OpenAi and Meta Prepare to Release AI Models Capable of Reasoning and Planning
Despite the significant advancements in artificial intelligence (AI) in recent years, there remain areas where machines still lag behind humans in capability. However, this may be on the verge of change, with reports suggesting that OpenAI and Meta are nearing the release of AI systems capable of reasoning and planning.
Leaders from both companies have hinted at the imminent launch of their latest AI models, promising unprecedented levels of power. According to their statements, the upcoming ChatGPT-5 and Llama-3 are poised to transcend mere text generation and exhibit qualities akin to human thought processes.
Joelle Pineau, Meta’s vice-president of AI research, emphasized the ongoing efforts to imbue these models with reasoning and planning capabilities, stating, “We are diligently working on enabling these models not only to converse but also to reason, plan, and retain memory.”
Meta has announced plans to share early versions of its Llama 3 model soon, alongside a comprehensive collection of models slated for release throughout 2024. Meanwhile, OpenAI executives have disclosed intentions to unveil GPT-5 in the near future, as reported by the Financial Times.
Presently, applications like ChatGPT utilize Large Language Models (LLMs) to anticipate the next words in a sentence. Through extensive data training, these LLMs excel at predicting word sequences, often exhibiting intelligence-like behavior. However, Yann LeCun, Meta’s chief AI scientist, acknowledges that even the most advanced AI systems can still err, remarking that they often generate responses without thorough consideration or planning.
Speaking at a Meta event in London, LeCun underscored the importance of imbuing AI with reasoning capabilities, enabling it to deliberate among potential solutions, strategize actions, and develop a conceptual understanding of their consequences. Similarly, Brad Lightcap, OpenAI’s chief operating officer, expressed optimism about the untapped potential of AI models in reasoning, foreseeing their evolution toward handling more intricate tasks in a nuanced manner.
LeCun revealed Meta’s ongoing endeavors to develop new AI “agents” capable of orchestrating and coordinating every aspect of a journey, from Paris to New York. These advancements signify a significant leap forward in AI capabilities and could potentially bring us closer to realizing artificial general intelligence (AGI) — a milestone in AI research aimed at achieving human-level cognitive abilities across diverse tasks.
AGI proponents envision machine learning systems capable of executing complex task sequences and accurately predicting outcomes. While some researchers remain skeptical about the feasibility of AGI, figures like Elon Musk have made bold predictions, suggesting that AI could surpass human intelligence by the end of the coming year.
How AI Learns Using Neural Networks
AI systems rely on artificial neural networks (ANNs), designed to emulate the brain’s functioning to facilitate learning. ANNs can be trained to discern patterns in various forms of data, including speech, text, or images, driving numerous advancements in AI technology.
Traditional AI methods involve inputting vast amounts of data to teach algorithms about specific domains, such as Google’s language translation services or Facebook’s facial recognition software. However, this process is time-consuming and limited to specific knowledge domains.
A new generation of ANNs, known as Adversarial Neural Networks, introduces a novel approach where two AI systems engage in a learning competition, enabling them to learn from each other’s strategies. This approach accelerates the learning process and enhances the quality of output generated by AI systems.