Jack Morris goes against the current: information theory, not agents or benchmarks

Latent Space profiles one AI PhD student every year. Two years ago it was Shunyu Yao, whose dissertation focused on language agents and who went directly to OpenAI. This year Latent Space chose Jack Morris, and the choice is deliberately different.

Morris ignores the loudest trends and asks what models actually represent

Morris is deliberately not working on agents, new benchmarks or VS Code forks. His area is the information-theoretic understanding of language models: how models encode and compress information, what the latent space reveals, and how embeddings become representations from which meaning can be read back.

Latent Space features him as a counterpoint to research that chases leaderboard results. The specific content of the interview is not fully available from the public excerpt. What is written here draws from the episode structure and Morris's prior work on embedding models and latent representations.

Why theory matters when building RAG, retrieval and interpretability tools

An information-theoretic view of LLMs is not academic curiosity. It affects practical things: how embeddings work and why they sometimes fail, what can be inferred about a model without access to its weights, and how to design retrieval systems with a better understanding of what the model sees during semantic search.

Without this foundation the market easily gets stuck on surface-level score comparisons. Much of today's RAG infrastructure is designed intuitively rather than from an understanding of how embeddings actually handle information. Morris is going where most researchers are not going during the current wave of agentic excitement.

Foundational focus: slow impact, broad reach

This interview is valuable as a signal that academic AI research includes people focused on foundations, not just the loudest product trends. The direct impact on existing products is less straightforward than agentic or retrieval topics.

Information theory for LLMs is an area where results are hard to translate into concrete recommendations for engineering teams. Morris's value will be more visible over several years, if his findings influence how embedding models or interpretability tools are designed.

Where Morris's findings will be in three years

Follow Morris's publications on embedding models and latent representations. The interesting question is whether an information-theoretic view translates into better diagnostics for retrieval failures or into tools for interpreting model behavior.

Latent Space PhD profiles are a useful signal of where research is heading outside the mainstream. Shunyu Yao was an agent thesis and went straight to OpenAI. Morris is about foundations. The question is where his findings will be in three years.

Lilith's verdict

In a moment when almost every researcher is building another agent or a new benchmark, it is worth watching the people who ask what models are actually doing under the hood. Morris's focus on information theory and latent representations is a quieter topic than Codex, but if it yields results it will reshape how embeddings and retrieval systems are designed for the next decade.