Andrej Karpathy — what to adopt for IKA

Audit of github.com/karpathy top-20 repos against IKA needs (Operations + Builder + Token stack). Compiled 2026-04-26. Companion to /ika-pro-architecture, /gstack, /graphify, /tokens.

TL;DR

One adopt, one watch, ignore the rest. Most of his repos are educational; the agent-orchestration ones are what matter for us.

ADOPT (pattern + possibly code): llm-council (17.7k★, Nov 2025). Karpathy's canonical implementation of multi-vendor LLM consultation — the same pattern we already plan to lift from gstack #3 (multi-vendor second opinion) and lock in v17's threshold table. Two independent strong signals (Karpathy + Garry Tan) for the same pattern = high confidence in the design. Use Karpathy's code as the implementation reference for our second-opinion middleware in Paperclip.

WATCH: reader3 (3.5k★, Nov 2025). "Read books together with LLMs" — small reference for our Note + Document + LLM interaction patterns. Worth a 1-hour code-walk before we build the doc-summarization skill, but not a direct lift.

IGNORE: nanochat, nanoGPT, llm.c, llama2.c, micrograd, minGPT, LLM101n, nn-zero-to-hero, build-nanogpt, makemore, minbpe — all educational implementations of LLM training/inference internals. We don't train models, we use them. Reference material if/when team members onboard to ML fundamentals; no adoption.

Top 20 repos (by stars) with relevance

Rank	Repo	Stars	Last commit	What it is	IKA relevance
1	autoresearch	76.7k	Mar 2026	"AI agents running research on single-GPU nanochat training automatically"	Pattern interesting (autonomous agent loop) but specific to ML model training; not directly applicable
2	nanoGPT	57.2k	Nov 2025	Training/finetuning medium GPTs	Skip — we don't train
3	nanochat	52.5k	Apr 2026	"The best ChatGPT that $100 can buy" — full self-hosted chat model	Skip — we use Anthropic
4	LLM101n	36.8k	Aug 2024	"Let's build a Storyteller" — educational course	Skip — pure education
5	llm.c	29.7k	Jun 2025	LLM training in raw C/CUDA	Skip — we don't train
6	minGPT	24.2k	Aug 2024	Minimal PyTorch GPT	Skip — education
7	nn-zero-to-hero	21.6k	Aug 2024	Neural Networks: Zero to Hero course	Skip — education
8	llama2.c	19.4k	Aug 2024	Llama 2 inference in pure C	Skip — we use Anthropic
9	llm-council	17.7k	Nov 2025	"LLM Council works together to answer your hardest questions"	ADOPT pattern + lift code — multi-vendor consultation; pairs with gstack #3 / v17 second-opinion middleware
10	micrograd	15.6k	Aug 2024	Tiny autograd engine	Skip — education
11	char-rnn	12.0k	Oct 2023	RNN language model (legacy)	Skip — outdated
12	convnetjs	11.2k	Jan 2023	Browser-based deep learning training	Skip — outdated
13	minbpe	10.5k	Jul 2024	BPE tokenizer reference	Skip — education
14	arxiv-sanity-preserver	5.7k	Nov 2021	arXiv browser/search	Skip — old
15	neuraltalk2	5.6k	Nov 2017	Image captioning (Torch)	Skip — old
16	neuraltalk	5.5k	Dec 2020	Multimodal RNN learning	Skip — old
17	build-nanogpt	5.0k	Aug 2024	Lecture: building nanoGPT	Skip — education
18	ng-video-lecture	4.7k	Jan 2024	Video lecture support	Skip — education
19	makemore	3.9k	Jun 2024	Char-level autoregressive model	Skip — education
20	reader3	3.5k	Nov 2025	"Read books together with LLMs" — doc + LLM interaction	WATCH — useful pattern reference for our Note/Document + LLM interaction; 1-hour code-walk before building doc-summarization skill

llm-council deep dive

What it is: Karpathy's reference implementation of "asking multiple LLMs and reconciling answers." For complex/ambiguous questions, you query a council (e.g. Claude + GPT-4 + Gemini), see the spread of answers, and synthesize a final response. Often catches mistakes that any single model would make alone.

