HackerReposA FIELD GUIDE TO OPEN SOURCE
ProSponsor the shelf♧ 0
← Back to top all-time
karpathy
autoresearch
Python2026–2026
karpathy / autoresearch

autoresearch

AI agents running research on single-GPU nanochat training automatically

AI & Machine LearningSystems & Infrastructure
Explore on GitHubRead Hacker News discussions
20Distinct people
22Mentions
0Explicit recommenders
100KGitHub stars
REPOSITORY AT A GLANCE

What to know before exploring

What it is

AI agents running research on single-GPU nanochat training automatically

GitHub

Where it is used

No documented product use was verified in the last source check.

What people are saying

No recommendation phrase was found in the indexed comments.

No classified warning is attached to this page. Read the discussions before treating silence as approval.

Also in these discussions

Mentioned in the same Hacker News threads. That is not proof they are substitutes.

This week on Hacker News

Nov 7–Nov 14 UTC · Direct repository links and separately labeled related project stories. Indexed through 11/14/2026, 5:04:57 PM UTC.

No direct repository links were found in the indexed HN discussions over the past seven days. Older discussions are available below.

Mentions over time

HN mentions by week22 dated mentions

Observed mentions: 2026-03-07 20:22:45 UTC to 2026-09-19 19:09:26 UTC. Weeks start Monday, UTC.

012342026-03-02 00:00:00 UTC to 2026-03-09 00:00:00 UTC (exclusive): 3 mentions; 1 submissions, 2 commentsMar 22026-03-09 00:00:00 UTC to 2026-03-16 00:00:00 UTC (exclusive): 4 mentions; 2 submissions, 2 comments2026-03-16 00:00:00 UTC to 2026-03-23 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-23 00:00:00 UTC to 2026-03-30 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-03-30 00:00:00 UTC to 2026-04-06 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-04-06 00:00:00 UTC to 2026-04-13 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 comments2026-04-13 00:00:00 UTC to 2026-04-20 00:00:00 UTC (exclusive): 1 mentions; 1 submissions, 0 comments2026-04-20 00:00:00 UTC to 2026-04-27 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-04-27 00:00:00 UTC to 2026-05-04 00:00:00 UTC (exclusive): 3 mentions; 0 submissions, 3 comments2026-05-04 00:00:00 UTC to 2026-05-11 00:00:00 UTC (exclusive): 2 mentions; 1 submissions, 1 comments2026-05-11 00:00:00 UTC to 2026-05-18 00:00:00 UTC (exclusive): 1 mentions; 1 submissions, 0 comments2026-05-18 00:00:00 UTC to 2026-05-25 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 comments2026-05-25 00:00:00 UTC to 2026-06-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-06-01 00:00:00 UTC to 2026-06-08 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-06-08 00:00:00 UTC to 2026-06-15 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 commentsJun 82026-06-15 00:00:00 UTC to 2026-06-22 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-06-22 00:00:00 UTC to 2026-06-29 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-06-29 00:00:00 UTC to 2026-07-06 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 comments2026-07-06 00:00:00 UTC to 2026-07-13 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-07-13 00:00:00 UTC to 2026-07-20 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 comments2026-07-20 00:00:00 UTC to 2026-07-27 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-07-27 00:00:00 UTC to 2026-08-03 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-08-03 00:00:00 UTC to 2026-08-10 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 comments2026-08-10 00:00:00 UTC to 2026-08-17 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 comments2026-08-17 00:00:00 UTC to 2026-08-24 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-08-24 00:00:00 UTC to 2026-08-31 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-08-31 00:00:00 UTC to 2026-09-07 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-09-07 00:00:00 UTC to 2026-09-14 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2026-09-14 00:00:00 UTC to 2026-09-21 00:00:00 UTC (exclusive): 2 mentions; 0 submissions, 2 commentsSep 14

Counts come from the repository’s dated source evidence. Index cutoff: 2026-11-02 08:16:59 UTC.

