OpenAI and Meta are preparing to release new AI models that they say will be able to reason and plan on their own, a key step toward achieving "superhuman cognition" in machines.

This week, Meta said it would begin rolling out Llama 3 in the coming weeks, while OpenAI said its next-generation model, expected to be called GPT-5, is imminent.

The media quoted Joelle Pineau, Meta's vice president of AI research, as saying: "We are working hard on how to make these models not only able to speak, but also capable of self-reasoning, planning and memory."

Brad Lightcap, chief operating officer of OpenAI, also said that the next generation of GPT will make progress in solving "hard problems" such as reasoning. "We're going to start to see AI being able to perform more complex tasks in more complex ways," he said. "I think we're just starting to scratch the surface of the potential of these models' ability to do inference."

Lightcap added that today's AI systems are "excellent at handling small, one-off tasks" but are still "fairly limited" in capabilities.

Meta's chief AI scientist, Yann LeCun, similarly pointed out at an event in London on Tuesday that current AI systems "just generate word for word, without real thought and planning."

LeCun said that because these systems struggle to handle complex problems or retain information for long periods of time, they still "make some silly mistakes." He explained that adding reasoning capabilities means the AI model will "explore possible answers," "plan sequences of actions" and build "a mental model of what the effects of its actions will be."

He also said that Meta is working on developing things such as full-process AI "agents" that can plan and book every step of the journey, and plans to integrate new models into WhatsApp and Ray-Ban smart glasses.

At the same time, technology giants such as Google, Anthropic and Cohere are also accelerating the development of this field, competing to develop more advanced generative AI, aiming to make the text, images, code and videos generated by software the same quality as humans. The outputs are indistinguishable.