RSSAmplifier

Blog

Leello Dadi

Recent content on Leello Dadi

saturdaygenfo.github.ioRSS feed ↗3 posts

Latest posts

trivial improvements on the re-masking strategy of llada

trivial improvements on the re-masking strategy of llada an auto-regressive (AR) language model generates a single token per function call or neural function evaluation (NFE). Language diffusion models are interesting because, in principle, they could generate more tokens per NFE. llada is, at the time of writing, the largest masked language diffusion model that is competitive with AR models. Its…

Generating Sparse Stochastic Processes

Test

The Symmetric gradient: an odd 40 year curiosity in matrix algebra

There shouldn’t be anything particularly difficult about differentiating with respect to symmetric matrices. Differentiation is defined over abstract spaces. And the set of real symmetric matrices $\mathbb{S}_n(\mathbb{R})$ is not special. And yet, this past semester, Paul and I, along with a student, Aleksandr, ran into problems. It turns out that the problem of computing gradients with…