{"slug": "autoresearch", "name": "autoresearch", "category": "ai", "category_name": "AI models and assistants", "description": "", "recommenders": 8, "commenters": 20, "warners": 1, "score": 9.02, "rising": -1.29, "first": "2026-03-08", "latest": "2026-09-19", "why": "which is extremely powerful when applied to problems that generalize well to it", "why_is_quote": true, "has_verdict": false, "sentiment": {"recommend": 8, "neutral": 8, "compare": 3, "warn": 1}, "spark": [0, 0, 0, 0, 3, 0, 1, 2, 1, 1, 0, 0], "listed": true, "loved": [], "caveat": "", "short_quote": "which is extremely powerful when applied to problems that generalize well to it", "kind": "", "icon": true, "domain": null, "github_repo": "github.com/karpathy/autoresearch", "official_url": "https://github.com/karpathy/autoresearch", "pricing": "", "aliases": ["autoresearch", "Autoresearch", "AutoResearch"], "verdict": null, "praise": [{"label": "", "summary": "", "quotes": [{"id": 788508640, "item_id": 49281790, "url": "https://news.ycombinator.com/item?id=49281790", "by": "lightbulbish", "date": "2026-08-12", "stance": "recommend", "quote": "maybe this?", "excerpt": "maybe this? https://github.com/karpathy/autoresearch", "context": "", "use_case": "", "switched_from": "", "story": "DeepSeek V4 Pro 0813"}, {"id": 779420000, "item_id": 48713750, "url": "https://news.ycombinator.com/item?id=48713750", "by": "binyu", "date": "2026-06-28", "stance": "recommend", "quote": "which is extremely powerful when applied to problems that generalize well to it", "excerpt": "This is particularly relevant because the guy used Andrej Karphaty's \"autoresearch\" [1] approach, which is extremely powerful when applied to problems that generalize well to it. [1] https://github.com/karpathy/autoresearch", "context": "", "use_case": "AI-driven research", "switched_from": "", "story": "Guy beats one of the best published quantum circuits for breaking ECDSA with AI"}, {"id": 774466512, "item_id": 48404157, "url": "https://news.ycombinator.com/item?id=48404157", "by": "hgoel", "date": "2026-06-04", "stance": "recommend", "quote": "This is already known to work.", "excerpt": "As usual, I find the AI-related discussion here to be hopelessly hysterical and conspiratorial. I get the impression that a large chunk of people have only read the title and assumed Anthropic is referring to recursive self-improvement in the runaway singularity sense. One of the examples they provide, of giving Claude the task of training a small AI model, then asking it to improve certain benchmarks, is essentially", "context": "", "use_case": "AI-driven model benchmark improvement", "switched_from": "", "story": "When AI Builds Itself"}, {"id": 767565920, "item_id": 47972870, "url": "https://news.ycombinator.com/item?id=47972870", "by": "ramon156", "date": "2026-05-01", "stance": "recommend", "quote": "Seemed like that one was safe", "excerpt": "I was actually about to run kaparthy's autoresearch tool. Seemed like that one was safe https://github.com/karpathy/autoresearch", "context": "", "use_case": "", "switched_from": "", "story": "Shai-Hulud Themed Malware Found in the PyTorch Lightning AI Training Library"}, {"id": 760821600, "item_id": 47551350, "url": "https://news.ycombinator.com/item?id=47551350", "by": "latentsea", "date": "2026-03-27", "stance": "recommend", "quote": "autoresearch managed to improve the harness performance", "excerpt": "I'm currently running autoresearch against my harness that autonomously builds SaaS against an enforced architecture, and autoresearch managed to improve the harness performance on my 'time-to-Realworld' benchmark which has Claude Code drive the harness to build an implementation of https://github.com/realworld-apps/realworld with the win condition that it must pass my rigorous postman collection + playwright test su", "context": "", "use_case": "Self-improving agent experiments", "switched_from": "", "story": "HyperAgents"}]}], "caveats": [{"label": "", "summary": "", "quotes": [{"id": 759903808, "item_id": 