OpenAI and Meta ready new AI models capable of ‘reasoning’

Stay informed with free updates

OpenAI and Meta are on the brink of releasing new artificial intelligence models that they say will be capable of reasoning and planning, key steps towards achieving superhuman cognition in machines.

This week, executives at OpenAI and Meta signalled that they are preparing to launch the next versions of their large language models, the systems that power generative AI applications such as ChatGPT.

Meta said it will begin rolling out Llama 3 in the coming weeks, while Microsoft-backed OpenAI indicated that its next model, expected to be called GPT-5, was coming “soon”.

“We are hard at work in figuring out how to get these models not just to talk, but actually to reason, to plan . . . to have memory,” said Joelle Pineau, vice-president of AI research at Meta.

OpenAI’s chief operating officer Brad Lightcap told the Financial Times in an interview that the next generation of GPT would show progress on solving “hard problems” such as reasoning.

“We’re going to start to see AI that can take on more complex tasks in a more sophisticated way,” he said. “I think we’re just starting to scratch the surface on the ability that these models have to reason.”

Today’s AI systems are “really good at one-off small tasks”, Lightcap added, but were still “pretty narrow” in their capabilities.

Meta and OpenAI’s upgrades are part of a wave of new large language models being released this year by companies including Google, Anthropic and Cohere.

As tech companies race to create ever more sophisticated generative AI — software that can create humanlike words, images, code and video of quality indistinguishable from human output — the pace of progress is accelerating.

Reasoning and planning are key steps towards what AI researchers call “artificial general intelligence” — human-level cognition — because they allow chatbots and virtual assistants to complete sequences of related tasks and predict the consequences of their actions.

Joelle Pineau, vice-president of AI research at Meta: ‘We are hard at work in figuring out how to get these models not just to talk, but actually to reason, to plan . . . to have memory.’ © Kimberly M Wang

Speaking at an event in London on Tuesday, Meta’s chief AI scientist Yann LeCun said that current AI systems “produce one word after the other really without thinking and planning”.

Because they struggle to deal with complex questions or retain information for a long period, they still “make stupid mistakes”, he said.

Adding reasoning would mean that an AI model “searches over possible answers”, “plans the sequence of actions” and builds a “mental model of what the effect of [its] actions are going to be”, he said.

This is a “big missing piece that we are working on to get machines to get to the next level of intelligence”, he added.

LeCun said it was working on AI “agents” that could, for instance, plan and book each step of a journey, from someone’s office in Paris to another in New York, including getting to the airport.

Meta plans to embed its new AI model into WhatsApp and its Ray-Ban smart glasses. It is preparing to release Llama 3 in a range of model sizes, for different applications and devices, over the coming months.

Lightcap added OpenAI would have “more to say soon” on the next version of GPT. 

“I think over time . . . we’ll see the models go toward longer, kind of more complex tasks,” he said. “And that implicitly requires the improvement in their ability to reason.”

At its event in London, Chris Cox, Meta’s chief product officer, said the cameras in Meta’s Ray-Ban glasses could be used to look at, for instance, a broken coffee machine, and an AI assistant — powered by Llama 3 — would explain to the wearer how to fix it.

“We will be talking to these AI assistants all the time,” LeCun said. “Our entire digital diet will be mediated by AI systems.”

Via

Leave a Comment

mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt mdt