Sub-title: Alex Zhang’s idea points beyond simply training bigger models.
I recently read Alex Zhang’s work on Recursive Language Models, or RLMs, and also watched his Scientific American July/August 2026 interview. The idea that stayed with me is not only about long context. It is about a different way to think about AI capability.
The usual instinct in AI is: bigger model, more compute, larger context window.
But Zhang’s RLM idea suggests another path: give the model a better working environment.
Instead of forcing an LLM to swallow a huge prompt or thousands of documents all at once, RLM treats the long context as something outside the model, accessible through a coding environment such as a Python REPL. The model can inspect the context, search it, split it, summarize parts of it, and recursively ask smaller model calls to work on selected pieces.
This feels important because it changes the model from a passive reader into something closer to an analyst with a notebook.
A coding environment gives the LLM useful “hands”: it can count, parse, search, test, compare, and organize. These are exactly the things human analysts do with tools. We do not solve every problem by staring harder at a giant document. We open files, run checks, filter data, and break the work into smaller pieces.
That is why I find RLM interesting. It is not just a clever long-context technique. It is a practical argument that AI capability may come not only from larger models, but also from better scaffolding around models.
This also connects to how I think AI should be built. The goal should not be simply replacing people’s jobs. The better goal is helping people do the work they already want to do, but better and smarter.
For coding, research, data analytics, validation, documentation, and quality control, we may not always need a giant oracle. Sometimes we need a careful assistant with a workbench.
Signoff: The next leap may be less about bigger mouths, and more about better workbenches.
Note: What is a Python REPL?
REPL stands for Read-Eval-Print Loop. It is an interactive shell where you type a line of Python, it runs immediately, and you see the result right away — no need to write a full script first. In the context of RLM, the REPL is the execution environment the model works inside: it can run code, inspect variables, and call itself recursively, all within that live session.
This post was written with assistance from ChatGPT/Codex.