47493988, "url": "https://news.ycombinator.com/item?id=47493988", "by": "dvt", "date": "2026-03-23", "stance": "warn", "quote": "likely not worth the $$$ spent on the tokens", "excerpt": "Ok, so looking at the commit log[1], I was mostly interested in seeing what the \"moonshot ideas\" implementations looked like, but basically everything is just hyperparameter tuning. Which is nice, but likely not worth the $$$ spent on the tokens. Am I missing something here? [1] https://github.com/ykumards/eCLIP/commits/main/autoresearch", "context": "", "use_case": "", "switched_from": "", "story": "I tried Karpathy's Autoresearch on an old research project"}]}], "who": [], "tags": [], "use_cases": ["autonomous ml research", "ai-driven research", "ai-driven model benchmark improvement", "self-improving agent experiments", "optimizing code benchmarks", "automated llm experiments"], "compared": [{"slug": "claude", "name": "Claude", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [{"id": 756846576, "item_id": 47302911, "url": "https://news.ycombinator.com/item?id=47302911", "by": "mattlangston", "date": "2026-03-08", "stance": "neutral", "quote": "My fork of Andrej's repo if you want to join in", "excerpt": "I had the same SETI inspiration yesterday of \"autoresearch swarms\". Andrej and I had no coordination - we just both had the same idea at the same time. Experiments already running. Write up: \"Autoresearch on DGX Spark: Error Bars, GPU Pools, and the Case for an Autonomous Research Swarm\": https://github.com/matt-langston/autoresearch/discussions/1 My fork of Andrej's repo if you want to join in: https://github.com/ma", "context": "", "use_case": "Autonomous ML experiments", "switched_from": "", "story": "Autoresearch next step: async collaborative for agents SETI home style"}], "threads": [{"url": "https://news.ycombinator.com/item?id=47493460", "title": "I tried Karpathy's Autoresearch on an old research project", "mentions": 3}, {"url": "https://news.ycombinator.com/item?id=49762687", "title": "Science Is Open Software", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49274600", "title": "DeepSeek V4 Pro 0813", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49184960", "title": "Discovery Loop", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48891267", "title": "Show HN: DOM-docx \u2013 HTML to native, editable Word docs (MIT)", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48713682", "title": "Guy beats one of the best published quantum circuits for breaking ECDSA with AI", "mentions": 1}], "sentiment_by_month": {"2025-11": {}, "2025-12": {}, "2026-01": {}, "2026-02": {}, "2026-03": {"neutral": 3, "recommend": 3, "warn": 1, "compare": 1}, "2026-04": {"compare": 2, "neutral": 2}, "2026-05": {"recommend": 1, "neutral": 1}, "2026-06": {"recommend": 2}, "2026-07": {"recommend": 1}, "2026-08": {"neutral": 1, "recommend": 1}, "2026-09": {"neutral": 1}, "2026-10": {}}, "quotes": [{"id": 796309008, "item_id": 49769313, "url": "https://news.ycombinator.com/item?id=49769313", "by": "jegp", "date": "2026-09-19", "stance": "neutral", "quote": "This seems strikingly close to autoresearch", "excerpt": "I actually disagree that software is a static program. If a program always have the same behaviour, it's a pretty bad program. You need inputs and, as demonstrated by the recent LLM inputs, parameters you can tune to correctly address a specific context. I would also contest the point about science being social. To me, science is the interplay between social constructions and physical reality. Which is why tests and ", "context": "", "use_case": "", "switched_from": "", "story": "Science Is Open Software"}, {"id": 788508640, "item_id": 49281790, "url": "https://news.ycombinator.com/item?id=49281790", "by": "lightbulbish", "date": "2026-08-12", "stance": "recommend", "quote": "maybe this?", "excerpt": "maybe this? https://github.com/karpathy/autoresearch", "context": "", "use_case": "", "switched_from": "", "story": "DeepSeek V4 Pro 0813"}, {"id": 786992048, "item_id": 