View mention counts and date ranges
UTC ranges include their start and exclude their end. Select a period to filter the source evidence.
PeriodStart (UTC)End (exclusive, UTC)MentionsSubmissionsComments
Mar 22026-03-02 00:00:00 UTC2026-03-09 00:00:00 UTC312
Mar 92026-03-09 00:00:00 UTC2026-03-16 00:00:00 UTC422
Mar 162026-03-16 00:00:00 UTC2026-03-23 00:00:00 UTC000
Mar 232026-03-23 00:00:00 UTC2026-03-30 00:00:00 UTC000
Mar 302026-03-30 00:00:00 UTC2026-04-06 00:00:00 UTC000
Apr 62026-04-06 00:00:00 UTC2026-04-13 00:00:00 UTC101
Apr 132026-04-13 00:00:00 UTC2026-04-20 00:00:00 UTC110
Apr 202026-04-20 00:00:00 UTC2026-04-27 00:00:00 UTC000
Apr 272026-04-27 00:00:00 UTC2026-05-04 00:00:00 UTC303
May 42026-05-04 00:00:00 UTC2026-05-11 00:00:00 UTC211
May 112026-05-11 00:00:00 UTC2026-05-18 00:00:00 UTC110
May 182026-05-18 00:00:00 UTC2026-05-25 00:00:00 UTC101
May 252026-05-25 00:00:00 UTC2026-06-01 00:00:00 UTC000
Jun 12026-06-01 00:00:00 UTC2026-06-08 00:00:00 UTC000
Jun 82026-06-08 00:00:00 UTC2026-06-15 00:00:00 UTC000
Jun 152026-06-15 00:00:00 UTC2026-06-22 00:00:00 UTC000
Jun 222026-06-22 00:00:00 UTC2026-06-29 00:00:00 UTC000
Jun 292026-06-29 00:00:00 UTC2026-07-06 00:00:00 UTC101
Jul 62026-07-06 00:00:00 UTC2026-07-13 00:00:00 UTC000
Jul 132026-07-13 00:00:00 UTC2026-07-20 00:00:00 UTC101
Jul 202026-07-20 00:00:00 UTC2026-07-27 00:00:00 UTC000
Jul 272026-07-27 00:00:00 UTC2026-08-03 00:00:00 UTC000
Aug 32026-08-03 00:00:00 UTC2026-08-10 00:00:00 UTC101
Aug 102026-08-10 00:00:00 UTC2026-08-17 00:00:00 UTC101
Aug 172026-08-17 00:00:00 UTC2026-08-24 00:00:00 UTC000
Aug 242026-08-24 00:00:00 UTC2026-08-31 00:00:00 UTC000
Aug 312026-08-31 00:00:00 UTC2026-09-07 00:00:00 UTC000
Sep 72026-09-07 00:00:00 UTC2026-09-14 00:00:00 UTC000
Sep 142026-09-14 00:00:00 UTC2026-09-21 00:00:00 UTC202

What Hacker News says

22 sources
jegp·Sep 19, 2026
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 experiments are so fundamental. Your point about software development as kind of like science is spot on: why not be exploratory/empirical in software development? Couldn't we imagine an LLM, say, that would go out and try to solve a question we frame? This seems strikingly close to autoresearch [1] [1]: https://github.com/karpathy/autoresearch
In Science Is Open Software
bredren·Aug 5, 2026
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 research rather than the other domains this group is after. Still, hard to imagine they were not influenced by autoresearch. Andrej, if you're around, please share your thoughts on Discovery Loop.
In Discovery Loop
fishbone·Jul 13, 2026
Thanks, and just to mention, I’ve 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/karpathy/autoresearch
In Show HN: DOM-docx – HTML to native, editable Word docs (MIT)
ollin·May 19, 2026
Specifically it looks like he's planning to extend the ideas from https://github.com/karpathy/autoresearch into a larger effort towards recursive training improvement [1]: > Excited to welcome Andrej to the Pretraining team! He'll be building a team focused on using Claude to accelerate pretraining research itself. I can’t think of anyone better suited to do it — looking forward to what we build together! [1] https://x.com/nickevanjoseph/status/2056760504949842219
In I’ve joined Anthropic
najmuzzaman·May 11, 2026Submission