Why it matters for IKA

IKA need	How llm-council answers it
v17 multi-vendor second-opinion (gstack #3) — high-stakes Operations actions need a second model to review before execution	This IS that pattern. Karpathy's implementation is canonical reference; lift the request fan-out + answer reconciliation logic.
Token stack Layer 3 (LiteLLM smart routing) — route low-stakes to Haiku, high-stakes to Sonnet	Council pattern extends routing to "ask multiple models when stakes are high; defer to single model when stakes are low"
Builder side — when I (Claude) face an ambiguous design call, asking a council instead of guessing	Direct application during v3.0 / v3.5 build for the hardest architectural decisions

What to lift from llm-council

Concurrent fan-out — issue the same prompt to N models in parallel; collect responses; track latency + token cost per provider
Reconciliation prompt — Karpathy's "synthesize these answers" prompt template (have one model read all the others' responses + produce a unified answer with conflicts noted)
Disagreement detection — flag when models disagree substantively vs trivially; route disagreements to human review
Cost cap — abort the council if cumulative cost exceeds threshold (already in our v17 governance model)

Operating notes

Tier: community-vetted (per v19.5) — Karpathy is a known author, code is small + readable, MIT license
Vendoring: lift small (likely <500 lines); pin to a specific commit; vendored under /vendored/karpathy-llm-council/ alongside our gstack-vendored prompt-injection classifier
Wiring: integrate into Paperclip's approval-gate middleware as an alternative to the simpler Codex-only second-opinion (v17 default); per-IKA setting picks single-vendor-second-opinion vs council
Cost: ~3× the cost of single-model per query (3 models in parallel). Enable per the v17 threshold table only — money/identity/permission changes; high-priority external recipients

reader3 — quick look

Small repo (~3.5k stars) demonstrating the pattern of "read a book + LLM together." Loads a PDF, splits into chunks, lets the user converse with the LLM about specific passages with the book always in context.

For IKA: relevant when we build the document-summarizer + note-companion skills (per v19.5 catalog). Don't lift the code wholesale (it's a demo, not a library); lift the UX pattern — keep the source document visible side-by-side with the conversation, let user click any passage to ground the next agent turn.

Effort: 1-hour code walk before building our own doc-companion skill. Maybe a few specific functions worth extracting (chunking strategy, citation format).

Why ignore the educational repos

nanoGPT, llm.c, llama2.c, micrograd, minGPT, LLM101n, nn-zero-to-hero, build-nanogpt, ng-video-lecture, makemore, minbpe, char-rnn, convnetjs, neuraltalk(2), arxiv-sanity-preserver — these are training/inference internals + courses. Excellent if you want to learn how transformers work or build your own model from scratch.

IKA's posture: we use Anthropic Claude as the model layer and don't train. So:

nanochat / llm.c / llama2.c (self-host alternatives) — only relevant IF we ever ditch Anthropic, which we're not planning
The educational courses — fine reference material when team members need to deepen ML understanding; not adoption candidates

Different posture would change this calculus. If we wanted full sovereignty over the model layer (no third-party LLM API), nanochat at $100 + llm.c for training would be the path. Out of scope for v0.

Recommendation

Action plan

Add llm-council to the gstack #3 / v17 second-opinion middleware spec in /ika-pro-architecture — note Karpathy as a second-source confirmation of the pattern + reference implementation. ~10 minutes of doc edit.
Vendor llm-council code under /vendored/karpathy-llm-council/ alongside the gstack prompt-injection classifier when we build out the second-opinion middleware. Pin to a specific commit. ~half day to integrate.
Code-walk reader3 before we build the document-summarizer / note-companion skills (per v19.5 catalog). Lift the UX pattern of side-by-side document + chat. ~1 hour.
Skip everything else. Karpathy\'s educational repos stay as ML reference, not infrastructure.

Open questions for Mats

Greenlight vendoring llm-council for the v17 second-opinion middleware (alternative to a Codex-only approach)?
For high-stakes Operations actions, do we want SINGLE second opinion (Codex per gstack #3) or COUNCIL (3+ models per Karpathy)? Council = ~3× cost but stronger guarantee. Per-IKA setting?
Anything in his Eureka Labs work (LLM101n, etc.) we want to use as onboarding material for Mats / IKA team? It's pure education, not for prosumers.

Sources: github.com/karpathy top-20 repos by stars (2026-04-26). Companion: /gstack, /tokens, /graphify, /ika-pro-architecture. On-disk at build/web/bi/main/karpathy/index.html.