Reinventing AI: Is It the Time for a New Paradigm?

Condemned to the sole intelligence that emerges from brute force?  Why should we imagine another science fiction movie? Haven’t we seen enough of them already? Is not the concern that arises from machines becoming aware of the world they live in to launch the ultimate challenge to the human race already enough? Doesn’t the hybridizations resulting from the marriage between biological and artificial neurons contribute enough to foresee dystopian scenarios that inject us with abundant apprehension? When science walks alongside science fiction it is sometimes not easy to separate reality from fantasy. However, it is very clear that since the dawn of the Web, the ideas of information retrieval from the 1970s have taken place in the large Web companies that have activated splendid information distribution services offered free of charge to the entire world. The other side of the coin is that those now indispensable services, that also contribute in a fundamental way to the growth of science, are based on the public heritage of Web data swallowed in their own belly (see Witten et al.⁹ for an early discussion on two sides of the coin). Something perfectly legitimate that, however, raises some legitimate concerns about the consequent centralization of power. The development of AI in the era of deep learning and transformers evokes a sense of instant déjà vu, perfectly mirroring what has just been described for search engines.

Personal agents and privacy issues.  Perhaps there is something even more revolutionary in the services that LLMs and foundation models can offer with respect to information search of search engines. Without particularly explicit announcements, intelligent agents already live with us and are capable of assisting us in many daily tasks. But they are also potentially in possession of our requests, our inquiries, our desires. All this seems to some an unacceptable invasion of our privacy, but the reverse of the coin is that we benefit of fruitful consultations! Perhaps it makes sense that each one is individually free to choose where to place themselves in front of this dichotomy. Interestingly, this is not the only possible perspective; a new community of agents that live for example in smartphones could become protagonists of a new way of manifesting themselves as personal assistants. They could live with us, sensorially participate in all the information in which we decide to insert them and assist us without exchanging everything that we consider confidential to the outside. This does not mean that the current “personal agents” already offered will no longer be there. It will certainly not be possible to compete on the extraordinary characteristic that they possess of universal knowledge that, precisely, derives from access to the entire Web. Two very different types of personal services could therefore be combined, chosen according to the specific conditions in which the consultancy is required. In this case, however, this new “social intelligence” would be on board the devices, not in the belly of the dragons.

Symbolic AI is collectionless.  The proposed scenario seems to belong to world fiction movies. How can that movie leave the screens of movie theaters and land in reality? Is a direct new declination of the methods we already have enough? Is only a new strong technological shock necessary, or is a profound rethinking of AI methods also needed? The cumbersome role of machine learning and neural networks in the last decade should not make us forget the extraordinary cognitive qualities that emerge from many artificial intelligence methods: The world of puzzles that relies on problem solving, the rich application domain covered by methods of constraint satisfaction, inference from knowledge bases, and automatic planning does not require data accumulation and bathing! Indeed, the seminal textbook by Norvig and Russell⁶ reminds us that symbolic AI methods are collectionless, whereas current machine learning methods do not seem suitable to be acting in our movie.

Current machine learning is collection-based and computationally cumbersome.  Can the movie director rely on the ideas of current machine learning? It soon becomes clear that we are faced with abysmally different cognitive processes, precisely because they are conceived in different universes. Let’s imagine using current machine learning methods: agents should learn online, without any possibility of conducting sophisticated pre-processing and without the possibility of randomizing data, a fundamental requirement for the success of SGD-based algorithms. Current LLMs also actually process online, but often after significant pre-training processes and, in any case, relying on architectures with huge time windows. This last requirement is definitely out of the movie director’s budget. Even if we imagined migrating modules built with learning with current learning methods on board the agents, we would face the director’s budget problems. Current machine learning is data- and computation-hungry and any attempt to inherit the current framework impacts at least in energy consumption problems.

