Explainers · measured July 2026
Karpathy called it vibe coding. The vibes still run grep.
Andrej Karpathy gave the moment its name: describe the software in English, let the agent build it, give in to the vibes, barely look at the code. It is a real shift, and a good one. But English never touches a disk. To act on the vibe, the agent shells out to the same primitives a 2006 laptop used (grep, ls, read) and walks millions of files to find the one it needs. On the machine below, that walk took 93.8 seconds and burned ~58% of the session's tokens. The vibe layer is only ever as fast as the retrieval layer under it.
The idea, taken seriously
The vibe is real. So is the disk under it.
In early 2025 Karpathy described a way of working he called vibe coding, where you “fully give in to the vibes” and “forget that the code even exists.” It followed his older line that “the hottest new programming language is English.” Both are real, both are his, and millions now work this way. The part the phrase gracefully hides is what happens the instant the vibe becomes an action.
The model does not know where your login page lives, which file holds the theme, or whether the config was ever committed to git. So it does the only thing it can: it searches. Here is the whole picture in one frame: the glossy sentence on top, and the loop that actually runs underneath.
Measured: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4.47M files.
One real task, measured end to end
We watched the vibe hit the disk.
One instruction. To carry it out without an index, Claude Code walked the tree, grepped bytes, read files, and grepped again: 71 tool calls before it touched the right one. With an always-warm index answering in its place, the same task needed a single call.
End to end (the whole detour, from the sentence to the file in hand), the measured run collapsed by about 7,200,000×. Tokens spent grepping are tokens not spent building. On a $200-a-month agent plan, half the bill can go to directory listings the model reads once and forgets a turn later. The token cost of file hunting is the least glamorous line item in vibe coding, and one of the largest.
Same machine · same drive · same queries
The same file, everywhere the vibe lands.
| The searcher | The task | Time | vs Interlinked |
|---|---|---|---|
| Interlinked | one indexed lookup | 85 μs | baseline |
| VS Code file search | same file, same drive | 35 s | 412,000× |
| Windows Search | median of 20 queries | 67 s | ~480,000× |
| ripgrep (what your agent runs) | content query, same corpus | 93.8 s → 16 ms | 58,625× on the set |
| Claude Code (no index) | the full find-one-file run | 6 m 57 s → 16 ms | ~7,200,000× |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Single-file lookups are a median 85 μs (best case 2 μs); the 20-query median across the whole drive was 139 μs. The ripgrep row is one content query timed both ways: 93.8 s walking the disk, 16 ms indexed; the 58,625× figure is the median speedup across the full content set. The Claude Code row is the end-to-end measured run: wall time, 71 → 1 tool calls, and tokens folded in. Full per-tool methodology in the 412,000× VS Code write-up and the ripgrep comparison.
Why the vibes stay slow
English compiles down to grep.
A vibe-coding tool can only be as quick as the tools it calls. Describe the change in a sentence and the agent still has to resolve it to files, and the resolver is a directory walk. VS Code indexes the folder you opened; Cursor indexes the repo you're in. Your machine is a dozen repos that ship together plus the 90% of the disk that was never in git: configs, downloads, the PDF with the answer. The vibe's questions don't respect folder boundaries; its tools do.
So the agent falls back to the floor: ripgrep, walking the disk on every single query, because nothing warmer exists across the whole machine. Give that floor a live, always-warm index (a save searchable in about a millisecond, under 30 ms even at its worst) and the vibe layer finally answers at the speed it always looked like it had. That is the backbone of the LLM economy: make the layer every agent stands on answer at memory speed.
The floor under the vibes
Speed finds the file. Memory knows why it matters.
The same index that answers in microseconds also carries a knowledge vault the engine builds by itself: files that change together, repos that share hundreds of identical files, notes any agent can leave on any file that every other agent reads. Keyed to each file's OS identity, so it survives renames and spans every repo: the memory the vibes forget the moment a folder is renamed. Not a replacement for RAG or your model's memory: the layer under both.
THE VIBE IS THE INTERFACE. THE INDEX IS THE FLOOR IT STANDS ON.
Do this today
Give your vibes a floor.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON editing, no keys.
Let it index. Filename search works immediately; content search fills in behind it. At rest the whole engine sits around 44 MB (less than a browser tab) and stays a millisecond behind your saves.
Scope when you can. Whole-machine when you don't know where something lives; scoped to a folder when you do. Scoped queries return in the microsecond range.
Let the vault grow. Day one it already knows your git history. Every week after, it knows your machine better, automatically, offline, and it never leaves the disk.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.