The year 2024 was marked by significant adoption of generative artificial intelligence (AI) within businesses. The 2024 AI Index report from Stanford University offers a comprehensive analysis of current trends in AI, covering various aspects such as research, development, economics, education, policy, and governance.
According to this AI Index report, 80% of decision-makers have experimented with generative AI, and 20% regularly integrate it into their professional activities. Additionally, one-third of companies now depend on this technology for specific tasks.
AI has surpassed human performance in several domains, including image classification, visual reasoning, and English comprehension. However, it still lags behind on more complex tasks such as advanced mathematics, common-sense visual reasoning, and planning. The training costs for cutting-edge AI models have reached unprecedented levels.
The report highlights a significant lack of standardization in responsible AI evaluations. Major developers, such as OpenAI, Google, and Anthropic, primarily test their models on different benchmarks, thus complicating systematic comparisons of AI model risks and limitations.
Despite an overall decline in private AI investments the previous year, generative AI funding nearly increased eightfold, reaching $25.2 billion. Major players like OpenAI, Anthropic, Hugging Face, and Inflection secured substantial funding rounds.
Generative AI, capable of generating original content from existing data, has transformed various sectors. Companies use it to automate content creation, enhance customer service through advanced chatbots, and optimize product design. This adoption reflects recognition of the competitive advantages offered by generative AI.
Limitations of Current Approaches and Yann LeCun’s Vision
Despite these advances, critical voices are emerging regarding the limitations of current AI approaches. Yann LeCun, Chief AI Scientist at Meta, considers generative AI, particularly large language models (LLMs), to represent a technological dead end. According to him, these models, while impressive, do not possess a deep understanding of the world and merely reproduce learned patterns without genuine contextual intelligence.
To overcome these limitations, LeCun proposes an approach called “Objective-Driven AI.” This method aims to develop AI systems capable of understanding the world, remembering, reasoning, and planning—skills that current LLMs have not mastered.
Objective-Driven AI draws inspiration from human learning, where individuals train themselves by observing and interacting with their environment. LeCun suggests that machines should be trained similarly, using data from various sources, such as videos, to predict missing parts and thus develop a deeper understanding of the world.
This approach requires a cognitive architecture composed of multiple modules, including those for perception, short-term memory, associative memory, strategy evaluation, and action. World modeling is at the heart of this architecture, enabling AI to predict the consequences of its actions and plan accordingly, thereby imitating human reasoning.
Yann LeCun advocates for an evolution of AI toward objective-driven systems, capable of learning and adapting autonomously, surpassing the limitations of current models based primarily on language processing.
The transition to objective-driven AI poses several challenges. It requires significant advances in self-supervised learning, world modeling, and hierarchical planning. Moreover, this approach demands a thorough understanding of human cognitive mechanisms to be effectively reproduced in artificial tools.
