Jack Morris is an AI researcher and co-founder of Engram. His work examines what language models remember, how text can be recovered from numerical representations, and how models can learn the context of an organization.[1][2]
Start with Text Embeddings Reveal (Almost) As Much As Text →
X Career
Morris received his PhD at Cornell Tech, working with Sasha Rush and Vitaly Shmatikov and supported by an NSF Graduate Research Fellowship. He spent part of the doctorate as a student researcher at Meta FAIR and previously was a Google Brain resident. His Engram announcement describes an effort to have models study a body of context repeatedly and retain understanding, rather than reread all of it for every question.[1][2]
Contributions
Embeddings are not anonymous text
Morris and collaborators treated embedding inversion as an iterative reconstruction problem. An embedding represents text as a numerical vector. Their model generates a candidate text, embeds it again, then edits the candidate to move closer to the target vector. The paper reports exact recovery of 92% of 32-token inputs in its experiments and recovery of names from clinical notes. This demonstrates a privacy risk in those tested settings; storing vectors does not automatically remove the information in the original words.[3]
Measuring memory separately from generalization
His memorization paper separates information about a particular training dataset from information about the process that generated the data. By removing generalization in controlled experiments, the authors estimate capacity near 3.6 bits per parameter for the GPT-style models they studied. The work connects capacity and dataset size to membership inference, the task of determining whether an example was in training. It is a measurement framework with experimental limits, not a universal capacity constant for every model.[4]
Tools, writing & talks
Writing, research & talks
Start here
A paper explaining how iterative generation and correction can reconstruct text from embeddings.
Begin with the embedding-inversion paper’s abstract and examples. Knowing what an embedding is helps, but the generate, compare and correct loop gives a concrete entry into his broader work on information stored by models.
Co-author[3]Train directly on ten thousand financial reports, drive the loss to near zero, and generation collapses. Jack Morris uses that to argue that against your own corpus the data axis is fixed and inference compute is the only axis left.
Catalog workOriginal sourceLatent Space interviews Jack Morris on information theory for language models and his research path, offering a rigorous framing for how LLMs encode and learn information.
Catalog workOriginal sourceA framework and experiments for separating dataset memorization from generalization and estimating model capacity.
Co-author[4]
Tools
- TextAttack
github.comA framework for composing adversarial attacks and evaluation utilities for text models. Morris lists it among the open-source projects he built.
Creator and contributor[5]