Yann LeCun's Test for AI Is a Baby in a High Chair

Yann LeCun says this is an experiment. A nine-month-old tosses each toy from the high chair to see it drop. He says that machines can take the same test. Feed an AI something that cannot happen, then check whether it's surprised. This video tracks that idea and the results we have so far. LeCun shared the 2018 Turing Award. He thinks machines must learn like a baby does, primarily by watching the world. He argues current chatbots will not reach human-level intelligence on their own. In September 2026, while Nvidia's Jensen Huang posted "AGI has arrived", LeCun wrote "LLM is today's technology" and "I work on tomorrow's revolution." The destination still stands; what he scraps is the label AGI. "a legitimate goal," he labels human-level AI, and he writes that machines will hit human smarts in every field. But "We're still far from that." "World model" now covers two system types. Google DeepMind's Genie 3 and World Labs' Atlas predict in 3D scenes or pictures. LeCun's kind forecasts via its own compact shorthand. It aims to help a machine plan. Both sides pitch their case, but this video crowns neither. LeCun co-authored a 2025 study where a model using his method identified the impossible video 98 out of 100 times on the classic test. It scored just over six in ten on two others. On a tougher test from the same Meta lab, it reached near 54. Humans hit about 96 there. A small system he helped design reacts to impossible events in basic simulations. An independent researcher rebuilt that small system. On one task, the result matched the report. However, the original model's performance relied heavily on target distance. It exceeded eight in ten when close, but dropped under one in ten far away. LeCun set up AMI Labs on this bet in March 2026. Investors put in just over a billion dollars. Its website showed no product by early October 2026. 0:00 The toy on the floor 1:58 The world itself 3:47 Don't draw the faces 5:45 Two kinds of model 7:41 The surprise test 11:17 Running the experiment WATCH NEXT Jensen Huang says AGI has arrived. At what, exactly?:    • Jensen Huang Says AGI Has Arrived. At What...   An AI record on Riemann's zeros, and the human who found a simpler proof:    • Claude Broke a Record on Riemann's Zeros. ...   SOURCES Yann LeCun at ETH Zürich, World Models (posted 9 Jun 2026):    • Yann LeCun: World Models: Enabling the nex...   LeCun on X, 25 Sep 2026: https://x.com/ylecun/status/210354394... Intuitive physics from self-supervised video (Feb 2025): https://arxiv.org/abs/2502.11831 IntPhys 2 (Jun 2025): https://arxiv.org/abs/2506.09849 LeWorldModel (Mar 2026): https://arxiv.org/abs/2603.19312 Independent reproduction of LeWorldModel (Aug 2026): https://arxiv.org/abs/2608.10145 V-JEPA 2 (Jun 2025): https://arxiv.org/abs/2506.09985 World Labs, Atlas: https://www.worldlabs.ai/blog/atlas Google DeepMind, Genie 3: https://deepmind.google/blog/genie-3-... AMI Labs: https://amilabs.xyz Subscribe for editorial explainers on the AI economy:    / @aiefficiencytalks   #YannLeCun #WorldModels #AGI #ArtificialIntelligence #JEPA #AIResearch #MachineLearning #LLM CREDITS Strobe photograph of a falling ball: Michael Maggs, CC BY-SA 3.0. Crown of Christian IV: photo by Ikiwaner, cut-out by Simetrical, CC BY-SA 3.0. Both via Wikimedia Commons. Wikipedia article pages, CC BY-SA 4.0. IntPhys benchmark videos: Riochet et al., CC BY 4.0. IntPhys 2 videos: Meta FAIR (Bordes et al., 2025), CC BY-NC 4.0. Earth at night: NASA GSFC / NOAA NGDC. Quarter: United States Mint. Footage used editorially: ETH Zürich's Computer Vision and Geometry Group (LeCun's talk), ACM, Yann LeCun (his 1989 demo), Johns Hopkins University, Boston College, the Beckman Institute at Illinois, Meta, World Labs, Google DeepMind, NVIDIA, Figure AI, Offcall, the Lex Fridman Podcast, BBC News, Computerphile, Peter Kiechle (slip-detection demo), a documentary segment on Renée Baillargeon's lab, Where's Waldo (Washington Square), and family videos by Seattle BabyBeeps, The JooJ, Ziad Bou-Zeid and Felisse Agatha Biay. Stock footage and photos: Pexels. Pages, posts and papers shown editorially: posts on X by Yann LeCun, Jensen Huang and AMI Labs; AMI Labs, Fortune, TechCrunch, World Labs, Google DeepMind and ACM; papers by Garrido et al., Bordes et al., Assran et al., Maes et al. (LeWorldModel, with its project videos and its code on GitHub), Singh, Mathieu et al., NVIDIA (Cosmos 3) and LeCun (2022). Logos are trademarks of their owners, shown for identification only.