#86 Inteligencia Artificial: Nuestro Co Piloto del Futuro, con Marc Riera

Thank you very much, Federico. It's a real pleasure. I feel at home chatting with you. We are going to touch on topics that, as you say, are not at all in vogue, right? No one is interested in them, but we are going to talk about it. Exactly, exactly.

Exactly, Federico. Phasing it is difficult with the speed it's reaching. I would classify AI into generations. As you mentioned, machine learning technology has existed since the 80s and 90s, used initially for classification and identification. Then came the next generations, like Siri and Alexa, acting as small agents to solve problems and execute mini-actions. But the difference with this new generation, let's call it third-generation AI, is the critical mass of data they have been trained on and the improving methods over time. When we implemented neural networks, we made a huge leap, but we needed data curation: selecting, unifying, structuring data and training models. That's what has been done in recent months by companies like OpenAI, Microsoft, Meta. Now we have LLMs, large language models, that have compiled all of humanity's knowledge in a structured way, combined with supervised learning where humans intervene. The only difference between previous generations and this one is the structuring of data, its quality, and the critical volume. We realized that going from 10,000 data points to 10 billion parameters was an exponential change. Siri and Alexa are limited agents programmed not to stray. But I have no doubt that those products will acquire this third-generation intelligence. The question is when. Right now we are touching philosophical issues, moral limits, bias. The great work these companies are doing is laying the ethical foundations for AI. We have previous-generation technologies coexisting with new ones, and it's a quantum leap.

It's based on predictions. Since working with words is complicated, these models tokenize words, converting each word into a number. Then you have a dictionary of words to tokens. With numbers, you can make predictions. You give me an input text, I convert it to numbers and compare with a database of numbers. It predicts what comes next. Models like ChatGPT have an extra training layer to act as a chatbot. Otherwise they are autocomplete. For example, if you say 'Hello, how are', the most probable next word is 'you'. The models compare what you entered with what could come next. OpenAI added a chat layer, converting an autocomplete model into a chatbot that predicts based on your request. If you ask 'Who is the president of Mexico?', it looks for data related to 'president' and 'Mexico'. It's a prediction within natural language. It works, it's intelligent and predictive. Sometimes it has more or less context depending on data quantity and curation. That's basically it.

You're familiar with the Turing test, right? And there's John Searle's Chinese Room argument. To quickly explain: The Turing test says if you communicate with an entity and don't realize it's a machine, it's intelligent. The Chinese Room is a thought experiment: suppose I put you in a room with Chinese symbols and give you instructions to respond to certain symbols with others. People outside might think you understand Chinese, but you don't. So my question is: Where do you think we stand? One can have almost metaphysical conversations with ChatGPT if you take the time. I think the Turing test, as originally conceived, made sense for previous technology. With ChatGPT, the method should be different. We get into murky waters: Is something that perfectly imitates intelligence but isn't intelligent? If a system behaves exactly like a person but doesn't understand, do we call it intelligent or an automaton? I think we need to compare neural studies in the brain with these computer models. My personal opinion is that we are talking to stochastic parrots. But what may differentiate a model that feels human is the critical amount of data. That's what happened with the generations: from first to second was a quantum leap, and the third has been universal. Now these tools are becoming interesting for all humanity. I'm practically convinced we haven't created life, but we may need to redefine concepts of intelligence.

In education, this technology can be used as a data consultation tool, like Encarta or Wikipedia. It can be a great boost for students. Structural education will take more time, but if we restructure and re-focus using these tools, education can become creative and personalized. In the workplace, we will be accompanied by copilots in all interfaces: email, PowerPoint, web pages. You can ask questions about your email, meeting summaries, suggestions. This affects the client end as well, offering more creative and dedicated service. But first we need an internal exercise to understand the technology before going external. In my sector, tourism, we are using AI internally, then applying it to customer interaction. I think we should be cautious: use these tools to gather suggestions, but human interaction should remain co-piloted. We shouldn't lose human touch. We will see our daily tools accompanied by copilots.

The criterion of each model depends on the brand behind it. But for jobs that may emerge, the so-called 'prompt engineering' is important: knowing how to construct queries correctly to get desired results. Since these models are stochastic parrots, the ability to transmit what you need is key. Also, RPA (Robotic Process Automation) can now be AI-enabled. Automating processes with a layer of intelligence gives analysis and conclusions. So two interesting roles: automation within organizations for faster decisions, and front-line translators who can 'translate' to LLM language - not engineers but ingenious in crafting prompts. For example, a financial analyst will know what specific questions to ask their spreadsheet. In marketing, data analysis will be very important, but we must be cautious about biases in the models.

I don't like to set things on fire. Everything changes quickly. As I said in my previous appearance on your podcast, we will co-pilot. We will live better, work more creatively. The question is until when? We will start supervising processes, gaining quality of life. But eventually, tasks susceptible to automation will be automated. We may need to rethink everything. In Europe, there is talk of universal basic income. That could be an answer, but it depends on the technology state. In the short term, five to ten years, we will co-pilot, be happier, solve 90% of daily tasks. In 20 years, I can't say how we will manage as humanity. It could create gaps between countries. Technological level will be marked by AI, quantum advances, genetics. Not all countries will have the same capacity. The solution might be global or local. Even within Spain, there are differences between the service-based south and industrial north. Competing with the US, South Korea, Japan, which are technology-savvy, Europe may be more moderate. That gives us security but a disadvantage in evolution.

In such technical fields, political regulation does more harm than good. Technical regulation would be much smarter. Italy banning was more of a political tantrum related to data protection laws. OpenAI didn't have time to set up servers in Europe as required. The privacy issue is an excuse to slow down and let Europe catch up, or maybe due to ignorance. Regulation should be technical. If we play with technology Europe doesn't truly understand but tries to use hypocritically, we could end up ten years behind the rest. Being ten years behind industrially is one thing, but in AI it's a transversal impact on intellectual level, decision-making quality. These companies serve their own environment; we might get a dumbed-down service. We may have models that don't let us advance as fast as others.

You've found a very techy person, Federico. I see any technological limitation as something coherent, given proper use. My whole life has been in this sector. A writer might have a different opinion. My work relies on technology. For a writer, this technology can help or not. The interesting thing is to see, in months or years, how human-supervised publications compare to those fully generated by AI. Imagine a newspaper entirely generated by AI. It could create bias, manipulation, fake news, or it could be very neutral and accurate. Neutrality is rare. We will solve many problems efficiently but generate others we didn't expect so soon, which were already in philosophy books.

In the next five years: co-pilot, supervise, and increase creativity. Humanity will perceive it in selling products, offering services, communicating. We will see companies run or co-piloted by AI, like a newspaper that is just a website fed content. These online business models will be the first adopters. In seven years, we might see crazy things like hotels managed by AI. In Japan, there are hotels with robot receptionists. If we give them these models, it becomes more coherent. In my sector, tourism, we could see hotels managed this way. Already on cruise ships there are robotic bartenders. Why not in five or seven years see vending machines, hotels, autonomous driving with natural language interaction? You'll talk to them, they'll know your usual Martini and ask about your wife. It's combining sequential mechanical processes with AI context for personalized service. We are at a point where I'm just using a crystal ball. It's moving at the speed of light. New papers come out every week. We haven't even touched multimodality, combining multiple AIs to generate images, avatars, real-time interaction. When that goes into robotics, the same hardware can become a new generation just by connecting to AI. The next five years will be very interesting.