pack.cxml
Single-file CXML (karpathy/rendergit format). Paste into Claude or ChatGPT — the model gets every doc with source attribution.
↓ Download pack.cxmlResearch pack · updated 2026-04-18 · built from the repo
Everything we've learned
about cutting LLM costs.
Every finding from the KostAI research loop — the throughput-and-cost ledger, 10+ research docs, the prior-art harvest from six flattened upstream repos (vLLM, llama.cpp, LLMLingua, LiteLLM, Medusa, DSPy) — bundled as a single CXML file you can paste into Claude or ChatGPT. The model gets the whole corpus in one shot.
14
documents in the pack
145.2 KB
total, raw markdown
147.1 KB
CXML download
146.9 KB
markdown download
pack.md
Same content, plain markdown with a TOC. Safer for models that don't parse XML well, or for grepping on your own machine.
↓ Download pack.mdmanifest.json
Per-file SHA-256 + byte count. Re-download when
generatedAt or hashes change.
Browse the corpus
Each link opens the source on GitHub so you can see the raw markdown, file history, and commit that last touched it.
Project overview
Evidence & ledger
Research — 2026-04-17
- Gemini Deep Research — Round 2 (2026-04-17)
- R06 — Token-level speculative decoding for LLM APIs
- R07 — Context persistence via KV-cache offload and prefix reuse
- R08 — Swarm verification: many cheap verifiers vs. one expensive one
- R09 — Stream severing and programmatic early-stop
- R10 — Hierarchical planner (frontier, rare) + executor (local, many)
- Gemini Deep Research — Round 3 (Ranked Prompt Pack)
- R36 — VQToken: extreme token reduction via neural discrete codebooks
- R37 — Prior-art harvest from 6 flattened open-source projects