Andrej Karpathy recently shared a gist he calls an “LLM Wiki” framed explicitly as an example of an “idea file.” The document itself is useful, but the meta-point he’s making is the more interesting one: the practice of maintaining a curated, living document about a technical domain is a skill worth modeling and adopting.
This landed on Hacker News with 200+ points, which tells you something. It wasn’t a paper, a library, or a benchmark. It was a text file. And people found it worth discussing at length.
What an Idea File Actually Is
An idea file is not a dump of bookmarks. It is not a reading list. It is closer to a personal wiki page for a single topic, written in a voice that reflects how you currently understand that topic, updated as your understanding changes.
The distinction matters. Bookmarks preserve links to where knowledge lives. An idea file preserves the knowledge itself, distilled and reorganized according to how you actually think about it. When you write a bookmark, you are outsourcing the synthesis to a future version of yourself who will, predictably, never do it. When you write an idea file entry, you are forcing synthesis now.
Karpathy’s LLM Wiki follows this pattern. It collects mental models, practical findings, gotchas, and conceptual anchors about large language models. The value is not that it is exhaustive; it is that it reflects a specific, coherent perspective built up over sustained engagement with the material.
The Fast-Moving Domain Problem
LLMs are an unusually hard domain to maintain a stable mental model of. The field publishes hundreds of papers per week. Techniques that were best practice twelve months ago are now considered naive. Benchmarks get saturated and replaced. New architectural ideas emerge, get hyped, get refined, and sometimes quietly disappear.
Traditional documentation does not handle this well. A blog post from 2023 on prompt engineering reads like a different era. A wiki maintained by a committee drifts toward committee-speak and becomes hesitant about anything contested. A course curriculum goes stale between iterations.
An idea file sidesteps these problems because it carries no obligation to be authoritative or comprehensive. It is one person’s working document. When something changes, the author updates it. When something turns out to be wrong, it gets corrected or deleted. The document serves its author first, and anyone else who finds it useful second.
This is actually the approach Karpathy has taken across his career. His nanoGPT repository is not a production library; it is a teaching artifact that reflects how he thinks a transformer should be written for clarity. His llm.c project grew out of wanting to understand, and then share, what GPT-2 training looks like in pure C. The idea file is the written equivalent of this: a record of how he currently understands the territory.
Knowledge Management in Software Development
The idea file sits in a long tradition of personal knowledge management practices that developers have used to handle information overload. The Zettelkasten method, popularized by sociologist Niklas Luhmann, centered on atomic notes linked by concept rather than chronology. Tools like Obsidian and Roam Research have built large communities around similar principles, with graph views and backlinks meant to surface connections across notes.
The idea file is simpler and less structured than any of these. It does not require atomic notes. It does not require a linking system. It requires only that you write down what you know, in one place, and keep it honest.
This simplicity is a feature. One of the recurring failure modes of personal knowledge management systems is that the system becomes the project. You spend time optimizing your vault structure, your tag taxonomy, your template library, and the actual capture of insight gets crowded out. A single markdown file in a gist has none of these affordances. There is nothing to configure. You either write in it or you do not.
What Makes the LLM Domain Suited to This Format
Not every technical domain benefits equally from an idea file approach. Stable reference material is better served by proper documentation with versioning and search. A domain where you are a beginner is better served by a structured curriculum that someone else has already curated.
LLMs sit in an interesting middle zone. The foundational concepts, attention mechanisms, tokenization, in-context learning, the scaling hypothesis, are stable enough to write down with confidence. The operational details, which models perform well on which tasks, how much context you actually need, when fine-tuning beats prompting, change frequently but follow patterns that experienced practitioners recognize. An idea file can hold both: stable concepts that serve as anchors, and provisional findings that get updated as evidence accumulates.
Karpathy’s document appears to reflect this structure. Core explanations sit alongside practical observations and things to watch. The format accommodates uncertainty without requiring the author to hedge everything into uselessness.
The Public Dimension
Sharing the document publicly changes its function in a useful way. A private idea file is a thinking tool. A public one is also a teaching artifact and an invitation to correction.
Karpathy has been explicit in the past about his view of public learning. His blog posts, YouTube lectures, and educational repositories all share a common thread: the act of explaining forces precision, and sharing the explanation invites feedback that makes the understanding better. The LLM Wiki extends this into an ongoing, living format rather than a one-time artifact.
This is different from a traditional blog post in a subtle but important way. A blog post is frozen at publication. An idea file can be updated, which means it can be wrong for a while and then become less wrong, reflecting the actual shape of how understanding develops rather than presenting a finished, polished perspective.
Adopting the Practice
The practical takeaway is straightforward. Pick a domain you are actively working in and that is moving fast enough that you keep having to relearn things. Open a new markdown file. Write down what you currently believe to be true about that domain: key concepts, useful heuristics, things that surprised you, things you got wrong. Keep the file somewhere you will actually open it again.
The discipline is in maintenance. A file you started eight months ago and never touched is an artifact, not an idea file. The value compounds when you return to it, notice where your understanding has shifted, and update accordingly. That act of comparison, what I thought then versus what I think now and why, is itself a form of deliberate learning that passive consumption of new content does not provide.
Karpathy’s LLM Wiki is useful as a resource. It is more useful as a demonstration that even researchers at the frontier find value in writing things down plainly and keeping the document honest over time. That is a practice that scales down to any technical domain, at any level of expertise.