Yann LeCun: LLMs are the best example of self-supervised learning — the next big challenge is video

At a recent FT Live event, Turing Award winner and Meta Chief AI Scientist Yann LeCun shared another powerful insight:
👉 The most successful real-world application of self-supervised learning so far is the Large Language Model (LLM).

He explained that LLMs became powerful because they learned language patterns by predicting the “next token” — a pure form of self-supervision trained on massive unlabelled data.

But the most exciting part was what he said next:

🔥 The key challenge for the coming years is applying self-supervised learning to video and multimodal sensor data.
🎯 Why video?

LeCun pointed out:
📌 Image AI rose through ImageNet and supervised learning
📌 Language AI exploded through LLMs and self-supervised learning
📌 But the real world is continuous and dynamic — made of video and sensor streams

Video is far more complex than text:

Huge amount of visual information

Temporal dependency

No simple “predict the next word” task

He believes this will spark the next generation of AI breakthroughs — and is essential for building AI systems with true common sense.

🧩 Self-supervised learning = the core of future AI

LeCun has long argued:
Real intelligence cannot rely on heavily labeled datasets.
Self-supervised learning is the only scalable path for AI to truly understand the world.

LLMs are just step one.
The next step is AI that can see, perceive, and understand reality.

🌟 My thoughts

LLMs proved the power of self-supervision — but extending it to video and multimodal sensors feels like the next turning point in AI evolution.🔥
We might be witnessing the start of AI’s next era.

#ai #vlog #technology #elonmusk #news #llm #largelanguagemodels #ImageNet #supervisedlearning #SelfSupervisedLearning #videoai #multimodalai #meta #aithoughtleaders #AITechTrends

Видео Yann LeCun: LLMs are the best example of self-supervised learning — the next big challenge is video канала Artificial Intelligence