Large technology companies are busy upgrading their own large-scale language models. According to reports, OpenAI and Meta are about to release their respective new versions of artificial intelligence (AI) models, claiming that these models will be able to perform reasoning and planning, which is a key step in achieving superhuman cognition in machines.
This week, executives from OpenAI and Meta announced that the two companies are preparing to launch the next version of their large-scale language models, which will power generative artificial intelligence applications.
Meta stated that Llama 3 will be launched in the coming weeks; OpenAI, on the other hand, stated that its next model, expected to be called GPT-5, will be "soon launched".
Inference function
Joelle Pineau, Vice President of Meta Artificial Intelligence Research, said, "We are working hard to figure out how to make these models not only speak, but also truly reason and plan... with memory."
OpenAI Chief Operating Officer Brad Lightcap said in an interview that the next generation of GPTs will make progress in solving "difficult problems" such as reasoning. "We will start to see artificial intelligence being able to handle more complex tasks in a more precise way, and I believe we are just beginning to touch the surface of these models' reasoning abilities."
Lightcap added that today's artificial intelligence systems are "very good at one-time small tasks," but their capabilities are still "quite limited.".
Paving the way for AGI
As technology companies increasingly roll back technology, the pace of progress is accelerating as they compete to create more complex generative artificial intelligence - models that can create text, images, code, and video that are no different from humans, and their quality is constantly improving. This year, companies such as Google, Anthropic, and Cohere have successively released a new wave of big language models.
The reasoning and planning functions are the necessary path to General Artificial Intelligence (AGI). AGI possesses human level cognition, allowing chatbots and virtual assistants to complete sequences of related tasks and predict the consequences of their behavior.
Meta's chief artificial intelligence scientist Yann LeCun stated at an event in London on Tuesday (April 9th) that current artificial intelligence systems "generate words one after another without thinking and planning," making it difficult for them to handle complex problems or remember information for a long time, so they still "make foolish mistakes.".
He stated that increasing reasoning ability means that artificial intelligence models will "search for possible answers", "plan the order of actions", and establish a "psychological model about what impact their actions will have.". LeCun added that this is an important missing part of the current model, and these capabilities will enable the machine to reach the next level of intelligence.
OpenAI's Lightcap indicates that OpenAI will soon have more to say about the next version of GPT.
Lightcap said, "I believe that over time, we will see these models move towards longer and more complex tasks, which requires them to improve their reasoning abilities."