Google I/O 2024

The Rise of Multi-Modal Reasoning Models

The current generation of artificial intelligence has moved far beyond simple pattern recognition. Major players including Meta and Mistral have introduced architectures capable of much more sophisticated tasks. These models can now perform multi-step reasoning, generate intricate code, and analyze massive documents with ease. Meta’s release of the Llama 3.1 405B model in July 2024 provided a high-parameter open-weights alternative to proprietary systems, while Mistral AI has continued to iterate on its large-scale reasoning capabilities. These advancements are driving massive infrastructure investment; in its July 2024 earnings report, Microsoft noted a significant increase in capital expenditures dedicated to the build-out of data centers to support AI workloads.

We are also seeing a shift toward multi-modal capabilities, where a single system can seamlessly combine text, image, and voice. For instance, Le Bonhomme Picard highlighted how tools like Gemini can transform a basic text description into a fully realized image in mere seconds. This capability extends to video; the recent emergence of Google’s Veo 3 has flooded social media with increasingly realistic synthetic footage.

However, this technical proficiency often creates a psychological illusion of sentience. While these systems appear to “understand” their instructions, their internal logic remains purely mathematical. According to Gaming Campus, even AI models that can “think out loud” lack any inherent subjectivity or intention. The core of these modern systems is statistical and probabilistic rather than cognitive.

• Multi-step reasoning: Solving problems through sequential logic.
• Code generation: Writing and debugging complex software.
• Document analysis: Processing and summarizing long-form text.
• Multi-modal integration: Combining text, voice, and visual data.

The EU AI Act and Global Regulation

As the distinction between human-like output and machine-driven statistics blurs, governments are moving to establish boundaries. The era of unregulated experimentation is closing. The European Union, the United States, and China are all actively developing frameworks to manage the risks posed by these technologies. In the United States, the regulatory landscape is currently guided by President Biden’s Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, signed on October 30, 2023, which mandates that developers of powerful AI systems share safety test results with the federal government.

The EU AI Act stands as a primary example of this regulatory shift. The legislation does not treat all AI equally; instead, it categorizes applications into three distinct risk levels. The Act, which received final approval from the EU Council on May 21, 2024, establishes a compliance timeline where bans on “unacceptable risk” applications—such as social scoring or certain biometric categorization—are expected to take effect by early 2025. This tiered approach imposes specific, heavy obligations on systems deemed to have a “high impact” on society, ensuring that the most powerful models are subject to the strictest oversight, with high-risk mandates set to be enforced over the coming years.

Projections for AGI Emergence

The most debated concept in the industry remains Artificial General Intelligence (AGI)—the hypothetical point at which an AI can perform any intellectual task a human can. While the hype is constant, the actual timeline for such a milestone is a subject of intense scientific disagreement. The debate often centers on the validity of “scaling laws,” the theory that increasing computational power and data volume will inevitably lead to emergent intelligence. While OpenAI CEO Sam Altman has suggested the transition to AGI may be approaching, other industry leaders remain skeptical of current trajectories.

Current expert estimates suggest AGI could emerge sometime between 2027 and 2035. However, these dates are far from certain. Yann LeCun, Chief AI Scientist at Meta, has argued that current large language model architectures lack the fundamental “world model” required for human-level reasoning, suggesting that a new architectural paradigm is necessary. The rapid development of embedded AI—technologies integrated into autonomous vehicles, military drones, and personal assistants—suggests that while AGI might be years away, the practical, specialized impact of AI is already deeply embedded in our physical infrastructure.

Michel Cabaret’s Upcoming Paimpol Lecture

The tension between AI as a tool for progress and AI as a potential societal threat will be the central theme of an upcoming public forum in Brittany. On June 9, the Rotary Club of Paimpol will host a conference at the Ciné Breiz to explore these exact questions.

The event features Michel Cabaret, the director of the Espace des Sciences de Rennes/Morlaix, who is widely recognized for his work in scientific communication. As Actu.fr reported, Cabaret’s goal is to bridge the gap between complex research and public understanding, helping citizens navigate the debates surrounding AI’s role in business and daily life.

The lecture is scheduled for 8:00 PM on Tuesday, June 9. Admission is set at 5€, with proceeds supporting the Rotary Club’s youth programs. For a society grappling with the implications of machines that can reason, code, and create, such discussions represent a necessary step in moving from technical awe to informed civic engagement.