autoresearch
The self-improving lab notebook
You named it autoresearch, which is either a bold research agenda or the moment the experiment started assigning itself homework. Either way, the premise has excellent mad-scientist energy.
A
Ship machine
silver medal
Overall
0.0/ 100
Roasts
llm.c
CUDA with a side of humility
llm.c takes the grand language-model conversation and translates it into the language of kernels, memory, and compile errors: apparently even giant ideas must eventually fit through a pointer.
llm-council
The committee meeting
llm-council turns model comparison into a parliamentary system, because one language model was apparently not enough bureaucracy for a single prompt.
hn-time-capsule
The internet time machine
hn-time-capsule preserves old discussion as though online discourse were a fine wine. A charmingly ambitious attempt to make the comment section historically significant.
llm101n
Archived before graduation
llm101n sounds like a complete curriculum, while its archived status gives it the energy of a semester that ended with an impressive final project and no sequel syllabus.
Category breakdown
Impact
25% weight
The portfolio centers on ambitious, distinctive educational and experimental machine-learning projects, including implementations, training-oriented work, and research-oriented tools.
Consistency
15% weight
The supplied repositories show repeated substantial activity from 2024 through 2026 across multiple related projects, though the evidence does not establish a complete contribution history.
Quality
25% weight
Non-fork ownership, substantial project scope, clear specialization, and continued updates provide strong maturity signals; source-level engineering quality cannot be assessed from the available metadata.
Depth
20% weight
Multiple major repositories show sustained ownership and meaningful scope, with activity distributed across foundational neural-network education, language-model systems, and research tooling.
Breadth
10% weight
The work spans Python, CUDA, and Jupyter Notebook, with educational, implementation, experimentation, community-oriented, and information-retrieval project types, though the domain focus is concentrated.
Community
5% weight
The supplied follower and star counts indicate substantial public attention, but community evidence is used only as a weak positive tie-breaker and does not raise the other categories.
Stats
52-week commit heatmap
151 active daysLanguage distribution
4 langs
- Python67%
- Jupyter Notebook17%
- Cuda8%
- Unknown8%
Based on the bounded repository sample saved with this analysis.
Numbers
Owned repos
Non-fork sample
12
Commits
Last 12 months
367
Followers
Accepted fact
215796
Joined GitHub
Profile date
Apr 2010
Top repos
Scores marked Profile rating use bounded public repository metadata. Full repo analysis appears only when a separate accepted repository analysis exists.
karpathy /
autoresearch
92/100
An ambitious Python research and automation project with very recent activity and unusually broad apparent experimentation scope.
OwnedActiveProfile ratingPython
★ 93515 · 5mo agoOpen on GitHub ↗
karpathy /
llm.c
89/100
A substantial CUDA project targeting language-model work at a lower systems level, giving the portfolio notable computational depth.
OwnedActiveProfile ratingCuda
★ 30765 · 1y agoOpen on GitHub ↗
karpathy /
nanogpt
88/100
A focused Python implementation project centered on compact language-model experimentation, with sustained ownership and recent updates.
OwnedActiveProfile ratingPython
★ 62008 · 9mo agoOpen on GitHub ↗
karpathy /
nanochat
87/100
A large, active Python project focused on accessible language-model experimentation and tooling, with strong apparent ambition.
OwnedActiveProfile ratingPython
★ 57078 · 8d agoOpen on GitHub ↗
karpathy /
nn-zero-to-hero
84/100
A substantial notebook-centered neural-network learning curriculum with clear educational scope and continued activity.
OwnedActiveProfile ratingJupyter Notebook
★ 23919 · 2y agoOpen on GitHub ↗
karpathy /
micrograd
83/100
A compact foundational autodiff and neural-network project presented through notebooks, combining conceptual clarity with meaningful implementation scope.
OwnedActiveProfile ratingJupyter Notebook
★ 17067 · 7d agoOpen on GitHub ↗
karpathy /
llm-council
82/100
An active Python project exploring multi-model coordination or evaluation, with an unusual and distinctive premise.
OwnedActiveProfile ratingPython
★ 23813 · 9mo agoOpen on GitHub ↗
karpathy /
mingpt
80/100
A maintained Python language-model implementation project that reinforces the portfolio’s sustained focus on understandable, compact systems.
OwnedActiveProfile ratingPython
★ 24782 · 2y agoOpen on GitHub ↗
karpathy /
makemore
76/100
A Python project centered on generative modeling and language-oriented experimentation, contributing foundational depth to the portfolio.
OwnedActiveProfile ratingPython
★ 4173 · 2y agoOpen on GitHub ↗
karpathy /
arxiv-sanity-preserver
73/100
A long-running Python research-information project that broadens the portfolio into literature discovery and preservation.
OwnedActiveProfile ratingPython
★ 5758 · 5y agoOpen on GitHub ↗
karpathy /
llm101n
72/100
An archived language-model education project with substantial apparent scope, though its archived status limits evidence of ongoing maintenance.
OwnedArchivedProfile ratingUnknown
★ 37507 · 2y agoOpen on GitHub ↗
karpathy /
hn-time-capsule
66/100
A smaller Python project built around preserving or exploring historical online discussion, showing useful thematic breadth beyond model implementation.
