June 15, 2026
Old math, new AI, fresh drama
The 90-year-old idea behind JEPA models: Canonical Correlation Analysis
Turns out AI’s ‘new’ trick may be a 1936 hand-me-down, and the comments are smug about it
TLDR: A new article says a core idea behind today’s JEPA AI models may actually come from a 1936 math method called CCA. Commenters split between smug “everyone should already know this” lectures and amused jokes that AI keeps reinventing old ideas with better hardware.
The big reveal in this piece is delightfully chaos-inducing: one of today’s buzziest AI training ideas may trace back to 1936. The article argues that a math method called Canonical Correlation Analysis, or CCA, laid the groundwork for what newer systems like JEPA are doing now—basically, teaching a model to match up different views of the same thing. In plain English: a lot of people are reacting like the AI world just found out grandma already invented the family recipe.
And oh, the community had feelings. One camp was in full professor-mode, practically scolding the room: if you work in unsupervised learning and don’t know CCA, what are you even doing here? That sparked the classic tech-comment-section energy: part history lesson, part gatekeeping, part victory lap for the “nothing in AI is truly new” crowd. Another commenter took the more playful route, joking about an alternate timeline where these ideas took off earlier and wrecked GPU supply chains even harder, because apparently reality wasn’t dramatic enough already.
There’s also a side serving of old-school AI credit-war drama, thanks to the ongoing JEPA invention debate. The article gently suggests the real ancestor might be Harold Hotelling, not the modern names fighting over ownership. Translation: the community isn’t just debating math—it’s debating who gets the legend status. And readers seem weirdly thrilled by the possibility that the hottest “new” thing is really a 90-year-old comeback story.
Key Points
- •The article argues that Harold Hotelling’s 1936 Canonical Correlation Analysis is the theoretical and intuitive foundation for modern embedding-prediction methods such as JEPA.
- •It states that JEPA and CCA share the objective of extracting common signal across two views, though in JEPA the second view is another view of the same data.
- •The article says JEPA-based models have been described in recent literature as a non-linear generalization of CCA.
- •It distinguishes CCA from JEPA by noting that CCA is linear and does not require a shared encoder, while JEPA uses joint embeddings with an encoder and predictor.
- •The technical section derives that, under CCA’s whitening constraints, maximizing cross-correlation is equivalent to minimizing mean squared error between paired embeddings.