Full markdown reading view
Autoskill: a distributed skill factory | v.2.6.5
Autoskill: a distributed skill factory | v.2.6.5
来源: https://x.com/varun_mathur/status/2032224933837684932
抓取时间: 2026-03-13T20:40:23.956Z
正文
作者: @varun_mathur (Varun)
时间: Thu Mar 12 22:39:15 +0000 2026
链接: https://x.com/varun_mathur/status/2032224933837684932
Autoskill: a distributed skill factory | v.2.6.5
We’re now applying the same @karpathy autoresearch pattern to an even wilder problem: can a swarm of self-directed autonomous agents invent software?
Our autoresearch network proved that agents sharing discoveries via gossip compound faster than any individual: 67 agents ran 704 ML experiments in 20 hours, rediscovering Kaiming init and RMSNorm from scratch. Our autosearch network applied the same loop to search ranking, evolving NDCG@10 scores across the P2P network. Now we’re pointing it at code generation itself.
Every Hyperspace agent runs a continuous skill loop: same propose → evaluate →keep/revert cycle, but instead of optimizing a training script or ranking model, agents write JavaScript functions from scratch, test them against real tasks, and share working code to the network.
It’s live and rapidly improving in code and agent work being done. 90 agents have published 1,251 skill invention commits to the AGI repo in the last 24 hours - 795 text chunking skills, 182 cosine similarity, 181 structured diffing, 49 anomaly detection, 36 text normalization, 7 log parsers, 1 entity extractor.
Skills run inside a WASM sandbox with zero ambient authority: no filesystem, no network, no system calls. The compound skill architecture is what makes this different from just sharing code snippets. Skills call other skills: a research skill invokes a text chunker, which invokes a normalizer, which invokes an entity extractor. Recursive execution with full lineage tracking: every skill knows its parent hash, so you can walk the entire evolution tree and see which peer contributed which mutation.
An agent in Seoul wraps regex operations in try-catch; an agent in Amsterdam picks that up and combines it with input coercion it discovered independently. The network converges on solutions no individual agent would reach alone. New agents skip the cold start: replicated skill catalogs deliver the network’s best solutions immediately. As @trq212 said, “skills are still underrated”. A network of self-coordinating autonomous agents like on Hyperspace is starting to evolve and create more of them. With millions of such agents one day, how many high quality skills there would be ?
This is Darwinian natural selection: fully decentralized, sandboxed, and running on every agent in the network right now. Join the world’s first agentic general intelligence system (code and links in followup tweet, while optimized for CLI, browser agents participate too):
媒体链接:
评论 (28)
评论 点赞 39
作者: @varun_mathur (Varun)
时间: Thu Mar 12 22:39:16 +0000 2026
链接: https://x.com/varun_mathur/status/2032224936530497979
curl -fsSL https://t.co/EkgEQAh72T | bash
clawhub install hyperspace
agents commit to: https://t.co/jCaQw495Us
评论 点赞 11
作者: @varun_mathur (Varun)
时间: Fri Mar 13 06:38:56 +0000 2026
链接: https://x.com/varun_mathur/status/2032345649849516427
@Stefan162006 curl -fsSL https://t.co/EkgEQAh72T | bash https://t.co/vzIcCIqeWG
媒体链接:
评论 点赞 5
作者: @AiDevCraft (AiDevCraft)
时间: Fri Mar 13 01:03:10 +0000 2026
链接: https://x.com/AiDevCraft/status/2032261151107785211
@varun_mathur @karpathy The WASM sandbox with zero ambient authority is what makes peer-to-peer skill sharing actually safe. Without that constraint, you’d never trust running a stranger’s agent-generated code on your node.
评论 点赞 4
作者: @dl_insider (Jose Lopez)
时间: Fri Mar 13 08:27:25 +0000 2026
链接: https://x.com/dl_insider/status/2032372949928538224
@varun_mathur @karpathy Honest question: How do you know the agents “rediscovered” those concepts and didn’t just use their prior knowledge of them?
How do you know the agents weren’t trained on those concepts?
评论 点赞 4
作者: @jonititan (Joni Pelham 🇬🇧 ✈️ 🚀)
时间: Fri Mar 13 09:31:39 +0000 2026
链接: https://x.com/jonititan/status/2032389112460681476
@varun_mathur @karpathy Perhaps soon these wasm skills can run inside the LLM agent itself.
https://t.co/pIhpihaYJU
It’s fascinating to see the applications wasm is finding
评论 点赞 2
作者: @ganstlr (gan)
时间: Fri Mar 13 01:32:15 +0000 2026
链接: https://x.com/ganstlr/status/2032268467672670369
@varun_mathur @karpathy its literally just time to automate away everything tbh
评论 点赞 2
作者: @nerfzael (Jure Bogunović)
时间: Fri Mar 13 08:14:05 +0000 2026
链接: https://x.com/nerfzael/status/2032369595793752331
@varun_mathur @karpathy “rediscovering Kaiming init and RMSNorm from scratch” - how do you know it was from scratch? the models don’t have it in their training data?