OwnedActiveProfile ratingPython
★ 668 · 8mo agoOpen on GitHub ↗
Timeline
- Joined GitHub
- Created karpathy/arxiv-sanity-preserver - Web interface for browsing, search and filtering recent arxiv submissions
- Created karpathy/micrograd - A tiny scalar-valued autograd engine and a neural net library on top of it with PyTorch-like API
- Created karpathy/mingpt - A minimal PyTorch re-implementation of the OpenAI GPT (Generative Pretrained Transformer) training
- Created karpathy/makemore - An autoregressive character-level language model for making more things
- Created karpathy/nn-zero-to-hero - Neural Networks: Zero to Hero
- Created karpathy/nanogpt - The simplest, fastest repository for training/finetuning medium-sized GPTs.
- Created karpathy/llm.c - LLM training in simple, raw C/CUDA
- Created karpathy/llm101n - LLM101n: Let's build a Storyteller
- Created karpathy/nanochat - The best ChatGPT that $100 can buy.
- Created karpathy/llm-council - LLM Council works together to answer your hardest questions
- Created karpathy/hn-time-capsule - Analyzing Hacker News discussions from a decade ago in hindsight with LLMs
- Created karpathy/autoresearch - AI agents running research on single-GPU nanochat training automatically
- Most recent push to karpathy/micrograd
Compare
KARPATHY · 6DMEDIAN CODER
Rubric
How this score was produced
Overall = Σ(category × weight) + deterministic top-end curve
| Category | Weight | Score | Contribution |
|---|---|---|---|
| Impact | 25% | 92 | 23.00 |
| Quality | 25% | 78 | 19.50 |
| Depth | 20% | 89 | 17.80 |
| Consistency | 15% | 82 | 12.30 |
| Breadth | 10% | 68 | 6.80 |
| Community | 5% | 50 | 2.50 |
Impact · 92/100
What it measures. Project usefulness, ambition, originality, and coherence. Popularity is not impact.
Evidence for karpathy. The portfolio centers on ambitious, distinctive educational and experimental machine-learning projects, including implementations, training-oriented work, and research-oriented tools.
Quality · 78/100
What it measures. Engineering and project-maturity signals supported by the available public metadata.
Evidence for karpathy. Non-fork ownership, substantial project scope, clear specialization, and continued updates provide strong maturity signals; source-level engineering quality cannot be assessed from the available metadata.
Depth · 89/100
What it measures. Sustained ownership, meaningful scope, and continued maintenance rather than one-shots.
Evidence for karpathy. Multiple major repositories show sustained ownership and meaningful scope, with activity distributed across foundational neural-network education, language-model systems, and research tooling.
Consistency · 82/100
What it measures. Recency and contribution patterns across the supplied activity window.
Evidence for karpathy. The supplied repositories show repeated substantial activity from 2024 through 2026 across multiple related projects, though the evidence does not establish a complete contribution history.
Breadth · 68/100
What it measures. Language entropy and project-type diversity across owned repos.
Evidence for karpathy. The work spans Python, CUDA, and Jupyter Notebook, with educational, implementation, experimentation, community-oriented, and information-retrieval project types, though the domain focus is concentrated.
Community · 50/100
What it measures. A weak positive tie-breaker for supplied community signals, never a popularity penalty.
Evidence for karpathy. The supplied follower and star counts indicate substantial public attention, but community evidence is used only as a weak positive tie-breaker and does not raise the other categories.
- Raw total
- 81.90
- Top-end curve
- +0.00
- Final overall
- 81.9
Tier thresholds
S 90-100 Mass-producing humansA 80-89 Ship machineB 70-79 Solid engineerC 60-69 Getting thereD 40-59 README enthusiastF 0-39 GitHub tourist
I. How this saved profile rating was produced
- 01
Validate. The server validates the login, signed browser session, attempts, cooldown, daily ceiling, and the single provider permit.
- 02
Collect. The VPS collector reads bounded public profile metadata and up to 12 recent repository metadata rows in volatile memory.
- 03
Rate. One pinned provider returned a closed profile-rating object with no tools, browsing, shell, or repository-content access.
- 04
Verify. Server checks bound the subject, rejected unsafe or echoed text, validated the schema, and recomputed rubric-v3 final arithmetic.
- 05
Save. Only a valid completion and allowlisted post-completion facts commit atomically; the public projection is then rebuilt.
This saved profile analysis is metadata-only and did not read repository trees, README files, or source code. Availability of new analysis is separate from this historical result.
Rated 2026-08-10 · rating-rubric/3
II. Data sources & caveats
- Scores summarize one accepted, subject-bound analysis; they are not a security audit or endorsement.
- Rank, percentile, and median use 1095 currently accepted public profiles and can change as the leaderboard changes.
- Repository and language details reflect the bounded analyzed sample, not necessarily every repository.
- Missing data is shown as unavailable or not evaluated instead of being inferred as zero.
- Analysis completed Aug 10, 2026, 12:12 AM UTC and is marked aging.