New AI models from OpenAI and Meta, which they claim will be able to reason and plan—two essential steps toward obtaining superhuman cognition in machines—will soon be available.
The next iterations of the massive language models—the engines behind generative AI apps like ChatGPT—that OpenAI and Meta executives hinted at this week were about to be released.
Vice-president of AI research at Meta Joelle Pineau stated, “We are hard at work figuring out how to get these models not just to talk, but actually to reason, to plan, to have memory.”
The Financial Times was informed by Brad Lightcap, Chief Operating Officer of OpenAI, that the next generation of GPT will demonstrate advancements in addressing “hard problems” like thinking.
Though still “pretty narrow” in scope, today’s AI systems are “really good at one-off small tasks,” according to Lightcap.
The new large language models that Google, Anthropic, and Cohere are releasing this year include the updates from Meta and OpenAI.
The speed of innovation is rising as tech companies rush to develop even more advanced generative AI, or software that can produce human-like texts, images, code, and video of a quality that is indistinguishable from human output.
As AI researchers progress towards “artificial general intelligence,” or cognition on par with humans, reasoning and planning become increasingly crucial because they enable chatbots and virtual assistants to accomplish related tasks in sequence and anticipate the outcomes of their choices.
Current AI systems, according to chief AI scientist Yann LeCun of Meta, “produce one word after the other really without thinking and planning,” during a Tuesday speech at an event in London.
They continue to “make stupid mistakes” he claimed, since they find it difficult to handle difficult inquiries or to retain information for extended periods of time.
They still “make stupid mistakes,” he claimed, because they find it difficult to handle challenging inquiries or to remember knowledge for an extended amount of time.
Incorporating reasoning would include an AI model “looking up potential solutions,” “arranging the flow of events,” and creating a “mental model of how [its] actions will be perceived,” according to him.
He continued, saying that there is a “huge missing piece that we are working on to get machines to get to the next level of intelligence.”
LeCun reported that it was developing artificial intelligence (AI) “agents” that could, for example, schedule and plan every leg of a trip, from a person’s office in Paris to another in New York, including traveling to the airport.
WhatsApp and Ray-Ban smart glasses will both include Meta’s new AI model. It is getting ready to release Llama 3 in the upcoming months in a variety of model sizes for a range of devices and applications.
Regarding the upcoming GPT version, Lightcap stated that OpenAI will have “more to say soon.”
“I believe that in due course, we will witness the models progressing towards more extended and intricate tasks,” he remarked. “And that implies that their capacity for reasoning needs to be strengthened.”
At a London event, Chris Cox, the chief product officer of Meta, mentioned that the cameras in the company’s Ray-Ban glasses might be used to look at things like a broken coffee maker, and an AI assistant powered by Llama 3 would instruct the wearer on how to fix it.
LeCun stated, “We will be communicating with these AI assistants constantly.” “AI systems will mediate our entire digital diet.”
