Apply AI or Die?

– You often come to give lectures and presentations on applied artificial intelligence. What are your impressions—are these technologies making their way into Bulgarian business?

– Definitely not. Alas, there aren't many specialists with deep knowledge of this technology, and I am endlessly disappointed, because I don't even see a discussion. It slides entirely onto the plane of large language models, which require no fundamental knowledge of AI and have limited capacity to generate added value.

I believe that artificial intelligence in business is not a fashionable topic in Bulgaria, and that is exactly where we need to put the emphasis.

I expect much more talk about practical applications. I think it is important to draw attention to this, for companies that have already started the process to talk about their experience. It has to begin with the vision of the people who make the important decisions in a company. If they don't grasp the significance of artificial intelligence, that it is a huge prospect, that from now on they can't use the old tactic—“let's keep our heads down and the storm will pass”—if they look at artificial intelligence as some temporary phenomenon…

No, the storm won't pass. On the contrary, it will rage with ever greater force. And if we don't act right away, within 5 years it may well blow us away.

To move forward, the people who make decisions have to be convinced. They have to understand that they need to invest seriously in artificial intelligence.

We have a golden opportunity! It's good that we talk about startups, but with them you have seeds—you scatter them and wait for them to sprout. Only you need fertile soil, some readiness on the part of Bulgarian business. That will be a longer battle, one I have been fighting for many years.

I think your platform is a way to reach business, to popularize the subject and get more young people to take this path…

– I try to look for such examples myself. I often host people from business, and I don't see things happening as quickly or on as large a scale as expected either. Meanwhile, for a year and a half the world hasn't stopped talking about artificial intelligence and ChatGPT. But I read that you call the hero of the day a “chat-pat artificial intelligence.” What do you mean?

– The great misconception of the last two years is that ChatGPT is a synonym for artificial intelligence, when it is only one of its possible forms. One of the reasons this has been communicated so aggressively, especially in Bulgaria, is that there are almost no people who understand real, core artificial intelligence. In Bulgaria, those who know it in depth number no more than a few hundred. So ChatGPT appeared on November 30, 2022, and by December quite a few people, especially in the IT sector, could see what it was about—that this technology could be very fast and could generate code. That's when the message popped up that froze the blood of many: so it's going to take the bread out of our mouths. It soon became clear that this was only the first step in the long process toward a fully generated IT project.

ChatGPT is one possible form, a very popular one, which brought artificial intelligence closer to the masses, and that's why it became so popular.

– But in a way it also harmed the subject, insofar as people really began to associate the term only with it and all its imperfections.

– The big question concerns business. My experience is in applied artificial intelligence, in how its capabilities are used in business, and in the “heavy artillery” at that—industry. I started as a postdoc at the University of Delaware, where I had projects with the chemical giant DuPont. After I got my green card, I was offered a job at their main competitor, Dow Chemical. Now I consult for the largest paper manufacturer in the US, Georgia-Pacific.

In other words, I know industry. In December 2022 it was extremely confused by the lightning-fast interest in ChatGPT. In the very first days, the vice president for innovation at the company I consult for wanted me to answer him at lightning speed, precisely and clearly: should we use ChatGPT? I examined its capabilities in as much detail as I could, until one answer tipped the scales toward a categorical “no.” To my question “Are there Shopi in Bulgaria?”—the Shopi being the people of the region around Sofia—it replied: “Yes, there is shopping, supermarkets.”

Well, how can I say “yes” after an answer like that? That's when the wordplay came to me, drawing on the richness of the Bulgarian language: this is a “chat-pat artificial intelligence,” one that gets things right now and then. There are areas where it is very successful, very suitable. But for business that is rarely the case, simply because it can't tell a lie from the truth. It's no accident that “hallucination” was the word of the year in English for 2023. Thank God,

the developers of large language models realized that unless the problem of hallucinations is solved, the future won't be very bright.

There are already thousands of software tools that tackle this, step by step. And yet the problem is fundamentally unsolvable because of the very architecture of the large language model.

– Why is that?

