109M vs 72B: Why Better Training Pairs Matter

What if better AI starts with better training pairs, rather than just a bigger model? This short visual explainer looks at a 2026 preprint by Randall Balestriero and Yann LeCun. Their 109M-parameter text JEPA achieved stronger alignment with image representations than the language models evaluated, including Qwen2.5-72B, on the reported WIT-1024 test. That is a specific representation-alignment result—not proof that the smaller model is generally smarter or can replace a conversational assistant. A dog, a bicycle and three descriptions help explain good views, representation space and why the relationships between examples matter. Real paper excerpts and a redrawn chart show where the reported numbers come from. READ THE FULL ARTICLE https://abvx.xyz/writing/lecun-jepa-109m-vs-72b-platonic-representations CHAPTERS 00:00 109M vs 72B: the question 00:21 The paper and JEPA 00:49 How training pairs work 01:01 The alignment results 01:16 What this does not prove 01:26 The takeaway RESEARCH Randall Balestriero & Yann LeCun — Proving and Improving the Platonic Representation Hypothesis with JEPAs, arXiv:2610.39302v1, 29 October 2026: https://arxiv.org/abs/2610.39302 Minyoung Huh, Brian Cheung, Tongzhou Wang & Phillip Isola — The Platonic Representation Hypothesis, ICML 2024: https://proceedings.mlr.press/v235/huh24a.html VISUALS & PRODUCTION PDF excerpts: Balestriero & LeCun, CC BY4.0. Cropping/highlights added; chart redrawn from Table 3. The point-space geometry is explanatory, not measured embeddings. The realistic dog-and-bicycle scene and thumbnail are AI-generated illustrations. English narration is synthetic; it does not impersonate either researcher. Original instrumental music is reused from our studio library. Produced with Framewright: https://github.com/markoblogo/framewright #JEPA #MachineLearning #AIResearch