrencies, database success stories, banking transactions in e-commerce, where you don’t want to lose money, you want precision and trust, in the context of Tau, it is also quite relevant

Smitha:
That’s great. This next question applies to both of you; anyone can answer. What is the single biggest problem you think Tau is going to be solving?

Ohad:
You can imagine the Russian dolls. The matryoshka. In that, we have problems hiding bigger problems. We started with the problem of blockchain defined by its users. Inside this, you have a bigger problem. How can any software be controlled by its users? After all, the software is there for only one thing: the users. It’s only for the users, but users have had no voice in what the software should do or how it should be. The next time you install software on your phone, you can’t choose what the software update contains.

Moreover, software development scales very poorly. You know the saying, “what one programmer can do in one month two can do in two months.” When you have 10 times more programmers, you don’t get 10 times more efficiency. It’s tough to create or specify a program collaboratively, let alone by non-programmers. So, from blockchain defined by its users, we have the problem of any software defined by its users, but then we reach an even bigger problem. How can any large group of people make a collaborative decision? This is a fundamental problem on its own. We call this large-scale decision-making.
So we now have a bigger problem. How can a large group of people have a meaningful discussion? If you have a discussion with too many people, you don’t get more utility, you get less utility. How can we collaboratively combine the brainpower of a million people? How can a million people have a meaningful discussion? A solution to all these problems and others arising from those solutions, like creating a knowledge economy, the solution is rooted in using Logical AI to improve communication between people. This is our fundamental paradigm which we call Human-Machine-Human Communication. It’s all about human-human communication but in languages which the machine can understand and therefore can help the discussion.

The industry of Logical AI deals with various solutions to various problems, and maybe in contrast to all others, we specifically deal with using Logical AI to improve human-human communication. One very important example is the concept of an opinion map. Let’s talk about why do small-scale discussions succeed. Everyone says what they have to say, they don’t need to vote. Who agrees and disagrees with who is obvious to everyone, which is what we call the opinion map. In a large-scale discussion, we lose the ability to comprehend the opinion map because we are just humans. We cannot process so much information, but if it were written in languages that machines could understand, the machine would be able to recover the opinion map. So, you could imagine, for example, Twitter or Facebook where the feed displays this user said that and this user said this. Imagine you could see that “this” opinion is held by this and this and that user, which is in contrast to that other opinion held by that user, and it’s a special case of that opinion” and so on. This is not magic if we make assumptions that people will speak in logic.

Smitha:
Who would you say is the ideal user for Tau? Would you say it’s everyone, or are you trying to target a specific group of people?

Ohad:
Here should come the snowball effect of software controlled by its users. If a decentralized blockchain network is defined by its users, then, maybe at the beginning, it will be usable and interesting to a very small group of people. What will this small group do? They will just improve this very software and make it interesting for more and more people, and we get this snowball effect. We have a social human mechanism for a technological creature to evolve. So we only need to start the snowball, and then it will go all by itself.

Smitha:
Initially, you are targeting a small group of people, and you’re hoping it’s going to be adopted by a much larger group of people as more iterations and improvements are made to Tau? How do you plan to make Tau a self-evolving, self-improving software? Where does that capability come in?

Ohad:
Imagine that people say what they have to say, just like we already do now. We already have uploaded our minds to the cyber. People do so all the time. They just write whatever is in their minds, but if they did so in a language that was readable by machines, at one point, I could ask Tau how to get what I want. It would infer from what I’ve said so far about what I want. It will also infer the methods of how to get how I want when I ask it so. I could say I want Tau to be good and useful. According to what good and useful means, according to what I’ve said so far. The more information, knowledge is acquired within the system, the greater the leaps that the system can make and become better.

Smitha:
How are the contracts used within Tau different from other blockchain smart contracts?

Ohad:
Existing smart contracts are nothing but code, just machine instructions. You can only say what to do, not what not to do. They are also Turing complete which implies it’s mathematically impossible to fully reason and answer questions about the outcome of contracts. On Tau, you will be able to write contracts in languages much closer to real-life concepts and languages, less about machine instructions and more about objects in real life. Existing contracts don’t know what money or property is. They only know how to add a number. In Tau, you will be able to speak in concrete human terms about what the contract should do. Another thing that existing Logical AI does not answer and does not help to answer ourselves is that in real life, there are conditions. What happens if we wish to change the contract?

