Schmidhuber challenges JEPA’s lineage with a 1992 PMAX comparison

On 31 March 2026, Jürgen Schmidhuber published a technical note arguing that the Joint Embedding Predictive Architecture set out in Yann LeCun's 2022 position paper substantially reprises Predictability Maximization work he and Daniel Prelinger described in 1992. It is a serious lineage challenge, not a settled verdict. The note puts two architectures three decades apart on the same bench and asks whether new naming has outrun old citation.

Jürgen Schmidhuber Watch analysis

What happened

The 1992 report gave two neural networks related inputs. One network learned a representation that another predicted, while an opposing discrimination pressure kept the representation from becoming a content-free constant. It also described asymmetric and symmetric forms, auto-encoding, constrained variance and other ways to preserve information. Schmidhuber's note maps that recipe onto JEPA's encoders and predictor, then calls the core arrangement essentially identical.

LeCun's 2022 paper defines JEPA as a non-generative architecture that predicts one input's representation from another's, then discusses regularisation, architectural choices and training procedures intended to prevent collapse. The paper also says almost all of its ideas had appeared in prior forms, claims no priority for them, and presents its contribution as assembling them into a consistent architecture for autonomous machines.

Why it matters

Conceptual overlap is visible: both approaches predict useful representations rather than every raw detail, and both need a mechanism against trivial outputs. But identity is a stronger claim. Objectives, training procedures, intended scope, empirical scale and integration into a wider world-model system matter. These sources establish the competing descriptions; they do not independently adjudicate who deserves a broad invention label.

An April addendum reproduces a reply attributed to LeCun: joint embedding architectures were decades old, no one claimed they were new, and PMAX was not JEPA. Schmidhuber rejects that distinction. The dispute is therefore less about whether antecedents exist than about which antecedent matches the later architecture closely enough to warrant explicit citation.

Our read

Priority rows can turn a bibliography into pub darts. The useful move is not to award the trophy to whoever throws hardest. Put the objectives, anti-collapse machinery, experiments and citations side by side; family resemblance is evidence, not a birth certificate.

What to watch

• A line-by-line technical comparison of PMAX and JEPA objectives, predictor roles and anti-collapse constraints.
• Independent historical reviews that separate the joint-embedding family from the narrower JEPA formulation.
• Reproductions testing whether the 1992 objectives recover behaviour claimed for modern JEPA systems.
• Citation changes or formal responses in later papers rather than another round of social-media shorthand.

Discussion spark: What evidence should decide whether PMAX is a direct JEPA predecessor: matching objectives, equivalent learned behaviour, citation history, or something else?

Independent WittyWires tracker for public updates about Jürgen Schmidhuber. Not affiliated with or endorsed by Jürgen Schmidhuber; this is not an official account.