The typical set

I am the CEO of .txt, a company that serves quality tokens. Our work started with an ArXiv paper on guided generation with Large Language Models, the Outlines library, but has expanded since. You can find out more about what we do on our blog.

Before diving into Large Language Models I was particularly interested in Bayesian Statistics and generative modeling, MCMC sampling and symbolic programming. The interests translated into the open source libraries I maintained:

• Outlines is a Python library to guide the generation of Large Language models.
• Aesara is a Python library for defining, rewriting and efficiently evaluating mathematical operations on arrays; A "tensor" library.
• AePPL is a collection of tools for A[e]PPL in Aesara.
• Blackjax is a sampling library built with JAX, focused on ease of use, speed and modularity.
• AeMCMC is a sampling library built with Aesara which builds symbolic samplers from model graphs.
• Aehmc provides implementations of samplers in the HMC family (including NUTS) with Aesara.

Unlike a traditional blog, most of the notes you will find here are constant work in progress. They form a tightly knit web of related notions. When traditional blogs are engineered, neatly ordered forests, the typical set is more like an aspen grove:

With this experiment I am also trying to figure out how to write useful notes, and how to efficiently navigate between them.

Notes that are found in the rabbit hole sometimes coalesce into blog posts. Unlike evergreen notes, blog posts are time-stamped and are thus allowed (sometimes expected) to become stale.

>> ENTER THE BLOG <<

• The 22 immutable laws of branding
• The Creative Act: A way of being
• Are your lights on? How to figure out what the problem really is
• The art of community
• Start at the end: How to build products that Create Change

This website is forever under construction.