How kids can make AI truly intelligent and avoid a Matrix-like scenario

Artificial intelligence has made huge breakthroughs, but at the cost of energy-hungry training methods based on massive data collection. Yet data centres already consume as much power as entire countries, and experts warn: "We're heading towards a Matrix-like future, where people will have to generate energy just to keep the machines running.

We might not be trapped in pods yet, and wired into the circuitry of The Matrix, but if we swallowed the red pill as Keanu Reeves did, we might realise that our reality could soon resemble the 1999 blockbuster that made him famous. "AI has made great breakthroughs over the past years, but it has all only been possible due to massively scaling AI models up," says Hector Gonzalez. "Today, data centres consume as much energy as entire countries, and we are not too far from a future where all globally produced energy will be needed to satisfy their demand. According to experts, we are heading towards a Matrix-like scenario, where people will have to generate energy just to power the machines."

Hector Gonzalez is the co-founder and CEO of SpiNNcloud, the company behind the world's largest brain-inspired supercomputer: an eight-rack system with more than 5 million processing elements embedded in over 30,000 microchips, designed to perform extremely compute-intensive tasks. Yet, more than its impressive numbers, what could prevent us from becoming enslaved by machines like in the Hollywood saga is its neurological inspiration.

"The human brain is the most advanced supercomputer in the universe," explains Gonzalez. "It can do things that artificial intelligence can today only dream of. Yet it's very lazy and only consumes as much energy as a light bulb." One of the characteristics that makes it so energy-efficient is that it doesn't rely on a single processing element, but on over 85 billion neurons. Its workload is therefore split among many different units that are never active all at once, but always in "a very strategic way." He notes: "We took inspiration from this large number of processing elements, and from the way communication works within the human brain, in order to design energy-proportional AI systems."

If Gonzalez's SpiNNcloud today markets its systems to high-performance computing companies and collaborates with the U.S. Department of Energy, it is also thanks to the funding and support it received from EIC Communities, a European innovation programme aimed at identifying, developing, and scaling breakthrough technologies. As programme manager for artificial intelligence, Hedi Karray oversees a portfolio of projects at different levels of technological readiness. "Good ideas and good technologies alone are not enough to make visionary projects real," he says. "We provide these projects with a background from academia and research, but also with advice and networks. My role consists of a strategic follow-up to make sure they deliver their impact, because what we want to build are AI champions."

One of these is a company that he describes as a "success story of European funding." Emerging as a spinout of the Human Brain Project—an EU initiative that promoted a new paradigm in brain research at the interface of computing and technology—it works on neuromorphic infrastructures to optimise AI models and cut their energy consumption. The company's name is Superintelligence Computing Systems (SICS), and Karim Nouira is its CEO. "Today's AI does not rely on structured world models. It's based on mathematical ones, very unfit to capture the inherent structure of the information needed. It's a 'brute force paradigm' that current approaches try to compensate for with large models containing trillions of parameters and ever-growing amounts of data," he says.

Hence, the idea to replace this "brute force" with an AI with training costs "virtually close to zero," not based on reinforcement learning but on how children learn. "Current AI systems scrape up all the data on the internet, shuffle it, and feed it to the model, but this is not how we humans learn," Nouira explains. "Children learn about object permanence at six months of age. Between six and nine months, they discover how gravity works, and so they gradually build up their world model and develop behaviours requiring dexterity and more complex cognitive functions."

From these principles, his company has developed a universal AI-robot brain capable of controlling any type of robot, regardless of size, shape, or function. The first pilot customer to test it is an e-commerce company where, upon the arrival of each order, the AI controls an industrial robot that picks and ships the right product. "For traditional robotics it's a very challenging environment, as warehouses may contain hundreds of thousands of products, with new ones being added and piled up each month, adding further complexity," he says.

If this first use case mainly addresses the need to introduce automation in unstructured and complex environments, what inspires SICS is the broader goal of overcoming some of key limitations of artificial intelligence as we know it: "The core problem with today's AI is that it's neither truly intelligent nor reliable. And this severely limits its application. Our system is not only more robust, it also provides access to information, thus ensuring more transparency," Nouira adds.

For users, transparency means above all understanding processes and reasoning – in other words, the "why" behind its answers. This is one main problem that "semantic AI" or "cognitive AI" seeks to solve. "If you ask an artificial intelligence system what an apple is, it can provide you with all possible definitions, but it will never truly understand what it is," explains Karray. "This is because it's based on prediction, which means it essentially does statistics. However, semantic AI is based on knowledge representation and aims for AI systems to develop a real understanding of things. So, if you tell them: 'Okay, give me a red fruit with this and that characteristic,' it gives you an apple, because it does know what it is."

If transparency and trust are two of the goals of the AI of the future, reducing the energy consumption of its infrastructure could significantly help achieve another one: democratisation. "For artificial intelligence to be a force for good, it's important to keep it geographically decentralised and to lower its operational footprint. Since not every country can generate the energy needed to power big data centres, this would help make it more accessible and prevent monopolies," Gonzalez explains. "We need to be driven by a vision of AI for good, as a real ally for people, not as a tool that can be turned against humanity. We already have regulations setting out these principles, but we must also ensure that the whole lifecycle of AI is responsible—from how we build it, to how we use it, and how much energy it consumes," Karray stresses.

Beyond the technology, he adds, it is crucial to study the impact this ongoing revolution will have on the world system as a whole: "Let's imagine AI were deployed everywhere. How would our society, our economic system, and our democracy look? All of those questions remain open, and they are something we must definitely focus on."

Coordination EIC Communities Project:

• Stefania De Santi (APRE)—desanti@apre.it
• Jessica Bonanno (APRE)—bonanno@apre.it

Communication Officer:

• Cesar Giovanni Crisosto (ICONS)—cesar.crisosto@icons.it
• Caterina Falcinelli (ICONS)—caterina.falcinelli@icons.it

Project website: https://deepsync.eu

LinkedIn: DEEPSYNC

Provided by iCube Programme