AGI: General artificial intelligence, myth or revolution in the making?

AGI, IAG, GenAI... Are you confused? No problem.

When it comes to artificial intelligence, there's a recurring confusion, even among experts: in French, the acronym IAG can refer to two very different notions... and yet they're often confused. On the one hand, Generative Artificial Intelligence (such as ChatGPT or DALL-E), on the other, General Artificial Intelligence, that technological Holy Grail still largely hypothetical.

In English, things are a little clearer:

-AGI (Artificial General Intelligence) refers to general AI,

- while GenAI (Generative AI) refers to generative AI.

But in French, this linguistic ambiguity - IAG for both - reveals a deeper misunderstanding: we often confuse what AI does today, with what we imagine it will do tomorrow.

So, what exactly is AGI? Why are there so many fantasies and debates surrounding this notion? And above all, where do we really stand?

In this article, we untangle the truth from the false to better understand the challenges, hopes and limits of artificial intelligence.

‍

🧠 What is AGI?

AGI refers to artificial intelligence capable of performing any cognitive task that a human is capable of performing. In other words, an AI that is generalist, versatile, capable of understanding, reasoning, learning, planning, transferring skills between different domains, adapting to new contexts, and even demonstrating a certain amount of common sense. And, let's face it, once it reaches that level, it will surpass it!

Unlike so-called "weak" or "specialized" AIs (such as those used to recommend movies, sort e-mails or play chess), AGI would aim for global intellectual autonomy, comparable to that of a human. Such an AI could move from one task to another without being reprogrammed, and solve problems in fields in which it has never been specifically trained.

In other words, it would have not only skills, but also a form of flexible, contextual intelligence.

🔬 An ancient concept, a living ambition

The idea of general artificial intelligence is not new:

• 1950: Alan Turing wonders whether a machine can think. He laid the foundations for thinking about simulated intelligence, with his famous test to determine whether a machine could imitate a human in conversation.
• 1956: Dartmouth Conference, the official starting point for the field of AI. Researchers dreamed of a machine capable of imitating human intelligence in all its diversity.

But technical realities soon put the brakes on ambitions. AGI became a distant, almost theoretical idea, relegated to the status of a scientific utopia. The following decades focused on highly specialized AIs, often performing well in a single field, but unable to move beyond it.

Until deep learning, big data and advances in computing power revived this ambition in the 2010s. The ability to train models on immense volumes of data and to advance them by learning opens up new perspectives.

🔍 Where are we today?

We are not yet in the presence of a true AGI, but some recent models are sowing doubts and fuelling debate:

• GPT-4, Claude, Gemini... These AIs are capable of performing a wide range of tasks (language, coding, logic, summarizing, creating...), without being explicitly programmed for each one. They sometimes seem to improvise, reason, demonstrate logic... blurring the boundaries.
• Multimodal models: some AIs can process text, images, audio and even physical commands. This ability to handle several types of information in parallel is seen as a step towards a more "global" understanding of the world.

Some researchers refer to this as "emerging AGI" or "proto-AGI". But there is no clear scientific consensus on the criteria that need to be met before we can say "this is an AGI".

Some propose alternative tests to the Turing test, others criteria of generalization, autonomy or consciousness. But it's all very subjective.

⚠️ Controversies and fears surrounding AGI

1. The alignment problem

How can we be sure that ultra-powerful AI will pursue objectives that are compatible with our human values?

That's what AI alignment is all about. It's still a wide-open field, with many unknowns. If it's difficult to define human values, how can we translate them into an algorithmic system?

2. Existential risks

Some researchers (Nick Bostrom, Eliezer Yudkowsky) believe that AGI represents an existential risk for humanity. If it becomes impossible to control it, it could pursue its own objectives to the detriment of our own. Dystopian scenarios are regularly evoked, where a misaligned AGI would make decisions contrary to human interest... or even fatal.

Others (e.g. Daniel Andler) retort that this discourse is alarmist, even counter-productive, as it diverts attention from very real and current problems: algorithmic biases, mass surveillance, job automation, inequalities in access to technologies...

3. Societal upheavals

An AGI could have a colossal impact on the organization of work, education, creation, politics... By radically simplifying cognitive tasks, it would call into question entire professions.

It could also influence human behavior, political decisions and cultural values. And concentrate enormous power in the hands of those who control it.

4. Regulatory vacuum

No current laws or regulations specifically provide for the supervision of an IGA. Global governance initiatives are still in their infancy, and governments are struggling to coordinate their efforts. However, if an IGA were to appear in an unregulated context, the consequences could be irreversible.

📊 A blurred but central subject...

There is no universal definition of AGI. For some, it will be an AI capable of passing the Turing test with flying colors. For others, it should possess some form of consciousness, or decision-making autonomy. Some believe it should even develop emotions or a moral conscience.

But one thing is certain: the idea of an AGI is already shaping debates, guiding the strategies of major technology companies, fuelling public discourse and shaping the collective imagination.

AGI fascinates, frightens and intrigues. It raises questions that force us to redefine intelligence, consciousness, work... and even what it means to be human.

That's why, whether you believe in it or not, distrust it or desire it, it's essential to understand what AGI is, to follow its evolution, and above all to talk about it in an open, ethical and critical way.

At Learning Robots, we campaign for clear, nuanced AI acculturation that is accessible to all. The debate surrounding AI should not be a subject reserved for a technical or philosophical elite. It's a civic, collective affair that concerns us all.

Want to find out more? Follow our publications to continue exploring the challenges of AI... without falling into fantasy.

‍