49187003, "url": "https://news.ycombinator.com/item?id=49187003", "by": "bredren", "date": "2026-08-05", "stance": "neutral", "quote": "hard to imagine they were not influenced by autoresearch.", "excerpt": "This seems to be an institutional, massively scaled version of https://github.com/karpathy/autoresearch. In March Karpathy described this direction: The next step for autoresearch is that it has to be asynchronously massively collaborative for agents (think: SETI@home style). Tweet is protected but in SERP caches: https://x.com/karpathy/status/2030705271627284816 Seems like Karpathy was largely focused on ML / SWE re", "context": "", "use_case": "autonomous ML research", "switched_from": "", "story": "Discovery Loop"}, {"id": 782281296, "item_id": 48892581, "url": "https://news.ycombinator.com/item?id=48892581", "by": "fishbone", "date": "2026-07-13", "stance": "recommend", "quote": "", "excerpt": "Thanks, and just to mention, I\u2019ve had awesome results using Autoresearch loops on other things like SQL performance. Prompt example: You are a SQL performance researcher. Run the following SQL query to establish a baseline, then come up with hypotheses to improve performance. Score each result and run 5 iterations. Avoid any regressions, each result must contain the exact same rows and columns. See https://github.com", "context": "", "use_case": "", "switched_from": "", "story": "Show HN: DOM-docx \u2013 HTML to native, editable Word docs (MIT)"}, {"id": 779420000, "item_id": 48713750, "url": "https://news.ycombinator.com/item?id=48713750", "by": "binyu", "date": "2026-06-28", "stance": "recommend", "quote": "which is extremely powerful when applied to problems that generalize well to it", "excerpt": "This is particularly relevant because the guy used Andrej Karphaty's \"autoresearch\" [1] approach, which is extremely powerful when applied to problems that generalize well to it. [1] https://github.com/karpathy/autoresearch", "context": "", "use_case": "AI-driven research", "switched_from": "", "story": "Guy beats one of the best published quantum circuits for breaking ECDSA with AI"}, {"id": 774466512, "item_id": 48404157, "url": "https://news.ycombinator.com/item?id=48404157", "by": "hgoel", "date": "2026-06-04", "stance": "recommend", "quote": "This is already known to work.", "excerpt": "As usual, I find the AI-related discussion here to be hopelessly hysterical and conspiratorial. I get the impression that a large chunk of people have only read the title and assumed Anthropic is referring to recursive self-improvement in the runaway singularity sense. One of the examples they provide, of giving Claude the task of training a small AI model, then asking it to improve certain benchmarks, is essentially", "context": "", "use_case": "AI-driven model benchmark improvement", "switched_from": "", "story": "When AI Builds Itself"}, {"id": 768817520, "item_id": 48051095, "url": "https://news.ycombinator.com/item?id=48051095", "by": "mkw5053", "date": "2026-05-07", "stance": "neutral", "quote": "the most viral lately is https://github.com/karpathy/autoresearch", "excerpt": "I feel like the most viral lately is https://github.com/karpathy/autoresearch", "context": "", "use_case": "", "switched_from": "", "story": "AlphaEvolve: Gemini-powered coding agent scaling impact across fields"}, {"id": 767565920, "item_id": 47972870, "url": "https://news.ycombinator.com/item?id=47972870", "by": "ramon156", "date": "2026-05-01", "stance": "recommend", "quote": "Seemed like that one was safe", "excerpt": "I was actually about to run kaparthy's autoresearch tool. Seemed like that one was safe https://github.com/karpathy/autoresearch", "context": "", "use_case": "", "switched_from": "", "story": "Shai-Hulud Themed Malware Found in the PyTorch Lightning AI Training Library"}, {"id": 767101504, "item_id": 47943844, "url": "https://news.ycombinator.com/item?id=47943844", "by": "pteetor", "date": "2026-04-28", "stance": "neutral", "quote": "In case you are unfamiliar with Karpathy's Loop[1]", "excerpt": "In case you are unfamiliar with Karpathy's Loop[1], it is a genetic algorithm[2] where