The future, Enrico Ruggeri sang, is a hypothesis. And in this hypothesis there is also an artificial intelligence capable not only of beating man at chess or Go, but of writing novels, compose symphonies. Do scientific discoveries, even feel emotions. This is the mirage of artificial general intelligence, or AGI: systems with general cognitive abilities similar to and superior to those of humans. A goal that seems to be getting closer thanks to advanced linguistic models such as GPT-4 and now also ol. But there are still many theoretical and practical knots to untie.
The AGI Dream
For decades, thegeneral artificial intelligence (AGI) is the Eldorado, the promised land of artificial intelligence researchers. The idea of creating a machine capable of equaling and surpassing human cognitive abilities in all fields (from abstract reasoning to the creativity, From planning to the generalization of skills) fascinates and scares at the same time.
An AGI system would in fact be able to autonomously solve extremely complex problems such as climate change, the future pandemic, finding cures for devastating diseases like cancer andAlzheimer. It could advance science and technology at an unprecedented pace, revolutionizing areas such as space exploration, clean energy, transportation, education.
On the other hand, a super AI would have enormous and potentially uncontrollable power if not properly aligned with human values. “Terrible things could happen because of the misuse of AI or because we lose control of it,” he warns. Joshua Bengio, one of the fathers of deep learning and a pioneer of AGI research. Apocalyptic scenarios that echo those imagined by scientists of the caliber of Stephen Hawking and visionary entrepreneurs like Elon Musk.
AlphaGo and the Limitations of Current AI Systems
Until a few years ago, AGI seemed like a distant mirage. Progress in “narrow” artificial intelligence was impressive but limited to specific tasks. Consider, for example, AlphaGo, the AI system developed by Google DeepMind capable of beating the world champions of the board game Go. A historic result, which however does not make AlphaGo intelligent in a general sense: it only knows how to play Go, it cannot improvise as a writer, composer or scientist.
This is the limit of current AI systems, even the most advanced ones: they are “wise idiots”, extremely good at a narrow domain (be it playing chess, recognizing images, or generating text) but unable to transfer those skills to different domains. They lack that ability to generalize e abstract, which is the measure of human intelligence.
The revolution of language models
The recent revolution of the linguistic models like GPT-3 of OpenAI, TheMDA of Google, DALL-E (also by OpenAI), Stable Diffusion by AI stability, to name the most famous, has changed the rules of the game. Or rather, it has shown us a representation of what a general artificial intelligence would be like, closer than we thought until recently.
The feature that makes these models so promising is their multifunctionality: unlike “narrow” AI systems, they are able to tackle very different tasks, from writing texts to generating images, from conversation to problem solving, with often impressive results.
A multifunctionality that in some ways recalls that of the human mind and has made some researchers talk about “general artificial intelligence already achieved”. In particular, the recent announcement of ol, the latest model from OpenAI, which (in some cases, few to be honest) would boast capacity of reasoning and learning much more human-like than its predecessors, has reignited the debate.
The knots to be untied towards general artificial intelligence
As emphasized Francois Chollet, computer scientist and creator of the Keras AI framework, today's large language models still suffer from major limitations that make them “not good enough for AGI.”
One of the main problems is the generalization: Even trained on huge amounts of data (terabytes of text and images), these systems struggle to apply what they have learned to situations even slightly different from those on which they were trained. “Language models cannot really adapt to novelty because they do not have the ability to recombine their knowledge on the fly to adapt to new contexts,” explains Chollet.
Also connected to generalization is the notion of “learning from a few examples” (few-shot learning), one of the key features of human intelligence. Today, language models require a huge amount of data and expensive retraining to “learn” new tasks, while we humans we are often able to grasp a concept from one or very few examples.
Impossible Roads to General Artificial Intelligence
As some experiments show, the “internal representations” that linguistic models construct of reality are often superficial and inconsistent. For example, a team of researchers from Harvard trained a model on taxi routes in New York City, and it was able to accurately predict a destination given a starting point. But when researchers looked at the “mental maps” the system developed, they found that they were completely meaningless, “with streets with physically impossible orientations and overpasses that go over other streets.”
Finally, the current models still lack a mechanism for feedback like the one in the human brain, where information flows bidirectionally between different layers of neurons, enabling the integration of perception, reasoning and action. While on the one hand sensory information rises upwards to build abstract representations of the world, on the other hand those same representations can influence perception and guide the acquisition of new relevant information. A dynamic that enables key functions such as imagination, planning, formulating hypotheses to test. Current models? They are nothing like this. An artificial general intelligence (AGI) that could emerge somewhere in the near future? Maybe. Who knows.