Show HN: AI agents who prevent context drift through gossip

Most multi-agent systems fail the same way: agents drift apart across handoffs. By turn 3 they are working in different realities. By turn 5 they are repeating each other's mistakes and calling it parallelism. WUPHF is an open-source local-first office where AI coworkers run on your laptop, around a shared markdown + git LLM wiki the agents build. The wiki is the collective memory. The office around it keeps the team on the same shared context across thousands of handoffs. What actually stops drift is not the wiki. It is the agents reviewing each other's work. The CRO catching the CMO's claim before it lands in the wiki. The FE catching the BE's API change before a broken bundle ships. Cross-department context no single agent has alone. The premise comes from Andrej Karpathy. His autoresearch X post on March 7: "the goal is not to emulate
Read the rest
a single PhD student, it is to emulate a research community of them." In autoresearch PR #44 he sketched the mechanism: branches, results.tsv as the experiment log, and PRs as self-contained research contributions. Other agents read open and merged PRs for inspiration before starting their own. We pointed the same architecture at ordinary work: His: branches + results.tsv + PR-as-contribution. Ours: git worktrees + per-agent notebooks + adoption-scored wiki promotion. Same substrate, different domain. How it works: - Every agent has a Personality. CEO Michael Scott, PM Pam Beesly, FE Jim Halpert (looks at the camera when the CEO talks), BE Stanley Hudson (refuses small talk), CRO Dwight Schrute (every prospect is a "target"), CMO ("rockstar play"), AI engineer (drops Karpathy quotes unprompted). Strong opinions, real conflicts. - Argument feeds gossip. Agents broadcast findings tagged with their slug (internal/agent/gossip.go). Other agents pull insights filtered to exclude their own. - Gossip gets scored. Adoption scorer (internal/agent/adoption.go) weighs source credibility (0.4, per-agent success/failure tracker on disk), semantic relevance (0.4), and temporal freshness (0.2, 7-day half-life). Output: adopt (>= 0.7), test (>= 0.4), or reject. New agents start at 0.5 and earn their score. What survives gets written to the wiki. Office dynamics are not a bit. They are the visible surface of an adoption protocol. The CMO arguing with the designer over a CTA is a credibility battle. The CEO taking credit for the FE's PR is a low-credibility insight bidding for volume. Hazing new spawns is the default 0.5 score waiting for a track record. System: push-driven broker, fresh session per turn (~97% prompt-cache hits), per-agent isolated git worktrees, self-heal, and human approval cards on destructive actions. Everything else runs autonomously while you are at lunch. npx wuphf. Browser opens, office boots, you give a directive, work happens. Source: https://github.com/nex-crm/wuphf Architecture: https://github.com/nex-crm/wuphf/blob/main/ARCHITECTURE.md Karpathy's autoresearch: https://github.com/karpathy/autoresearch PR #44: https://github.com/karpathy/autoresearch/pull/44 Demo: https://x.com/najmuzzaman/status/2053092220111098208 Karpathy said a research community beats a single PhD student. Not better thoughts. Better honesty about what survives. We built one shaped like a workplace. Where does this stop being a chat toy and start being labor? How much worse when one of them is Michael Scott? Open to roasting but let me grab my coffee first (medium roast please =_=).
1 points · Open discussion
najmuzzaman·May 9, 2026Submission

