Video summary

🇫🇷 Ce Français vient de faire TREMBLER la Silicon Valley

Main summary

Key takeaways

Technology

Summary

The video examines a research paper published online on October 29, 2026, by Randall Balestriero and Yann LeCun. It reports that a 109-million-parameter model, trained using a different approach from standard language models, outperformed much larger open models on one specific test. The result challenges the assumption that making models bigger is the only route to better AI—but it does not show that the small model is generally more capable than ChatGPT or other conversational systems.

The competing approaches

  • Conventional language models, such as GPT-style systems, are trained largely to predict the next word or token in text. LeCun argues that text alone does not give a system the grounded, physical understanding people and animals gain by observing the world.
  • The smaller model uses a JEPA-style learning approach: it learns to place different descriptions of the same scene near one another, focusing on shared meaning rather than predicting the next word.
  • The test compares how a text model groups descriptions with how a vision model groups corresponding images. In the video’s analogy, two librarians independently organize related photos and captions; the researchers then compare their “neighbors.”

Reported results

On this measure of similarity between text and visual representations, the paper reports:

  • Alibaba’s 72-billion-parameter model: 15.4%
  • Meta’s 70-billion-parameter Llama: 15.1%
  • A Mistral model: 14%
  • The 109-million-parameter model: 18.4%

The authors describe their model as 660 times smaller than the largest model tested and say it used over 100 times less text. In a separate comparison using the same small model and the same text, the JEPA-style method scored 18.4%, versus 11.1% for next-token training, while reaching that level using 42 times fewer examples.

Important limitations

The video emphasizes that this is not a general intelligence test. The small model was evaluated on how it organizes concepts, not on conversation, writing, translation, summarization, or calculations. ChatGPT, Claude, and Gemini were not included; the comparison used older open models.

The paper is also a preprint and, at the time discussed, had not yet been independently replicated. Another caveat is that many training descriptions were generated by LLaVA, an image-and-text system based on Meta’s Llama. The paper reports a substantial drop in performance on one dataset without machine-generated descriptions. The video notes that this caveat affects the comparison, though not the separate same-data comparison between training methods.

Industry and company context

LeCun, a prominent deep-learning researcher and former head of Meta’s FAIR research lab, has long argued that language-only training is insufficient for building systems that understand the physical world. He left Meta in 2025 as the company moved toward large-scale investment in models, computing infrastructure, and data centers.

His Paris-based company, AMI Labs, announced a $1.03 billion fundraising round in 2026. The video contrasts this with Meta’s planned $135–145 billion in investment that year, much of it directed toward infrastructure. Its central question is whether continued scaling will remain the best strategy, or whether different learning methods can produce better results with fewer resources. The video concludes that the paper is an early piece of evidence, not a verdict.

Reviews, guides, or tutorials: None are presented; the video is an analysis of a research paper and its implications.

Main speakers and sources

  • Main speaker: FĂ©lix Media narrator
  • Research paper: Randall Balestriero and Yann LeCun
  • Other analysis cited: Mayank Kejriwal
  • Additional sources referenced: Financial Times, TechCrunch, Maddyness, CNBC, Engadget, Fortune, ACM, Yann LeCun’s X post, and the World Economic Forum

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video