the genetic \"mutations\" are clever-but-random ideas generated by an LLM agent, aimed at improving a system. (1) Let the LLM randomly perturbate the system. (2) Measure the system's performance. (3a) If the perturbation improved performance, keep the change. (3b) Otherwise, don't. (4) Repeat [1] https://github.com/karpathy/autoresea", "context": "", "use_case": "", "switched_from": "", "story": null}, {"id": 763972738, "item_id": 47748296, "url": "https://news.ycombinator.com/item?id=47748296", "by": "pravj", "date": "2026-04-12", "stance": "neutral", "quote": "Experimented with auto-research for prompt fine-tuning", "excerpt": "Grit Garden: https://grit.garden (https://github.com/pravj/wordle-garden) Recently shipped this personal art project that turns daily Wordle attempts into gritty / struggle-filled stories, kinda similar to the emotional arc of the Wordle game play. You can upload your own Wordle game screenshot to generate one for yourself. In addition to completing what was once in the idea list, I got to learn about - Prompt fine-t", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: What Are You Working On? (April 2026)"}, {"id": 763730416, "item_id": 47733151, "url": "https://news.ycombinator.com/item?id=47733151", "by": "coppsilgold", "date": "2026-04-11", "stance": "compare", "quote": "yet it's but a highly simplified version of AlphaEvolve", "excerpt": "LLMs are wordsmith oracles. A lot of effort went into trying to coax interactive intelligence from them but the truth is that you could have probably always harnessed the base models directly to do very useful things. The instruct tuned models give your harness even more degrees of freedom. A while ago, the autoresearch[1] harness went viral, yet it's but a highly simplified version of AlphaEvolve[2][3][4]. In the cy", "context": "", "use_case": "", "switched_from": "", "story": "Small models also found the vulnerabilities that Mythos found"}, {"id": 763320610, "item_id": 47707538, "url": "https://news.ycombinator.com/item?id=47707538", "by": "dataviz1000", "date": "2026-04-09", "stance": "compare", "quote": "The difference is autoresearch's val_bpb measures what the agent built", "excerpt": "(Sorry to spam.) I'm working on this also from a different angle. Hopefully sharing adds to the conversation. First, about the loop, Claude's (coding agent) context and attention is big enough to self-reflect. Agent Tuning shows a technique that not only demonstrates this but a way quantify it. [0] The difference is autoresearch's val_bpb measures what the agent built; Agent Tuning's p\u0302 measures the agent itself. > C", "context": "", "use_case": "", "switched_from": "", "story": "Research-Driven Agents: What Happens When Your Agent Reads Before It Codes"}, {"id": 760821600, "item_id": 47551350, "url": "https://news.ycombinator.com/item?id=47551350", "by": "latentsea", "date": "2026-03-27", "stance": "recommend", "quote": "autoresearch managed to improve the harness performance", "excerpt": "I'm currently running autoresearch against my harness that autonomously builds SaaS against an enforced architecture, and autoresearch managed to improve the harness performance on my 'time-to-Realworld' benchmark which has Claude Code drive the harness to build an implementation of https://github.com/realworld-apps/realworld with the win condition that it must pass my rigorous postman collection + playwright test su", "context": "", "use_case": "Self-improving agent experiments", "switched_from": "", "story": "HyperAgents"}, {"id": 759979824, "item_id": 47498739, "url": "https://news.ycombinator.com/item?id=47498739", "by": "Xx_crazy420_xX", "date": "2026-03-23", "stance": "compare", "quote": "Autoresearch is nothing new, big players are already in the game with more sophisticated solutions", "excerpt": "Autoresearch is nothing new, big players are already in the game with more sophisticated solutions: - https://arxiv.org/abs/2602.02660 (MARS) - https://arxiv.org/abs/2601.14525 (Execution-grounded automated AI research) - https://arxiv.org/abs/2601.10402 (ML-Master 2.0) The mostly used benchmark for automated AI engineering/ research is: https://github.com/openai/mle-bench", "context": "", "use_case": "", "switched_from": "", "story": "I tried Karpathy's Autoresearch on an old research project"}, {"id": 759906704, "item_id": 47494169, "url": "https://news.ycombinator.com/item?id=47494169", "by": "bethekind", "date": "2026-03-23", "stance": "recommend", "quote": "the ability to say: \"optimize this benchmark and commit only if you pass these tests\" is nice", "excerpt": "I used it to speed up an codecompass-like repo from 86 files per second to 2000. Still haven't used the repo in production, so maybe it secretly broke things, but the ability to say: \"optimize this benchmark and commit only if you pass these tests\" is nice", "context": "", "use_case": "Optimizing code benchmarks", "switched_from": "", "story": "I tried Karpathy's Autoresearch on an old research project"}, {"id": 759903808, "item_id": 47493988, "url": "https://news.ycombinator.com/item?id=47493988", "by": "dvt", "date": "2026-03-23", "stance": "warn", "quote": "likely not worth the $$$ spent on the tokens", "excerpt": "Ok, so looking at the commit log[1], I was mostly interested in seeing what the \"moonshot ideas\" implementations looked like, but basically everything is just hyperparameter tuning. Which is nice, but likely not worth the $$$ spent on the tokens. Am I missing something here? [1] https://github.com/ykumards/eCLIP/commits/main/autoresearch", "context": "", "use_case": "", "switched_from": "", "story": "I tried Karpathy's Autoresearch on an old research project"}, {"id": 758284177, "item_id": 47392761, "url": "https://news.ycombinator.com/item?id=47392761", "by": "dothack", "date": "2026-03-15", "stance": "neutral", "quote": "Take a look at the GitHub repo: \"forked from karpathy/autoresearch\"", "excerpt": "Take a look at the GitHub repo: \"forked from karpathy/autoresearch\"", "context": "", "use_case": "", "switched_from": "", "story": "Autoresearch Hub"}, {"id": 757855808, "item_id": 47365988, "url": "https://news.ycombinator.com/item?id=47365988", "by": "ozeron", "date": "2026-03-13", "stance": "recommend", "quote": "great find, create a claude code plugin", "excerpt": "great find, create a claude code plugin https://github.com/ozeron/autoresearch", "context": "", "use_case": "", "switched_from": "", "story": "Pi-Autoresearch"}, {"id": 757441088, "item_id": 47340068, "url": "https://news.ycombinator.com/item?id=47340068", "by": "janalsncm", "date": "2026-03-11", "stance": "neutral", "quote": "worth having a look at the changes the LLM did apply", "excerpt": "I don\u2019t disagree, but it\u2019s worth having a look at the changes the LLM did apply. https://github.com/karpathy/autoresearch/blob/master/progress.png My opinion is you\u2019d have to go pretty far down the x axis to get to anything that\u2019s not things like tinkering with bs, lr, or positional encodings. There are so many hyperparameter knobs already exposed that duplicating layers is unlikely to be proposed for a long time. I ", "context": "", "use_case": "automated LLM experiments", "switched_from": "", "story": "Show HN: How I Topped the HuggingFace Open LLM Leaderboard on Two Gaming GPUs"}, {"id": 756772784, "item_id": 47298299, "url": "https://news.ycombinator.com/item?id=47298299", "by": "karpathy", "date": "2026-03-08", "stance": "neutral", "quote": "when I had the model write up the results for that session", "excerpt": "So the interesting part about this one is that when I had the model write up the results for that session: https://github.com/karpathy/autoresearch/discussions/32 Look at its comment about this \"improvement\": \"\"\" Surprising non-results: - Changing random seed from 42\u2192137 improved by 0.0004. Seed 7 was worse. Make of that what you will. \"\"\" So the model knows! It knows that this is a weird thing to do after the fact. ", "context": "", "use_case": "", "switched_from": "", "story": "Autoresearch: Agents researching on single-GPU nanochat training automatically"}]}