Blog
Notes from the lab
What we are building, what we measured, and what went against us.
22 Oct 2026
A surprise map of a book
We let a byte-level model read a 19th-century memoir and coloured every character by how surprised it was. It found the names, the index, and four typing errors in the source.
21 Oct 2026Streaming is not memory
A model with a fixed-size state can read five million bytes in one pass. That is not the same as remembering them. What we measured, and what it means.
20 Oct 2026Why phase?
Radio receivers, radar, MRI scanners and coherent optics all measure complex numbers. Most neural networks throw half of that structure away. Here is why we do not.
14 Mar 2026CNet: complex-valued neural networks in C++ and CUDA
A tour of CNet, our open-source framework for complex-valued networks: every layer complex by construction, explicit Wirtinger gradients, and a CPU reference checked against every CUDA kernel.