New · DyNoSeD & BPPB seminar
Published research explained · October 8, 2026
How can unordered measurements constrain a rule of motion? An interactive example, the intuition behind the Fokker–Planck approach, and what DyNoSeD reveals about parameter ambiguity. Based on our published paper and my October 2 BPPB talk.
Read the explainer & try the example →
New · Research notes
The intuition & proof · October 8, 2026
Why replaying a trajectory backward differs from generating it stably. A guide to the geometric obstruction, from connected unstable leaves and disconnected stable fibers to an entropy contradiction, with a schematic and expandable technical details.
Read the explanation →
Santa Fe Institute · Part I of the talk · May 2026
A written summary of the first half of my SFI talk: recurrent models forecast the dynamical
world by synchronizing to it — forming a delayed embedding and holding it
transversely stable through invertible generalized synchronization. Includes the
slides.
Read the summary & get the slides →
Santa Fe Institute · Part II of the talk · May 2026
A self-contained derivation showing that single-head self-attention is exactly one step
of the Concave–Convex Procedure (CCCP) on a Gaussian-mixture energy landscape — geometrically,
a "hop" along the surface normal followed by a fixed linear readout. Includes the tanh and
linear kernel variants and a live interactive demo.
Read the write-up & open the demo →