public demo + playground demo: read-only · playground: rate-limited write

Karpathy’s llm-wiki — without the vault dying at Telegram scale

Same loop: raw immutable · agent compiles wiki · Obsidian as IDE. Different substrate: holographic index so queries don’t re-scan hundreds of thousands of messages.

Compare / contrast (measured)

~609 chats · ~321k msgs · 25 MB raw · 20 queries · full writeup in docs

Arm A — markdown scan

Classic Obsidian / agent greps the vault
p50 latency~122 ms
p95 latency~158 ms
Grows with corpus?yes
Exact string needlestrong

Arm B — hologram + hybrid

winner for scale query
p50 latency~0.022 ms
Speedup vs A~5,100×
Index size~1.1 MB
md roleraw + human IDE
Why: at research scale the pure-md wiki is fine. At “whole shebang” life dumps, scan cost and context pressure dominate. Keep Obsidian. Put a holographic brain underneath. Hybrid keyword for exact phrases.

How / why install for yourself

  1. Why: you want Karpathy compounding without your corpus destroying context / agent time.
  2. Engine credit: leCore by undacappn — we ship the product, not rebrand the engine.
  3. Self vaults on this machine:
# open as Obsidian vaults
open -a Obsidian /Users/stacc/brain-ab/vault-a
open -a Obsidian /Users/stacc/brain-ab/vault-b

# vault-a = md + redteam wiki · vault-b = hologram projection + plugin
  1. Pull your active TG (conv turns):
# api_id/hash: https://my.telegram.org/apps
python3 /Users/stacc/brain-ab/code/tg_pull_active.py --dialogs 150 --per-chat 500
python3 /Users/stacc/brain-ab/code/tg_pull_active.py --live   # forward-looking
  1. Or play here: query demo-telegram (read-only) or teach the public playground (rate-limited writes).

Query

loading…
Type a query — nothing is pre-filled.
Shoulders of giants: leCore by undacappn · pattern: Karpathy llm-wiki · product/market/TG/hosted: stacc
Read: POST /v1/{tenant}/query · Teach: POST /v1/playground/teach · Owner reload: X-Owner-Key on /admin/load