Time is the protagonist of learning.  Machine learning took off under the umbrella of statistics, which has definitely favored the current undisputed assumption of separating the learning process from the testing process. This separation finds its perfect place in statistics. However, if you learn while you live you face a new world where we immediately realize that the process of evaluating learning intersects daily with learning itself. In the movie a new protagonist appears: Time! Every environmental perception, every action is marked by time just like it happens to humans and every other biological species in nature. Time produces something different from collections of examples. In a video, for example, time “takes pictures.” However, the frequency with which it takes them appears to be inextricably linked to the learning processes themselves. To produce meaningful examples one must presumably take pictures only when the video changes in some way, while it is certainly not useful to focus attention for too long on still images. So, time produces sequences, but those same sequences, to be cognitively relevant, are the fruit of a precious focus of attention mechanisms. The methodologies at the basis of learning can still be driven by optimization theories, but one must bear in mind that, while in statistical machine learning we are relying on averaging the errors on the basis of probability distributions that are estimated from data collection, our movie imposes an estimation over time. Interestingly, while we can process online information and potentially take into account any past event, we cannot access the future, an issue which was very well described by Danish theologian, philosopher, and poet Søren Kierkegaard: Life can only be understood backward; but it must be lived forward. Interestingly, this intersects with the longstanding dilemma of biological plausibility underlying neural propagation mechanisms. In 1989, Francis Crick published a seminal work² in which he dampened the enthusiasm for the growing connectionist wave emerging from algorithms such as backpropagation by clearly highlighting their lack of biological plausibility. It is not in fact only a question of biological emulation, but from the perspective of computer science the conquest of algorithms that operate locally in space and time is a question of computational efficiency. Intelligence on small devices working continuously in the environment do require to face Crick’s challenge seriously. This likely requires a paradigm-shift⁴ in the training schemes that govern the neural propagation. This might be a nice field to look for “survival strategies for depressed AI academics.”⁸

Developmental learning.  What is perhaps most surprising about current machine learning schemes compared to what happens in nature³ is that we cannot speak of a real development of current artificial agents. It is more properly learning in the laboratory and the interaction in the world that produces at most refinements, but not substantial modifications. How do the agents in the movie enter the stage? How can we profitably model the precious social interactions that can take place amongst agents and humans? We might need to model those relation following the spirit of early graph neural network models,⁷ where also learning mechanisms on a single graphical domain were proposed, but this time we need to restrict those mechanisms to the local objectives of the agents. Moreover, they must be likely very different with respect to what has been primarily proposed in machine learning. The agent in our movie experiences a virtual life and poses new challenges. They can born during the movie and begin the process of learning from an initial configuration. However, how can already mature agents, therefore already concretely active in the environment, learn? Most importantly, what is their neural structure? Here the movie takes an additional dose of stimulation from nature and truly becomes science fiction! We can think of generations of agents that are born and die. Death, after all, can be an excellent idea to regenerate new agents, perhaps from genetic crosses and transfer learning. Interestingly, when thinking of the recent decades of investigation in the field of evolutionary computation¹⁰ we realize that the mentioned evolution of the movie might not be a matter of fiction only.

Which tasks could be faced.  How could the distribution of intelligence to the unspecified number of billions of CPUs available on the planet be useful? How could the writing of algorithms operating mainly at the meta-level of learning re-direct many services? What tasks will those agents be able to perform in addition to that of the personal agent? It is difficult to limit the imagination, but it is instead simple to understand the impact of this approach in the entire world of robotics. Robots could live in their environment progressively gaining cognitive abilities from environmental interaction and could undergo a real cognitive development that fits perfectly with the concept of “developmental robotics.” The development of robots would take place by relying exclusively on environmental interactions, exactly as happens in nature. They could learn to see, to speak, to manipulate objects, they could learn to act because of their life in the environment where we desire their presence and their action.

Actionable insights and technologies.  The aforementioned evolutionary computation style has already been shown to be successful when adopted in conjunction with LLMs for discovering solutions to hard problems.⁵ It is shown that LLM can be seen as a source of diverse programs with occasionally interesting ideas, which can concretely take off in the framework proposed in this Opinion column, where we emphasize the social component of billions of small intelligent agents. An actionable insight that is coherent with the view proposed here also arises from Behrouz et al.,¹ where long-term neural memory modules can efficiently and effectively learn to memorize at test time. It is pointed out that we need an online meta-model that learns how to memorize/forget the data at test time. In this setup, the Behrouz et al.¹ paper provides evidence that the proposed system, called Titans, can learn a function that is capable of memorization, but it is not overfitting to the training data, resulting in a better generalization at test time.

A research direction consistent with the perspective proposed in this Opinion column is described at https://collectionless.ai, where the idea of a world of agents whose intelligence emerges solely from interactions is promoted. Apart from the methodological foundations of learning processes, the aspect that is technologically most relevant is how to concretely implement environmental interaction. Instead of cultivating the development process of the agents described in this Opinion column directly through interactions with the real world, current research is investigating the creation of artificial environments based on LLM technologies and foundation models. This approach allows for an acceleration of the developmental mechanisms of the agents and their generations described in this Opinion column.