A text model with 109 million parameters, trained by Randall Balestriero and Yann LeCun, agreed with image models more closely than every language model they tested, up to Qwen2.5-72B, on the benchmark of the original Platonic Representation Hypothesis paper. Their paper, posted on arXiv on 29 October, puts the small model at .184 against .154 for Qwen2.5-72B: about 19.5% higher, with about 660 times fewer parameters, by our arithmetic.
The result is about one property, not general ability. The test scores how far a text model and a separately trained image model, DINOv2, agree on which captions and images sit closest together. The small model is an encoder; it does not write text, and the paper does not compare it with language models on chat, reasoning or coding.
Business Pill · Power Law Explained
A short explainer on power laws: why a few items take most of the outcome. It teaches the general idea only and says nothing about the paper in this piece.
The key insight: Balestriero and LeCun report a 109 million parameter text JEPA scoring .184 on image-text alignment against .154 for Qwen2.5-72B, a model about 660 times larger by our arithmetic, and .111 for the same encoder trained with next-token prediction on the same texts.
The paper, “Proving and Improving the Platonic Representation Hypothesis with JEPAs”, lists Balestriero at Brown University and AMI Labs and LeCun at New York University and AMI Labs. The hypothesis it takes up says models trained on different kinds of data come to represent the world in the same way as they grow; the authors set out to prove a version of it for Joint-Embedding Predictive Architectures, or JEPAs.
Their claim turns on what the model is trained to treat as a matching pair. A JEPA trained on two captions of the same image, which the paper calls good views, learns the same representation as an image JEPA up to one rotation, the paper argues. Next-token prediction, the objective behind today’s language models, is treated as a JEPA with bad views: parts of a single document.
For the headline test the authors trained a 109 million parameter BERT-B text encoder with 900M samples from the texts of 940M Recap-DataComp images, on 8 or 16 GPUs. That encoder read 116 billion text tokens, the paper says, while it gives 15 to 18 trillion tokens for Qwen2.5-72B and Llama-3-70B: about 129 to 155 times more, by our arithmetic.
Alignment with a DINOv2 ViT-B image encoder on the WIT-1024 benchmark (mutual 10-nearest-neighbour score; higher means closer agreement) against parameter count on a log scale. The 109M good-view text JEPA scores .184; the same encoder trained with next-token prediction .111; the Qwen2.5 family from 0.5B to 72B runs from .102 to .154; Llama-3-8B .145, Mistral-7B .140, Llama-3-70B .151. Source: Balestriero and LeCun, arXiv 2610.39302, Tables 3 and 16. JEPA value is the mean of three seeds; a fourth seed scored .121. The 660x size ratio is our arithmetic.
The Numbers Against 72 Billion Parameters
On the WIT-1024 benchmark, scored by mutual 10-nearest-neighbour agreement with a DINOv2 ViT-B image encoder, the good-view JEPA reached .184. Qwen2.5-72B scored .154, Llama-3-70B .151, Llama-3-8B .145, Mistral-7B .140 and CLIP’s text encoder .124, according to the paper’s Tables 3 and 16.
The same 109M encoder trained with next-token prediction on the same texts scored .111. Only the training objective differs between the two, so the good-view version scored about 66% higher, by our arithmetic. Across the Qwen2.5 family, going from 0.5B to 72B parameters, 144 times the size, moved the score from .102 to .154, by our arithmetic from Table 16.
The lead held on the other tests in Table 3: .230 against .210 for Qwen2.5-72B with CLIP’s ViT-H/14 image encoder, and .161 against .108 on 5,000 COCO test pairs. The authors write that their model “has over 600 times fewer parameters than Qwen2.5-72B and never saw pixels”.
The paper also reports a sample-efficiency figure: on the same corpus, a good-view JEPA reached with 21.6M training samples the alignment that next-token prediction reached with 900M, which the authors put at 42 times fewer samples.
What the Authors Conclude
The abstract reads: “next-token prediction is a remarkably inefficient route to platonic representations”. It closes: “Positive views, and not scale alone, are thus the lever to build platonic representations”.
As we read it, the business question is where the cost of a capability sits. The authors attribute the gain to how training pairs were defined, a data-curation choice, rather than to parameter count, a compute bill; on identical texts, their own control puts good views at .184 against .111 for next-token prediction. The paper does not price either.
As we read the paper, its sharpest comparison is the internal one: the same 109M encoder, the same texts and the same number of samples, with only the training objective changed, scored .184 with good views and .111 with next-token prediction, about 66% higher by our arithmetic.
As we read it, the comparison with the 72B model is the one that travels, and it is the one the authors qualify most: they write that the comparisons with language models at scale rely on captioner-written texts, and that without a captioner good views reached .060 against .107 with the recaptions.
Balestriero and LeCun, arXiv 2610.39302
“Positive views, and not scale alone, are thus the lever to build platonic representations”
Three Implications
FOR MODEL BUYERS The test measures one property, agreement with image encoders, and the small model is an encoder; the paper does not compare it with language models on generation, reasoning or coding.
FOR TRAINING BUDGETS The paper reports 116 billion text tokens and 8 or 16 GPUs for the large run, against 15 to 18 trillion tokens for Qwen2.5-72B and Llama-3-70B; it gives no cost in dollars for either.
FOR THE SCALING DEBATE Across the Qwen2.5 family, 144 times the parameters moved the score from .102 to .154, by our arithmetic from Table 16, while the paper puts the gain from good views at 42 times fewer samples on the same corpus.
The framework reads: “The transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it”.
As we read it, the paper is an argument about where that accumulation happens: in how training pairs are defined rather than in parameter count alone. The framework was published in July 2026 and does not discuss this paper.
The result covers alignment with image encoders on two benchmarks and two metrics. It does not show that a 109M model can do what a 72B language model does; the JEPA is an encoder and was not tested on generation, reasoning or instruction following. The paper is an arXiv preprint.
The authors list three caveats. A fourth training seed, with twice the effective SIGReg weight, reached only .121; all four seeds average .168, still above every language model, and about 9% above Qwen2.5-72B by our arithmetic. Most of the training texts were written by an image recaptioner that sees images through CLIP ViT-L/14. And at least 43 of the 1,024 WIT captions appear among the training alt-texts, though removing them changed none of the nine text models checked by more than .002.
Under known limitations, the paper says two of each image’s three texts on Recap-DataComp were generated by LLaVA built on LLaMA-3-8B. Without any image captioner, good views reached .060 on WIT against .107 with the recaptions, so the comparisons with language models at scale “rely on these texts”. We did not rerun the code or contact the authors.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Balestriero and LeCun report a 109 million parameter text JEPA scoring .184 on image-text alignment against .154 for Qwen2.5-72B, a model about 660 times larger, by our arithmetic. The authors attribute the gain to how training pairs were defined, and on identical texts good views scored .184 against .111 for next-token prediction; the paper says the margin over language models also relies on the captioner-written training texts.
A note on sourcing. All results, caveats and quotations are from Balestriero and LeCun’s arXiv preprint 2610.39302 (version 1, 29 October 2026) and its listing; we did not rerun the code. Scores are as given in the paper’s Tables 3 and 16. Size, score and token ratios are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