Another example, in real life, we have the law. We also have laws of changing the laws. To express the code to change the laws of changing the laws is surprisingly not addressed by existing logical frameworks. Hence, we had to invent a logical framework that supports such statements. Some logical paradoxes highlight the difficulty with this. When I say the law, I refer to the program’s behavior. We have laws, and we also have laws of changing the laws; otherwise, we can’t change the law or prevent anyone from changing the law. So the paradox becomes that we also need the laws of changing the laws of changing the laws of changing the laws. We would need infinity laws, to begin with. Perhaps the only escape from this infinite loop are laws that speak about themselves. Imagine a constitution that says that any new law that contradicts the constitution needs a 60% vote; otherwise, it needs a 50% vote. Here comes the logical trap. You can see within this constitution example there are two fundamental features.

One is that the constitution speaks about itself, and two, it speaks about there being a newly proposed law that contradicts the constitution. If you have those two language constructs or even just one of them, you are in big trouble — for example, the liar paradox. Imagine you have a sentence that says that “this sentence is false,” so this sentence cannot be true or false. Let’s assume it is true, but it says that it is false so let’s assume it is false, but then it becomes false that it is false. The liar paradox is an example of a sentence that cannot be true and can not be false. So to express this sentence all we need is a self-reference and to say if something is true or false.

Smitha:
Would you say that implementing these laws makes use of Logical AI?

Ohad:
Yes, but I’m pointing to an issue within existing logical frameworks that cannot express the laws of changing laws because they fall into similar paradoxes like this liar paradoxes precisely because they must contain the construct of self-referential metalogical statements because if they don’t contain those we get infinite statements of “the laws of changing laws” just to begin with. So we have no choice but to create logical languages which can support such statements.

Smitha:
Has this been implemented? How are you planning on implementing this?

Ohad:
We have implemented one language called TML, which is not intended to change the law. However, we have finished designing the language intended to do so and are implementing it now. We don’t currently use external tools, but perhaps we will in the future; for now, it is all written from scratch. Many reasoning tools available are not industrial quality but of academic quality, so they don’t have wide enough shoulders to carry a whole network and a whole economy. We need to write a lot ourselves and adhere to very strict engineering standards.

Smitha:
So a lot of the Logical AI software that exists is not scalable for enterprises and mostly for academic use.

Ohad:
Yes, but as Professor Enrico Franconi mentioned, databases are Logical AI. Of course, many of the database applications out there are of an industrial scale. However, they do not have self-referential metalogical statements, and they cannot discuss the law of changing the law as I’ve described.

Smitha:
In developing Tau and in making use of Logical AI to create Tau, what type of technologies have you made use of. Suppose there is someone out there that wants to learn Logical AI. What are some of the frameworks they should be looking at?

Prof. Enrico Franconi:
Independently of Tau, there is a very popular framework related to Logical AI which tries to put this more abstract logical AI than just databases and verification, AI which can reason instead of just answering queries. This is called knowledge graphs and the semantic web. You may have heard about that. That’s where ontologies which are just logical theories, logical expressions, can be written and reasoned upon. This technology is very powerful and has been around for 30 years. As Ohad mentions, there are two shortcomings. First of all, they don’t scale up to the level needed for Tau. Secondly, they don’t have this referential ability. We are taking many things from the knowledge graph and semantic web technology, but we are substantially improving upon the two dimensions: scalability and self-reference.

Smitha:
Are there any educational programs that teach Logical AI?

Prof. Enrico Franconi:
Most universities have a course on ontologies and semantic web, and there are many industry-strength courses on the knowledge graph. This idea of ontologies and knowledge graphs has become a standard by the WTC, the organization that runs the internet. 90% of the world’s e-commerce websites use Logical AI because they are forced by Google to mark up their webpages using a logical language called RDF. Behind RDF, there is an ontology defined by an organization called https://schema.org, which standardizes all the e-commerce terms. This is because Google likes to understand better what you are doing in your shop without having to parse the natural language on your web page. These web pages are marked up with logical formulae that explain what you are actually doing. Web designers do not see these Logical Formulae; they are hidden in a very nice format. The web designer will use this notation to mark up their website. All the search engines will understand their exact meaning, whatever they are doing, so they can make an indisputable conclusion about your website. They will index it in the proper way returning information in an exact way to the searcher. That’s a technology that exists, it’s standardized and it’s out there.

