In Silicon Valley, on October 7-8, GPT Party 2.0 took place—the largest Russian-speaking networking event dedicated to artificial intelligence. Over 300 people gathered at Plug and Play to meet with leading experts, entrepreneurs, and investors, discuss the latest trends in artificial intelligence, and gain practical knowledge. During his presentation, Nick Davydov talked about the impact of the economy on the field of artificial intelligence, demonstrated the workings of artificial intelligence and Chat GPT, and shared insights on how to invest wisely in AI.
My job is to find new and interesting startups. We dedicate all our time to it, but I think we still miss a lot. The tectonic shift is so massive that it’s impossible to keep up with everything. I realized that I need to learn a way of thinking that allows me to catch trends faster when they become noticeable, rather than trying to cover everything. In ‘Alice in Wonderland,’ there was a phrase, ‘running twice as fast,’ and that’s actually what’s happening in the field of AI. Why is now the best time to start startups? “Let’s start with some economics. During the pandemic, a lot of money was printed, inflation began, people were quitting their jobs and changing their workplaces, and there was a significant amount of credit card debt. At the same time, businesses had relatively little debt, but they faced a different problem—they couldn’t sell and hire people. We are in a remarkable situation where the shortage of labor is not where AI can help us. It’s mostly manual labor: there is a shortage of waiters, builders, nurses, nannies, teachers. What do businesses do in this situation? They survive. So, the situation is very favorable for startups because they are forced to deal with survival all the time. Larger businesses, in times of such uncertainty, start thinking about how to remain viable. And they start looking towards startups to acquire something. So, if you’re starting a startup, you’re in a good place at the right time.” The pace of change in the field of AI: “If you’re looking at something and trying to understand whether it’s AI or machine learning, there’s a simple framework: if it’s written in Python, it’s machine learning; if it’s written in PowerPoint, it’s AI. The speed of change has changed over the past few years. Why? Let’s take a look at the image below.”
“In the 1960s, people assembled the first Perceptron, which operated on one of the early computers. At that time, AI was created with chalk on a board. Many years have passed since then, but not much has changed. Everything started to change significantly when cheaper computers became available. Once everything transitioned to graphics cards, the rate of computer growth increased. There was Moore’s law: every two years, the computational capacity of all computers doubles. Now it has stopped working, and the speed is doubling every 3.5 months. We are entering an exponential era of computers, and in 10 years, we will have neurocomputers, which will be a completely different world. Let’s try not to get lost in it. Now is the time when computers have learned to understand text, languages, read materials, images, speech, and handwritten text better than average people consistently.” What is AI, and how does it work? “Let’s talk about how neural networks work. Our brain consists of a vast number of neural networks, and we ourselves don’t know how it is structured. However, we can understand how a neuron is structured: there’s a cell with multiple inputs and one output, and a chemical reaction occurs within it. Mathematicians created a mathematical model where a neuron is a function with input variables. Neural networks equivalent to the brain can be constructed using this model.
How can we turn this into AI? One of the simplest models you can look at is called MNIST, which is handwritten character recognition. For example, it recognizes the number nine. What is a nine? It’s a field of 28 by 28 pixels, representing black, white, or shades of gray. This way, we can transform the number nine into the first layer of neurons. Then we add several layers of interconnected neurons. The last 10 neurons correspond to the numbers from 1 to 9. When they are activated, it indicates, with a 96% probability, that it’s a nine, so it lights up. The first neurons represent the pixel color, the last neuron gives the answer, and what do the neurons in the middle do? In reality, people don’t know what’s happening there and don’t understand how the machine comes to conclusions. We can guess that, in the case of MNIST, the intermediate neurons learn to recognize patterns: some find loops, others find sticks. Then, a process of fine-tuning takes place to make them converge. After that, it needs to be done several times until, at some point, the coefficients in the trained networks stabilize, consistently producing a good output that can be worked with. As you can see, explaining how to recognize handwritten text is very complex and tedious. All other models are much more complex, with many models within a generative model.” How does ChatGPT work? “Let’s talk about how ChatGPT works. It’s a large language model trained on a vast amount of data and has numerous internal parameters. To teach the model to converse and understand text, embeddings need to be created. All words can be placed close or far from each other based on their meaning in this n-dimensional space. This space is not two-dimensional but n-dimensional, with many different dimensions. Words are represented as vectors in this space, and the model is trained on how these vectors relate to each other. All large language models work by predicting the most probable next word in a sentence. If I say to you, ‘In the forest, there was…,’ okay, your LLM works with a 90% probability because in your brain, you probably thought ‘a tree,’ and with a 30% probability, you might continue with ‘she was born.’ Then there’s the probability of the next token; in the English language, it’s almost always a word, while in the Russian language, it’s almost always a pair of letters. We have a network that predicts the next word in a sentence. When we ask our question, it is used to call the correct embeddings inside the network. This way, the network recalls the necessary area of knowledge.”
“ChatGPT4 consists of several models that are called a mix of experts, where one model excels in answering management-related questions, another in literature, and so on. When using the mix of experts, it triggers memories not only of specific information but also determines the domain of knowledge. An interesting experiment you can conduct is to ask Chat GPT to write a story that starts and ends with a single word. For it, this is a very challenging task, and in half of the cases, it might not succeed. When it starts writing, it doesn’t know how it will end. This is the difference between artificial intelligence and human intelligence. Humans understand the ultimate goal of their work.” Who are Digital Workers? “David Yang came up with a cool term, ‘digital workers,’ which are not organic employees. If you plug Chat GPT4 into Chat GPT4, they can communicate with each other, and sometimes, this dialogue leads to very interesting outcomes, especially if you give them different systemic instructions. For example, one can be a project manager who doesn’t execute what users tell them but instead creates a plan on how to do it. Other instances are given prompts that they are workers and should perform certain tasks. We also create a neural network that handles reflection, meaning it looks at completed tasks and analyzes how to do it better next time. This team of virtual employees built on the internal dialogue of artificial intelligence is needed for more complex tasks than writing poetry. It can handle tasks of employees or assist them, functioning as a co-pilot, agent, or simply a tool.”