– We come to the great “ideological divide,” about which, for some reason, Bulgaria keeps complete silence.

This is the message I insist my students understand well. At the moment there are two ideologies about the future of artificial intelligence. The first, the dominant one, holds that we can develop artificial intelligence that is not structured. It is based on the scaling law of large language models—the more parameters we have, the more accurate their answers. If, say, a model grows from 7 to 65 billion parameters, it starts working significantly better. But this way we can't understand why the model predicted the next word one way and not another. That is the only thing a generative language model does—it predicts the next word based on those billions of parameters.

The big problem is that we don't know what's going on inside this hypercomplex model. For neural networks we use the analogy of a black box, but here it's much scarier—we have a black hole. And we know very well from astronomy that black holes swallow matter.

But it is the cheap and fast way forward. To use this law—now they're already thinking about tens, hundreds of trillions of parameters, and the next-numbered GPT will predict better on the basis of what may come in a year or six months. Behind this approach stands a brilliant startup manager, the head of OpenAI, Sam Altman.

He doesn't come from science. What matters to him is to start a business with small investments and inflate it, and he has done that brilliantly.

However, there are some small details that are beginning to trouble people: the scientific team has left him. His chief scientist, Ilya Sutskever, one of the creators of deep neural networks, organized the coup, failed and withdrew.

– These topics are certainly not discussed enough, and they're the reason this site exists. Not long ago, with another of my guests, Viktor Botev, we talked in some depth about the clash between these two currents. On the other side of which stands Meta's technology leader and another of the “fathers” of AI, Yann LeCun.

– And thank God, because not only is he a great scientist, but behind him stands a tech giant—Facebook. I too stand firmly behind his philosophy, precisely because I am a scientist. Even though I come from industry, I can tell very well when something is solid and when we're just beating our chests that we're going to build general AI.

People, to do that you need so many technical solutions! The path obviously goes through the structured approach and the vision of how to get there. Yann LeCun sets the direction: let's start from biological intelligence, from human intelligence. The brain is structured; our knowledge isn't based only on text, a large part of it comes from visual sources. That's why he works slowly and methodically in this direction. There is an alternative model that has been in the works for 5 years now, and the road won't be short, it won't be next year. Maybe it will take 20 years, but that is the path.

The other path, Altman's unstructured one, is dangerous!

If in physics we had taken the easy road, without a theory of the atom, without understanding the physical forces behind it, and just started making bombs, I don't know how we would have carried on… That is exactly the road they are proposing we take with the unstructured approach. As a scientist, as someone who has been in this field for over thirty years, I see the danger, and I'm not alone. Almost all the big names in core AI have distanced themselves from the unstructured approach; the first was Geoffrey Hinton.

This, in my view, is the fundamental question facing the development of artificial intelligence. Naturally, big business stands behind the unstructured approach, investing tens of billions, and now the surplus money of the Arab countries is getting involved too. Still, I hope we won't end up with the negative consequences of not understanding the black hole…

– What could those be? On the one hand, we may be close to the limit of large language models—we have no more data to train them on. But if it does work out, what are the dangers of this black hole?

– In my view, there's no way to reach truly general artificial intelligence like this. We can't achieve deep knowledge and thinking, because we don't know the mechanisms of intelligence, of intellectual activity. What are we doing now? We compensate for this birthmark of large language models by injecting our own knowledge, through prompting techniques.

What Sam Altman says more and more often in interviews is that general AI will arrive “when we decide.” Because it depends on how the concept of “intelligence” is defined, and that's where the big disagreements arise; there's a long list of what any such system must have.

The main thing is self-awareness, but there's also enormous divergence over what we mean by that. We could get bogged down in endless debates about basic concepts, while practical people like Sam Altman and Satya Nadella will at some point bang on the table and say: “There, this is general artificial intelligence.”

At some point we'll plant the flag and announce that we've landed on the Moon… without having the Saturn V rocket. And we'll only think we've landed.

– But we're all in rather a hurry… Why would the structured approach take so much time?

– Because it is fundamental. Here the center of gravity shifts to a great extent from language models to dynamic, visual models, and developing those is orders of magnitude harder and more expensive.

