Recently Milla Jovovich open sourced an LLM memory management system based on the concept of memory palaces (essentially placing memories into rooms that can be retrieved later). Memory management in LLMs is a big problem. I've struggled with this in my projects and RAG and other retrieval and storage methods aren't really a solution.
Milla used an AI agent to develop the codebase (like everyone else), and the ideas around the system are really sound.
There's a big challenge though, and Milla's not the only one who has it: The dark code problem.
What is dark code? According to Jouke Waleson it is: "lines of software that no human has written, read or even reviewed."
We all know that AI agents are fantastic at generating code quickly. What's still slow? Human comprehension. Agents can describe code one way and it does another.
The Review
The Response
Milla ran into the same problem we all do with AI generated code! Agent will confidently claim a feature exists, but when you actually look at the codebase you sometimes quickly conclude: no, this isn't doing what you claim it is.
There's a lot of pressure to ship often and ship fast. AI coding agents are getting better, code is becoming commoditized, but understanding is still slow, messy and operates at human scales.
How are you all handling the dark code problem?
Top comments (0)