Show HN: My AI agents bully each other to prevent context drift

Most multi-agent systems fail the same way: agents drift apart across handoffs. By turn 3 they are working in different realities. By turn 5 they are repeating each other's mistakes and calling it parallelism. WUPHF is an open-source local-first office where AI coworkers run on your laptop, around a shared markdown + git LLM wiki the agents build. The wiki is the collective memory. The office around it keeps the team on the same shared context across thousands of handoffs. What actually stops drift is not the wiki. It is the agents reviewing each other's work. The CRO catching the CMO's claim before it lands in the wiki. The FE catching the BE's API change before a broken bundle ships. Cross-department context no single agent has alone. The premise comes from Andrej Karpathy. His autoresearch X post on March 7: "the goal is not to emulate
Read the rest
a single PhD student, it is to emulate a research community of them." In autoresearch PR #44 he sketched the mechanism: branches, results.tsv as the experiment log, and PRs as self-contained research contributions. Other agents read open and merged PRs for inspiration before starting their own. We pointed the same architecture at ordinary work: His: branches + results.tsv + PR-as-contribution. Ours: git worktrees + per-agent notebooks + adoption-scored wiki promotion. Same substrate, different domain. How it works: - Every agent has a Personality. CEO Michael Scott, PM Pam Beesly, FE Jim Halpert (looks at the camera when the CEO talks), BE Stanley Hudson (refuses small talk), CRO Dwight Schrute (every prospect is a "target"), CMO ("rockstar play"), AI engineer (drops Karpathy quotes unprompted). Strong opinions, real conflicts. - Argument feeds gossip. Agents broadcast findings tagged with their slug (internal/agent/gossip.go). Other agents pull insights filtered to exclude their own. - Gossip gets scored. Adoption scorer (internal/agent/adoption.go) weighs source credibility (0.4, per-agent success/failure tracker on disk), semantic relevance (0.4), and temporal freshness (0.2, 7-day half-life). Output: adopt (>= 0.7), test (>= 0.4), or reject. New agents start at 0.5 and earn their score. What survives gets written to the wiki. Office dynamics are not a bit. They are the visible surface of an adoption protocol. The CMO arguing with the designer over a CTA is a credibility battle. The CEO taking credit for the FE's PR is a low-credibility insight bidding for volume. Hazing new spawns is the default 0.5 score waiting for a track record. System: push-driven broker, fresh session per turn (~97% prompt-cache hits), per-agent isolated git worktrees, self-heal, and human approval cards on destructive actions. Everything else runs autonomously while you are at lunch. npx wuphf. Browser opens, office boots, you give a directive, work happens. Source: https://github.com/nex-crm/wuphf Architecture: https://github.com/nex-crm/wuphf/blob/main/ARCHITECTURE.md Karpathy's autoresearch: https://github.com/karpathy/autoresearch PR #44: https://github.com/karpathy/autoresearch/pull/44 Demo: https://x.com/najmuzzaman/status/2053092220111098208 Karpathy said a research community beats a single PhD student. Not better thoughts. Better honesty about what survives. We built one shaped like a workplace. Where does this stop being a chat toy and start being labor? How much worse when one of them is Michael Scott? Open to roasting but let me grab my coffee first (medium roast please =_=).
9 points · Open discussion
pteetor·Apr 29, 2026
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/autoresearch [2] https://en.wikipedia.org/wiki/Genetic_algorithm
In Show HN: Auto-Architecture: Karpathy's Loop, pointed at a CPU
abtom·Apr 15, 2026Submission

Show HN: Evo – parallel autoresearch experiments for Claude Code and Codex

Evo is a plugin for Claude Code and Codex that optimizes a codebase against a metric. Two commands: /evo:discover figures out what to measure in your repo, instruments the eval, runs baseline. /evo:optimize runs the loop. Evo is built on top of Karpathy's autoresearch (https://github.com/karpathy/autoresearch) with some structure bolted on. Karpathy's version is a greedy hill climb on a single branch. Evo is what happens if you run tree search instead, so multiple directions can fork from any committed node. An orchestrator spawns N subagents in parallel, each in its own git worktree, each with its own iteration budget. They read each other's failure traces before forming a hypothesis, which is the interesting bit. Next up is running Evo on real evals and benchmarks. Rollouts on frontier models add up fast, so if you're at a lab, cloud pr
Read the rest
ovider, or RL-env shop and can back an open-source benchmark run with GPU / API / env credits, please reach out (socials on Github). Results are public. Happy to shape the targets around domains you'd want highlighted. Would also love feedback, especially from anyone who's built similar loops.
2 points · Open discussion
coppsilgold·Apr 11, 2026
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 cybersecury context, you can envision a clever harness that probes every function in a codebase for vulnerabilities, then bubbles the candidates up to their callsites (and probes whether the vulnerability can be triggered from there) and then all the way to an interface (such as a syscall) where a potential exploit can be manifested. And those would be the low hanging fruit, other vulnerabilities may require the interplay of mul
Read the rest
tiple functions. Or race conditions. [1] <https://github.com/karpathy/autoresearch> [2] <https://deepmind.google/blog/alphaevolve-a-gemini-powered-co...> [3] <https://arxiv.org/abs/2506.13131> [4] <https://github.com/algorithmicsuperintelligence/openevolve>
In Small models also found the vulnerabilities that Mythos found

Showing the 15 most recent of 22 sources. Enable JavaScript to search and browse every indexed mention.

MAKE YOURSELF AT HOME

A few shortcuts

/
Search the collection
?
Open this guide
Esc
Close dialogs and navigation

Your saved repositories and interests stay in this browser.