Second, we already have a lot of research on the early stages of intelligence. Kahneman, the Nobel laureate in behavioral economics, defined the system of “genetic intelligence”—something we have in our brains and that is probably genetically built in. We all receive an initial intelligence, which ChatGPT doesn't have, even though it contains, compressed, all the information available in textual form. But how do we understand where the difference lies?

So, first, in the structured approach we have a sequence of scientific research before we arrive at a result better than what we have now. But what we have now satisfies, let's say, 50 percent of the users of artificial intelligence. Schoolchildren and students are jumping for joy, teachers to some extent too, people in marketing. I'm even surprised at the patience of Facebook's leadership, which at any moment could say: “Hold on, our competitors are making tens of billions of dollars, and here we are working on better concepts that might eventually put us ahead in 10 years.”

There is a problem, but they stand behind LeCun, because he is an undisputed authority. Although he isn't politically correct—he flat-out shoots down large language models. I don't agree with statements to the effect that in 2 years nobody will use them, that they don't work. That's not true—the fact that they are used by at least 300 million people means there's something to them.

– Well, I can say from direct experience that they undoubtedly help and certainly speed up the writing of computer code… Where do they have a place?

– At this stage, wherever they have reached the necessary level of maturity. I'll borrow a thought from a professor at Wharton, the most prestigious business school in the United States. He published the book Co-Intelligence, about “joint intelligence,” whose

main idea is that in the end we won't have a pure form of artificial intelligence, but the joint action of the artificial and the natural.

That is the reality, and I have to tell you that from a business point of view things have been like that for years.

The plus of generative artificial intelligence is that the model you're using right now is the worst possible one—in just three months there will be a better one. And that the progress behind it, precisely because of the scaling law, will help develop bigger and more complex models.

But the big minus is that it will never be able to reason deeply, to tell the truth from a lie.

These things will be a barrier to an automatic, seamless adoption of generative models, especially in business. You say you use their capabilities, and with every version they'll grow more powerful, but in any case you hardly trust it 100 percent. There's always an iteration of checking whether the generated solutions are correct.

– Of course. But even so it makes economic sense, because it makes me much faster and more efficient at my work.

– Naturally. But you also have to keep something else in mind—the time you lose on checking. And quite a few people forget that small effect. I'll tell you from personal experience: four years ago we changed the strategy for applying AI at the company I consult for. Before, it was what I call the “lumberjack” strategy—we have a problem, say the paper machines stop. We set out to solve that problem, then the next one. But we realized that the effect wasn't big until we solved the problem of the forest. To work out a common strategy for where artificial intelligence would support the company as a whole. To build something like a digital brain.

The pandemic pushed us in that direction too—at the start, this company faced an enormous challenge. Within a few weeks, demand for toilet paper rose by 73 percent; it was like that everywhere. In that difficult period the company needed fast and effective solutions. Before that, whenever we worked with artificial intelligence, we were “ideologically” pushed again and again toward data, data, data. Unfortunately, in this particular case we had no data—the last such situation had been a hundred years earlier, with the Spanish flu.

The only thing that saved us was knowledge. That's when, in my view, the whole field of applied artificial intelligence turned away from the paranoia about data and toward knowledge. Yes, it all comes down to balance, but that's when we started thinking about the second part.

We began with the only large language model of the time—BERT. We moved insanely fast, which showed us in practice that what matters is crystallizing specialized knowledge.

Business doesn't need a large language model on its own; nobody gains a competitive advantage with run-of-the-mill, off-the-shelf knowledge. It makes sense when other models with specialized knowledge are added—ontologies alone, language technologies alone won't do it! You need another method too, with knowledge graphs, which structures the data in a permanent form rather than a fluid one. Ask a large language model something 5 times and it will give you a different answer each time. Well, that's not how you do business.

The third factor is that when we make decisions, they have to be based on cause-and-effect relationships. These three components—specialized knowledge, structured with knowledge graphs, and cause-and-effect relationships—are the foundation of modern deep artificial intelligence, and they are what business believes in…