Forget generative AI, it’s not me who says it, but Yann LeCun, the French guru of artificial intelligence. On the long road to general AI, Meta’s chief scientist and vice president is banking on a goal-based model that can safely learn, reason, and plan. An entire program.
AI here, AI there. Artificial intelligence is making headlines in the media and social networks, Clubic first and foremost. So, when Meta asked me to come to its premises to talk about artificial intelligence with its in-house experts, I didn’t hesitate for two seconds. Yann LeCun, Joëlle Pineau, Naila Murray… The casting was tempting, as was the promise of seeing more clearly in this maelstrom of innovations and technologies.
ChatGPT, the surprise he didn’t see coming
When Yann LeCun, Vice President and Chief AI Scientist of Meta speaks, it immediately becomes exciting and informative. Turing Prize 2018 and member of the National Academy of Sciences in the United States, we instantly realize that our interlocutor does not really have his tongue in his pocket: “ ChatGPT ? A surprise, a popular enthusiasm that we were unable to anticipate even though it was not the most advanced generative AI model of the time “. Internally, I wince a little, remembering the bad buzz caused at the time by the putting online of Galactica, Meta’s language model immediately withdrawn from circulation at the end of 2022.
While Elon Musk predicts the emergence of an AI more intelligent than humans next year, Yann LeCun also denotes the ambient concert by explaining that the level of current generative AI is far too limited: “ they can’t think for themselves, they make basic mistakes, they hallucinate… “.
These should be limited to assistance with writing, editing or, even, assisting developers in coding their computer programs. For him, although LLaMA is also a generative AI developed by Meta (version 3 was officially announced this week), there would be much better to do while waiting to reach the level of a general AI which, by the way , would not be close to arriving. 10 years ? 15 years? We do not know ! The man with the connected Ray-Bans preferring to talk about AMI, the acronym forAdvanced Machine Intelligence.
Taking inspiration from humans or animals for learning
One of the fundamental points to improve would be the machine learning or machine learning in French. The idea would be to draw a little more inspiration from humans, or even animals, who learn tasks much more quickly than any existing machine, who understand the world around them in the blink of an eye and who can develop quickly come up with a plan. It is this common sense, this behavior based on objectives that should be copied. “ Even a cat is able to plan complex tasks » recalls Yann LeCun, justifying the work still to be accomplished. These are comments that literally flatter my ego as a cat owner.
Still in the demonstration, Meta’s Mr. AI reminds us that a 10-year-old child is capable of clearing a table and filling a dishwasher all by himself, which no robot has ever managed to do to date. The same goes for level 5 autonomous cars which are still not deployed on a large scale while a 17-year-old teenager can get their license after 20 hours of driving school lessons.
THE machine learning generative AI is not at the level
And if we want to push the comparison of the learning capabilities of a generative AI model a little further with a 4-year-old child, there too, there is no match. Yann Lecun teaches us that a child’s two million optical nerves – a bandwidth per nerve estimated at 10 bytes/sec – are capable of ingesting 50 times more video data than an LLM (large language modelgenerative AI model) with text over the same learning time.
Yann LeCun’s observation is concise: the machine learning of these models is “ null “. To push the level of generative AI which, in 10 years of research, has never succeeded in learning effectively with objects such as images and videos, we would need to review the approach and favor learning through the objective.
Understand the world, to reason better and plan
This is the path chosen by Yann LeCun with his objective-oriented AI model (Objective-Driven AI Architecture). A model that would not seek to reconstruct what is missing in a text or video, but could construct an abstract representation of reality and predict what happens in the submitted content. The objective is to develop a system capable of understanding the world, having a certain common sense and planning action sequences to achieve a particular goal.
A system capable of learning, for example, the physics of our world through videos and which would come to the conclusion that “ the waving leaves of a tree or the reflection of light on the surface of a lake are not relevant information to predict or process “. The perfect AI assistant in short, capable of answering all our questions, memorizing, helping us in our daily lives and understanding our preferences and interests. A first model has already left the Meta laboratories, called JEPA for Joint Embedding Predictive Architecture.
A long evolutionary process towards general AI
Open source just like LLaMA, Meta’s generative AI, JEPA is in its second version and would already make it possible to predict missing frames in a video using self-supervised learning (V-JEPA), but also to recognize the content of an image (I-JEPA) or, even, measure the height of the forest canopy using satellite images.
From there to predict the imminent arrival of artificial general intelligence? It wouldn’t really be right away according to Yann LeCun. Goal-based AI would be just one step in the evolution of artificial intelligence. An evolutionary process punctuated by new models “ always smarter » until reaching, one day or another, the Everest of AI. A race where new players with long teeth like OpenAI and Mistral will not be afraid to shake up the established order and push the limits of science ever further. In the meantime, my cat would remain smarter than ChatGPT…
I have been the editor-in-chief of Clubic since September 2023. Like the entire editorial team, I am passionate about high-tech and more particularly "old school" computing, from processors to graphics cards, including NAS and network routers. .
Read other articles
3