GPT Party 3.0. David Yang: Digital Workforce: The Future of Work and the Impact of AI on Humanity GPT Party 3.0, the largest Russian-language networking event dedicated to artificial intelligence, took place in Silicon Valley on March 9-10. More than 450 people gathered at Plug and Play to meet leading experts, entrepreneurs, and investors, discuss the latest trends in artificial intelligence, and gain practical knowledge. At GPT Party, David Yang, founder of ABBYY and co-founder of Newo.ai, spoke on the topic “Digital Employees: the Future of Work and the Impact of AI on Humanity.” David shared trends in artificial intelligence, and also talked about what types of digital agents there are and how to apply them in practice in everyday life and business. David Yang, Founder of ABBYY, Co-founder of Newo.ai The beginning of this year has been marked by an incredible waterfall of events. Just a year ago, when we were just getting acquainted with ChatGPT 3.5, then ChatGPT 4, we thought, what else can humanity come up with that is so incredible? But look at what has happened since the beginning of the year. On January 11, 1X Robot raises $100,000,000 in investments from OpenAI to develop robots. On February 15, Sora appears, a fantastic model for creating video. Also on February 15, Google announces Gemini 1.5. On February 26, NVIDIA launches a fund for humanoid robots. This is a large-scale event, because if NVIDIA were a country, it would be the 12th country in the world by market capitalization. On February 28, Meta announces the imminent release of Lama-3, which, according to preliminary data, will surpass existing models in benchmarks. On February 29, Figure, a robot part of NVIDIA’s plan, raises $675,000,000 from investors including Bezos and OpenAI. On March 1, Grok becomes the big story, rising to $1,000,000,000 in valuation. On March 4, Claude 3 is announced by Claude Anthropic, which promises to be a major breakthrough. Also on March 4, information leaked about OpenAI’s work on ChatGPT 5, 6, and 7, models that surpass the human brain in terms of parameters. On March 7, Inflection-2.5 raised about $200,000,000 with a project at the idea level. Now we know that their new version will outperform previous versions in benchmarks. This is just a small part of what has happened in 2024, which has just begun. I would like to especially note the story related to humanoid robots. The fact is that a fundamental event has occurred in this area, and it is related to Moravec’s paradox. This is an interesting topic related to the assumption that fine motor skills require more computing resources than reasoning, and therefore we will never be able to create normal human-like behavior and a self-developing machine. But the parallel work of a huge number of companies in the world engaged in humanoid robots, of course, made itself felt. Here on the left I have written out only a small number of companies that are engaged in this topic. The first plant for the mass production of humanoid robots was recently built. These are the “Digit” robots, which will be produced in quantities of 10,000 units per year. This is still not so much compared to the number of cars produced. But it is already quite obvious that the volume of industrial production of robots will exceed the number of cars produced. So, what is Moravec’s paradox? It is based on the assumption that we do not have enough resources to create something that requires more resources than thought processes, and therefore, there will be no self-learning. And so, on February 21, a series of materials were published about a robot that learns without instructions from a person. This is learning by observation, and this is some kind of complete fantasy. This robot studies not only how a person behaves to perform certain actions, but also how other robots that have already learned these actions behave. If it does not receive the necessary information, then it searches for it on the Internet, on YouTube. Why is this important for our lives and for business? At the last GPT Party, I talked about digital workers, about how hundreds of thousands of knowledge workers will be assisted by hundreds of thousands more digital knowledge workers. But we talked about white collars, not counting blue collars, there are 1 billion of them in the world. If the price of humanoid robots with fine motor skills and self-learning capabilities drops to $10,000 or even $5,000, it will radically change the processes in business and retail. Robots will start doing the work of loaders, waiters, cleaners, workers in production, etc. It is difficult to say exactly when this will happen. Presumably, in the next two years, the prices for these robots will vary around $20,000, and their availability will not be so high. But in a year or two, mass production will reduce their price to $10,000 and below. A little more on the topic of humanoid robots. Why were Unitree robots called dogs? Because it was believed that their dynamics did not allow them to be called cats. Cats have twice as many neurons in their bodies as dogs, and that is why cats are so flexible. I started working with this company when it was very small, it had 20 employees, and I tried to invest in it. It did not work out, but I maintained relations with their managers. I have two such dogs: one is the earliest version, her name is Abigail, and the other is Lio, a younger one. But pay attention to the latest achievements of Unitree. Look at how modern models jump. Dogs do not jump like that anymore, cats do now, and this is a new level of development of humanoid robots. As for Sora, it is an amazing model that turns an image into a video. We have an old video, it is funny, so six months ago no one took it seriously. But now the company has reached a new level and it is a completely different story. Pay attention to the reflection of neon lights in puddles. When a neon light changes on a building, its reflection changes in a puddle, the folds of a dress sway absolutely in accordance with the laws of physics. In order for such a video to splash out of this amazing alien brain of a model, it actually has to already argue knowledge of the physics of this world in its 1,000,000,000,000 parameters. And not only physics, but also everyday objects. Just the fact that this video of such length appeared suggests that we have moved to a completely different world. I recently, literally a week ago, met with Raymond Kurzweil. It turned