Artificial intelligence (AI) emerged in the 1950s when a group of visionaries wondered if machines could "think." Seventy years later, reality has surpassed any expectations: neural networks have conquered cognitive tasks that, for hundreds of thousands of years, were exclusive to living beings. Today, this is an indisputable fact. Neural network-based machine learning has solved challenges that seemed unattainable for machines: - Models like ChatGPT, Gemini, or Claude dominate natural language. - They possess encyclopedic and fluent knowledge. - They program code at a level superior to humans. - They describe images and transcribe audio with human-like accuracy. - They translate texts with quality comparable to that of a professional translator. - Other systems generate realistic images, predict weather phenomena, defeat Go champions, or drive autonomous vehicles. Researcher François Chollet summarizes it: "In the last decade, deep learning has represented a technological revolution." Each advance would be remarkable on its own; achieving them all with the same technique is like finding a master key for multiple doors. Why now? Three factors converged: algorithms, computing power, and large volumes of data. Behind each are key individuals: Geoffrey Hinton championed neural networks when others dismissed them; Jensen Huang, CEO of Nvidia, drove the development of high-performance chips; Fei-Fei Li created ImageNet, an unprecedented image database. In 2012, Hinton's students, Ilya Sutskever and Alex Krizhevsky, combined these elements and created AlexNet, a neural network that revolutionized image recognition. The success spread quickly: the formula of networks, data, and massive computing worked. The impact is profound. As Ethan Mollick points out, even if AI development stopped today, "we would have a decade of changes ahead in all industries." The future of these machines is uncertain. Between unbridled enthusiasm and skepticism, the essential is forgotten: current models are already astonishing. Their functioning resembles human intuition and raises deep questions about intelligence, both artificial and biological. Lesson 1. Machines can learn The most evident and least discussed lesson: machines learn. James Watt's centrifugal governor (1788) already adjusted the speed of steam engines without human intervention. Traditional programming defines explicit rules; machine learning, on the other hand, allows the system to discover rules from examples. As Chollet explains in Deep Learning with Python: "A machine learning system is trained, not programmed." Large language models, like ChatGPT, are neural networks with billions of parameters that are adjusted during training. The process is mathematical and has proven effective. Here arises the "bitter lesson": for decades, experts tried to encode knowledge into machines without success. What worked was creating the conditions for knowledge to emerge on its own. Lesson 2. Emergent skills in AI Complex capabilities can arise from simple processes, as occurs in biological evolution. Large language models handle language flexibly, detect sarcasm, and respond to changing contexts, without anyone explicitly teaching them grammar or irony. The training of "predicting the next word" turned out to have unexpected emergent power. To predict words, the model must grasp complex notions: geographical knowledge, mathematics, common sense, and empathy. Thus, a seemingly simple task contains all the others within it. For example, models like Gemini can solve unseen mathematical problems during training, implying the inference of complex algorithms from examples. Lesson 3. AI learns like a "rudimentary evolution" Unlike humans, who learn with few data, AI models require millions of examples. The process resembles evolution: tiny, repeated changes generate complex capabilities. As Andrej Karpathy notes, training a model is a "shoddy evolution," where intelligence emerges from the accumulation of knowledge on a large scale. Lesson 4. Automation of cognition Chollet prefers to speak of "cognitive automation" instead of artificial intelligence. Current models do not possess cognitive autonomy, but they automate cognitive tasks on a large scale. Although they lack deliberate reasoning, massive experience compensates for many of their limitations. Other experts, like Andrej Karpathy and Blaise Agüera y Arcas, believe these systems exhibit genuine forms of intelligence, capable of generalizing and learning actively. Lesson 5. More intuitive than rational Contrary to the classic image of the rational robot, current AI operates more intuitively, similar to the "System 1" described by Daniel Kahneman: fast, automatic, and pattern-based. Deliberate reasoning remains a challenge, so recent advances seek to equip models with reflective and step-by-step reasoning capabilities, improving their performance on complex tasks. Lesson 6. Humans are also patterns AI's success in capturing patterns leads us to question how much of our own abilities function similarly. Science has been dismantling the idea of our exceptionalism, and AI demonstrates that many human capabilities can be replicated through large-scale pattern learning. Lesson 7. A Cambrian explosion of AI The limitations of current AI are notable, but the combination of networks, data, and computing has opened an era of accelerated innovation. Laboratories around the world are exploring new directions: adaptive systems, models that understand the physical world, or AI that write and evolve their own programs. The potential is enormous, and while the future is uncertain, the transformation is already underway. The question is no longer whether intelligence can be replicated, but how far we will go. What we have already achieved is extraordinary and could be the greatest transformation of our era. Source: elpais.com