What’s happening in the economy now? “Let’s return to the economy. Every time a major technological revolution occurs, something becomes cheaper. For example, when the internet appeared, the cost of mobile communication plummeted. When the cloud revolution began, the cost of data storage dropped, and the cost of computers started to decline significantly. Currently, the primary thing that’s decreasing in cost is code. Developers may fear that their salaries will decrease, but I believe that humanity needs much more code than it has produced so far. This won’t only affect developers but any field of intellectual work related to information. This potential trend could lead to a small economic shock and significant structural unemployment. For example, if I manage a venture fund and manage to add an AI tool that simplifies the work of analysts, I won’t fire them; instead, I’ll hire more analysts so that together we become the most analytical fund in the world. However, larger companies, in times of uncertainty like this, start looking at startups to acquire and adapt to the changing landscape. So, if you’re starting a startup, you’re in a good place at the right time.”
“The greatest economic risk is associated with the rapid development of software engineers. As practice has shown, it is possible to distribute Chat GPT worldwide within a couple of months at such a pace that we can create and implement AI solutions. Anyone can reinvent their business in just a few months with the help of the developers they already have. Interestingly, statistics show that 80% of companies plan to implement AI functionality within the next year, but only 10% of them need to hire new employees to do so. Technologies are becoming simpler, and we are approaching a point where we can make employees more efficient and automate certain business functions. It is also possible to automate managerial decisions. We are constantly trying to digitize and create frameworks to move away from our human nature and make decisions without emotion. People often make managerial decisions incorrectly, while robots excel in this regard. Frameworks represent how people make decisions. If we combine this with large language models, we can automate managers. Then we can automate their employees, and that’s how we end up with fully autonomous companies.” About the Six Levels of Automation “Marina and I came up with our own six levels of company automation, which look like this:
“Right now, almost all businesses are at level zero. We can use CRM, RP, ChatGPT, calendar automation, and perhaps even implement an incredible system, but in 99% of cases, it’s trivial. Currently, only 1-2% of businesses can move to the first level. This is because such businesses have a digital twin. All actions that people take leave a digital trace. For example, when we make an investment decision, we fill out a lot of fields and numbers; we have a whole process. And the logic behind our decisions is preserved so that we can improve them and allow AI to learn from it. So, if you have a complete digital twin of your business, you can say that you’re at the first level. Then you can go to the second level, creating a fully autonomous enterprise and introducing co-pilots, AI assistants, robots, and so on. At the third level, there are agents that replace employees, and at the fourth level, autonomous departments, such as the logistics department. The fifth level is a fully autonomous business, like a box that needs money, electricity, and a task. This allows 2-3 people to create a business that sells products worldwide.” Why do people resist artificial intelligence, and what will it lead to? “There is a very small number of people comfortable with change; they are called irrational. The majority of people in society are rational, and they feel uncomfortable in uncertainty, so they begin to resist change. We can observe this happening with AI now, but something similar occurred in the past century when people were resisting electricity. They claimed that electricity was killing people and distributed propaganda leaflets. Society will always have internal resistance to change. In the end, we will reach a point where AI, like electricity, becomes ubiquitous.” How to invest in artificial intelligence? To understand how to invest in AI, you need to understand how AI works. First and foremost, AI starts with electricity because nothing functions without it. In many cases, it’s currently inefficient to implement AI because electricity costs more than hiring a human worker in places like the Philippines. We are already consuming a significant amount of electricity, and it will only increase fivefold in the next 10 years. Therefore, investments in anything related to electricity generation, transportation, and storage are promising. The next step is chips, mainly produced by Nvidia. They are incredibly expensive now, with Nvidia chips costing $1 trillion because no other company’s graphics cards work as well for large models. Supercomputers are currently a point of conflict because every cloud provider wants to buy more of them, but there are quotas, and it’s a highly political matter. Should you invest in chips and Nvidia? I’m not a stock analyst, so I won’t give you investment advice, but it seems to me that they are already priced in, and there won’t be substantial transformations. Next, there’s AI cloud software, like Nebius. Then, foundational models, such as ChatGPT 4. Is it worth investing in models? Possibly only for specific use cases, such as creating a charming AI boyfriend trained on romance novels, which won’t compete with ChatGPT. On top of this, there’s the largest layer in the inverted pyramid – AI Apps. I believe these will be essentially all businesses because in five years, businesses will either be AI-based or dead. This layer includes those who use AI to automate something or create a product.”
“The problem with investment investors in Silicon Valley is that all the major funds want to find small winners like Facebook. In AI, there won’t be winners who take it all; all of humanity will win. As a society, we will get three things. Firstly, democratization of knowledge, allowing anyone with internet access to have a conversation, for example, with a Stanford doctor and get advice. Secondly, it significantly accelerates science. We are already seeing the results of AI integration in scientific laboratories. Thirdly, AI will help us create technologies to solve humanity’s environmental problems. We currently have a major issue with planet pollution, but we lack the technologies to address it. We can develop these technologies only in collaboration with AI.”
