In 2021, Emily Bender and her collaborators coined the phrase "stochastic parrots" to describe Large Language Models (LLMs). Their argument was pointed: these systems, trained to predict the next token in vast corpora of text, lacked understanding, merely recombining previously seen patterns—much like a parrot randomly repeating phrases without grasping their meaning.
The term escaped academia into popular parlance. Stochastic parrot became a shorthand critique not just of LLMs, but of a nebulous notion of AI itself. It offered a concise, reassuring narrative to those both with serious and very valid critiques of the technology, as well as those without.
Yet, over the past few years, it has become clear that LLMs have far greater ability to generalize, reason, plan, and represent abstract structure in ways that are difficult to dismiss as mere mimicry. They may not reason like humans, but they do reason. I first encountered this as I learned about the field of interpretability, with its beautiful, though often brittle, results revealing real structure.
This important work, however, has largely failed to capture the public imagination. Public understanding of AI is still dominated by coverage from commentators who lack deep technical grounding, and whose critiques often default to knee-jerk skepticism or moral panic. There are extremely serious criticisms to be made of modern AI—about power, incentives, assumptions, and harm—but those are rarely the critiques that make headlines.
This blog hopes to bridge that gap.
I'll post about the incredible power of AI and how to best harness it, about its immense technical and political failings, about my own learnings and work, and I'll promote others' excellent work on the subject.
My aim is to make modern AI legible: to explain what these systems are actually doing, where they succeed, where they fail, and why both matter. I want to speak simultaneously to curious non-specialists and to researchers working at the frontier—showing not just what AI can be used for, but how it works under the hood, and what that implies for safety, alignment, and our understanding of intelligence itself.
Along the way, we may find that the parrot has something worthwhile to say after all.