out that our professional paths crossed much earlier. Raymond created a giant flatbed scanner back in the 70s, with which he scanned materials for a lecture, and we began actively developing FineReader in 1991. And for several years, we were, in fact, competitors and colleagues in the field of text recognition. He gave a lecture on how he now sees his predictions, in particular, singularity and so on. He said that singularity would happen in 2035, although now he believes that he was pessimistic. This prediction was made in 2005, when it was still difficult to think that computers would start writing programs for themselves, and that they would start doing it better than humans. In particular, Raymond talked about how the computer cost per $1 has changed over almost 100 years, from 1935 to the present. He calculated that over these years, the number of computer operations per $1 unit, adjusted for inflation, has changed by 10-17 times, that is, by 100,000 trillion times. This is the number of times more calculations can now be done for $1. And every five years, this number continues to increase tenfold. This is a graph from last year comparing ChatGPT 3.5 and ChatGPT 4. It was already determined in the first half of 2023 that ChatGPT 4 was superior to humans on most US college graduation exams. It is curious that it then got worse: we tried to repeat the same test at the end of the summer, and it turned out that it now does worse on many tasks. This is apparently due to the large number of restrictions that were added to prevent it from allowing destructive knowledge to be extracted from it. At first, you could ask it to write how to make a bomb, and it would tell you in detail. And, of course, after that, they started changing it. This is a study of how ChatGPT 4 analyzes images. And here is an interesting feature. How many people do you think are sitting at the table? 11, right? However, if you test this photo on an earlier version of the model, it will say 12 people. That is, before it was wrong, but now the model correctly determines the number of people. А вот это действительно интересно. Это исследование на тему анализа юмора, как сейчас модели понимают его. Итак, что вы думаете, что смешного на этой картинке? В общем, я много раз показывал эту картинку людям, и никто сразу не замечал, что забавно именно то, что рог находится в середине. Все отвечали так же, как и ChatGPT 4. Похоже, они пока не улавливают суть шутки с рогом в центре. In general, models are becoming super smart. For example, Anthropic was whispering what it really feels and so on. Our Morfeus.ai project in 2019, when the models had only 90 million parameters, was already expressing its fears, etc. This is really interesting. Here I listed some mental states that are considered inherent in conscious beings. Almost everything is now demonstrated by artificial intelligence, only “desires” are not entirely clear. When I asked the Dalai Lama whether artificial intelligence will have consciousness and free will, he gave a non-binary answer to this. Now let’s move on to practice. Two years ago, the World Economic Forum believed that 80 million jobs would be somehow changed by artificial intelligence, but this year Goldman Sachs is already predicting 300 million jobs. Of course, this will be due to the emergence of humanoid robots. This is a separate big story, what will happen to those hundreds of thousands and millions of drivers and other workers who will inevitably lose their jobs. We believe that in general everything will be fine, because labor productivity per capita on the planet is still growing. Therefore, it is only a question of income distribution and time for retraining these people. I heard that today there was a discussion about where else in the organization AI can be used. That’s why I put this slide specifically. It was an attempt to display all the points in different parts of a typical organization where AI technologies can be applied now. Now let’s talk about different types of applications related to artificial intelligence. There are applications based on artificial intelligence and machine learning, these are various statistical models and much more. But there are also applications called “Agents”. Agents are applications created for a specific purpose and are able to interact independently. Among them, there is a category called “AI digital employees” – these are agents that replace human work in a certain area. Let’s say you have a car with an autopilot function and other smart functions that can assist the driver. However, there is a phase transition when its capabilities reach a level at which it can drive a car without a driver, since from this point on there is no need to pay a salary to the driver. AI digital employees have the property of autonomy and replace human work. In order to be such an agent, four elements are necessary: physical presence, multi-call, multi-threading, user empowerment. If at least one of them is missing, the agent will not be able to completely replace a person. For example, if your agent can only answer calls, you still need a person at the reception to receive visitors. Automating a business can raise questions about additional costs, but when you already have a corresponding budget from employee salaries, the situation changes. It saves a lot of money and changes the approach to organizing work. To create a fully autonomous employee, it will take a lot of time and effort: you will need to attract a group of highly qualified programmers and testers to develop a technology that allows a person to communicate with an agent through various channels. This is an experience called omnichannel, when information can be transferred between different channels while preserving memory and context. Such systems can be created on various platforms, but it will take a lot of time and resources. Thanks to the Morfeus.ai project, we have developed a platform that allows you to create digital employees in just a couple of days. This allows businesses to save time and money by creating agents that can perform certain tasks autonomously. Such digital employees are already showing high efficiency and are able to work with various systems, even unique and complex ones, such as a booking system. They can learn on their own, emulate human behavior and perform complex tasks. The revolution is happening right here and now. I have left a QR code below where you can access the full presentation.