Smitha:
Users are the heart of this technology. How can users use this, and how can they make money using Tau?

Ohad:
It will look like any other social network where you just say what you have to say. You have comments and discussions, but in languages that computers can understand and create an opinion map. It will also take what users say about how the program should be, calculate their consensus and become what the users want it to be like. As for making money, one important example is the Knowledge economy. Because we are all about the ability to formalize knowledge, we can, for the first time, create an efficient Knowledge economy. You cannot buy or sell a single piece of knowledge right now. No one does this. You currently go to an expert and pay for what they know, but you never by yourself, or trade a single piece of knowledge. Once the knowledge is formalized for the first time, you will be able to do so. For example, if you take a law firm that makes the huge effort of formalizing in logic parts of corporate law. A company can have discussions within the system. The law firm can automatically comment in the discussion about legal aspects of what’s been raised within the discussions and charge a subscription payment for that.

Smitha:
What stage is Tau development at, and when do you expect to release Tau or a part of Tau for users?

Ohad:
We have finished implementing one logical engine called TML, which is used to create the internet of Languages within the system. Something we haven’t discussed today. The IoT is about allowing several to co-exist. Beyond that, TML will also support Proof of Execution. In existing smart contracts, all nodes need to execute all the smart contracts. It’s part of the block verification process. But using Proof of Execution, only one node will be needed to execute them and provide a short cryptographic proof that execution was correct using the sum-check protocol. This is possible because of the specific way that TML is implemented using BDDs (Binary Decision Diagrams). We are also working on implementing the system’s logic, which supports the law of changing the law and logical properties we’ve previously discussed. We also have a way to turn this logic framework into a framework that allows us to speak about programs. It’s all about programs and programs that change themselves. This is something that we are currently researching. I would guess a rough approximation a year, we will start seeing something that changes based on users’ discussions.

Smitha:
You mentioned the Internet of Languages. Can you elaborate on that?

Ohad:
The point is to not impose one language on the system but to allow languages to evolve. How can we make languages evolve? We can go in many directions with this statement. Suppose you take it to its full meaning. In that case, you will need languages that interpret each other; just like the distinction in computer science between compiler and interpreter, we need to implement one language into another. This is impractical, and there is a reason why in software development, people work with compilers and less with interpreters because it just doesn’t work well. The Internet of Languages is about allowing some ability of languages to evolve and co-exist by offering the ability to offer translators (compilers) from one language to another. To add a new language to the system, you need to specify how this language translates to an existing language. By that, to a very large extent, languages can coexist and not only knowledge representation languages but various formats. For example, You may have a knowledge base, and you wish to translate it into HTML or even Wiki format.

Smitha:
What are some of the most significant milestones that Tau has reached this last year?

Ohad:
I think the biggest thing over the last year is our team. We are significantly growing with leading researchers from the field and an excellent development team. In the first years of IDNI, I was on my own, but now that’s not the case, and we have a fantastic team working on Tau. Secondly, we managed to find a perfect solution to the logical problem that I presented before. Not only have we solved it but it is also a very good solution.

Smitha:
What further milestones do you wish to achieve moving forward next year?

Ohad:
To implement this language that allows metalogical reference statements. To implement the Proof of Execution through sum-check protocol and start a blockchain defined by its users.

Smitha:
Thank you, Ohad, that sounds amazing. I can’t wait to see what happens in the future. I’m sure lots of my viewers will be interested in what IDNI is developing and even using your product. Thank you so much.

Follow Tau for more updates. Learn more about the project at tau.net

Subscribe to Smitha’s Youtube channel for more AI content.

Join the Tau community here.

Read the Tau whitepaper here.

Get Agoras ($AGRS) here.