Not at the moment. Some of these features are currently obtainable only in a rudimentary way, by “adding from the outside” to the models ad hoc modules called “verifiers” that evaluate and correct the output. But this is an approach that is not very scalable and far from the efficiency of the human mind.
Artificial General Intelligence: The Next Steps Towards the Goal
Despite these limitations, the path to truly general artificial intelligence appears clearer and more viable today. From a theoretical point of view, there do not seem to be insurmountable obstacles: "Humans and some animals are living proof that it is possible to get there", he emphasizes Melanie mitchell, a professor at the Santa Fe Institute and an expert in AGI. The challenge is more engineering and implementation than conceptual.
Researchers are already working on next-generation AI systems that overcome the limitations of current ones and incorporate key features of human intelligence such as:
- Models of the world more sophisticated, coherent and adaptable to support high-level reasoning, planning and generalization. Representations similar to our “mental simulations” with which we imagine hypothetical scenarios to make decisions.
- Internal Feedback that allow a bidirectional and recursive flow of information, where abstract representations can guide the acquisition of further relevant data and the formulation of experiments to validate hypotheses.
- Greater efficiency in learning, developing the ability to actively select which information to “taste” to refine one’s models, rather than relying on an indiscriminate mass of data. A bit like a child who actively explores his environment in search of interesting stimuli.
- A form of “metamemory”, that is, the awareness of what one knows and does not know, to guide the exploration and targeted acquisition of new knowledge.
- Separate structures for short and long term memory to rapidly store, recall, and recombine previous knowledge and experiences, as occurs in biological memory systems.
- Some form of consciousness and self-model, to guide goal-directed behaviors and effective interactions with the environment and with other agents, artificial or human.
Interesting progress is being made on many of these points.
Yoshua Bengio, for example, is working on new neural network architectures, which he calls “generative flow networks”, capable of simultaneously learning to build models of the world and the modules to use them to reason and plan. An approach that in some ways resembles the functioning of the human brain.
Other researchers, such as Jeff Hawkins at Numenta, are trying to implement the principles of biological memory in AI systems, with separate structures for short- and long-term memory and processes for consolidating and retrieving information. The hypothesis is that it could help with generalization and abstract reasoning challenges.
Then there are those, like the theoretical neuroscientist Karl Friston of University College London, proposes that one key to getting closer to AGI is to build systems that don’t just passively “ingest” training data but actively decide how much and what information to sample to achieve their goals. A principle similar to “active sampling” in living things.
The ethical and security challenges of an AGI
If building a general artificial intelligence is theoretically possible, This does not mean that it is without risks or critical issues. As mentioned, a superintelligence out of control or not aligned with human values could represent an existential threat to our species. Not to mention the possible impacts on the economy, work, inequality, privacy, and information manipulation.
For this reason, researchers such as Stuart Russell of the University of Berkeley emphasize the importance of developing techniques of “alignment of values” (value alignment) to ensure that AGI systems have goals and behaviors compatible with ethics and human well-being.
Promising approaches in this direction are those that aim to bring out ethical behaviors and values from the learning dynamics of the system itself, without imposing them from the outside. An interesting proposal is that of“collaborative inverse reinforcement learning” (CIRL). With CIRL, the AI agent is rewarded for fulfilling human preferences, but these preferences themselves are partly inferred from the agent's behavior and refined over time.
Other researchers believe that it is essential that the development of AGI occurs gradually. I rephrase: “layered”, with increasing levels of autonomy and skill, so that the safety of the system can be tested and validated at each stage. A bit like the development of the human brain, which goes through stages of increasing sophistication from basic motor and perceptual skills to language and higher cognitive abilities.
And then there is of course the issue of regulatory e governance: who and how should control the development of technologies as powerful as AGI? Only private companies and laboratories or also governments and international organizations? Many are skeptical about the ability of current regulatory and institutional frameworks to keep up, and call for a “global governance” of AI but the contours of this future governance are yet to be defined.
Artificial General Intelligence: A Fascinating and Uncertain Future
All things considered, how close are we to achieving a general artificial intelligence comparable to or superior to that of humans? It is difficult to say for sure, as experts' estimates vary. for a few years several decades. What is certain is that the progress of recent years has made AGI a much more tangible goal and has given the field an impressive acceleration.
The road is full of obstacles, but the goal has never seemed so close. And if one day, hopefully not too far away, an artificial intelligence will be able to ask us questions, test hypotheses, learn from experience and even feel emotions… then we will know that we have arrived. And that the world, for better or for worse, will never be the same again.
