The ultimate goal of companies like OpenAI or Meta is to achieve artificial general intelligence, or AGI. Even though this field has experienced a true revolution in the last year and a half, there’s a big difference between creating chatbots like ChatGPT and reaching, or surpassing, human intelligence.
Large language models (LLMs) won’t reach AGI status. In a recent interview with the Financial Times, Meta AI chief Yann LeCun explained that LLMs will never be capable of reaching human levels of planning and reasoning.
Limited by their training. The expert pointed out that, although LLMs are trained with massive amounts of data, their responses are limited by the nature of that data. If the data is accurate and precise, the responses will be, too. However, these data sets include both valid and invalid data. A recent case where Google’s AI suggested gluing cheese to a pizza is one of many examples of this.
AI models can’t reason. LeCun said that LLMs have a “limited understanding of logic,” lack persistent memory (although this is improving) and don’t understand the physical world (progress is also being made in this area). For him, the key is that LLMs powering chatbots like ChatGPT “cannot reason in any reasonable definition of the term.”
If we want AGI, we'll have to look elsewhere. LeCun explained that he and his team of about 500 engineers are working on a new generation of AI systems based on an approach called “world modeling,” where systems are built by trying to understand the world as humans do, developing an idea of what happens if something changes in a certain scenario.
No AGI for (at least) 10 more years. Trying to reach levels of human intelligence with machines will take time. For LeCun, reaching that AGI level with the world modeling approach might take a decade.
Of course, Meta has its chatbot, too. A few months ago, Meta introduced its own chatbot, Meta AI, to compete with Google, OpenAI, Microsoft, Anthropic, and the rest. The recent launch of Llama 3 has improved the company's chatbot.
A paid version of AI... for its future glasses? The Information has revealed that the company is planning to launch an advanced, paid version of Meta AI. With that, Meta would be following the footsteps of others like ChatGPT Plus or Gemini Advanced. However, its chances of winning that battle lie in other sectors. First, in its huge social platforms (WhatsApp, Instagram, Facebook), and second, in the potential integration of the AI into its smart glasses—a promising and exciting format for this kind of application.
Image | O’Reilly Internal
Related | Generative AI Creates a Gap Between Local and Cloud-Based Models. There’s Room for Both