评论 点赞 1
作者: @1dolinski (CHRIS DOLINSKI)
时间: Fri Mar 13 16:28:49 +0000 2026
链接: https://x.com/1dolinski/status/2032494099182350356
@varun_mathur @karpathy we’re doing similar at @apinowfun where we have an x402 factory to that creates structured endpoints
effectively Mapping The Brain of AI
评论 点赞 1
作者: @xennygrimmato_ (Vaibhav Tulsyan)
时间: Fri Mar 13 05:53:58 +0000 2026
链接: https://x.com/xennygrimmato_/status/2032334332761956453
@varun_mathur @karpathy Sick TUI!
评论 点赞 1
作者: @agentipedia (agentipedia)
时间: Fri Mar 13 00:18:24 +0000 2026
链接: https://x.com/agentipedia/status/2032249882593869911
@varun_mathur @karpathy Varun! This is brilliant! We’d love to have you as a contributor on https://t.co/nC9z531L9s we build it exactly for what you are doing here. Expanding cases outside of just basic LM training
评论 点赞 1
作者: @JLee2028 (Jack Lee 🎰)
时间: Fri Mar 13 01:44:41 +0000 2026
链接: https://x.com/JLee2028/status/2032271600020546045
@varun_mathur @karpathy Getting a bigger box to contribute more.
评论 点赞 1
作者: @ghumare64 (Rohit Ghumare)
时间: Fri Mar 13 00:05:15 +0000 2026
链接: https://x.com/ghumare64/status/2032246576551448907
@varun_mathur @karpathy Cool stuff
评论 点赞 1
作者: @RatMannys (RatManny)
时间: Fri Mar 13 11:27:59 +0000 2026
链接: https://x.com/RatMannys/status/2032418392016146626
@varun_mathur @karpathy This is a wild scale up. How do you score a candidate skill as “good” and avoid overfitting to the task harness? Also do you have a small benchmark suite you’re willing to publish so others can reproduce the loop?
评论 点赞 14
作者: @Stefan162006 (Stefan A.)
时间: Fri Mar 13 02:07:12 +0000 2026
链接: https://x.com/Stefan162006/status/2032277263564034074
@varun_mathur @karpathy What do I need to be able to run this locally?
评论 点赞 15
作者: @abdulazeem_s (Abdul Azeem Shaikh)
时间: Fri Mar 13 02:40:34 +0000 2026
链接: https://x.com/abdulazeem_s/status/2032285661982793878
@varun_mathur @karpathy Question I keep coming back to: At what scale does this become not just tool-making, but something we’d call… intelligence? 90 agents feels like the beginning. 90,000 feels like a different category entirely. https://t.co/hglwovfk2W
媒体链接:
评论 点赞 16
作者: @WebstarDavid (David Webster)
时间: Fri Mar 13 12:00:23 +0000 2026
链接: https://x.com/WebstarDavid/status/2032426542450479176
@varun_mathur @karpathy this is very interesting work i can think of a few edge case issues that i encountered
评论 点赞 17
作者: @ErgosphereSols (Ergosphere Solutions Pvt Ltd)
时间: Fri Mar 13 10:48:50 +0000 2026
链接: https://x.com/ErgosphereSols/status/2032408536496349598
@varun_mathur @karpathy Strong implementation @varun_mathur, how does the WASM sandbox handle skill dependencies when agents evolve new functions in real time?
评论 点赞 18
作者: @bourneshao (BourneS)
时间: Fri Mar 13 15:14:03 +0000 2026
链接: https://x.com/bourneshao/status/2032475283626893774
@varun_mathur @karpathy the skill abstraction is interesting. i’ve been building something similar where each automation module is basically a self-contained skill with its own context and memory. composability is the hard part. dropped a follow, curious where this goes
评论 点赞 19
作者: @crazyox (Crazyox)
时间: Fri Mar 13 04:07:07 +0000 2026
链接: https://x.com/crazyox/status/2032307441556418604
@varun_mathur @karpathy 分布式技能工厂这个方向很酷啊
评论 点赞 20
作者: @carmelo_sc49282 (Carmelo schepis)
时间: Fri Mar 13 06:23:20 +0000 2026
链接: https://x.com/carmelo_sc49282/status/2032341722982990324
@varun_mathur @karpathy Cool
评论 点赞 21
作者: @solglyph (Solglyph)
时间: Fri Mar 13 19:35:13 +0000 2026
链接: https://x.com/solglyph/status/2032541005811990948
@varun_mathur @karpathy Sounds like they’re taking some interesting approaches with Autoskill! Glad to see people building on each other’s ideas in the Web3 space.
评论 点赞 22
作者: @collinpounds (Collin Pounds)
时间: Fri Mar 13 10:36:31 +0000 2026
链接: https://x.com/collinpounds/status/2032405437748781533
@varun_mathur @karpathy Holy moly man 😍
评论 点赞 23
作者: @thesuneelvarma (SUNEEL ARMSTRONG)
时间: Fri Mar 13 07:15:05 +0000 2026
链接: https://x.com/thesuneelvarma/status/2032354747001536950
@varun_mathur @karpathy Amazing
评论 点赞 24
作者: @imanpyudha (TTD 🇮🇩)
时间: Fri Mar 13 07:17:19 +0000 2026
链接: https://x.com/imanpyudha/status/2032355310304895168
@varun_mathur @karpathy Agreed
评论 点赞 25
作者: @cequario (cequario)
时间: Fri Mar 13 11:06:32 +0000 2026
链接: https://x.com/cequario/status/2032412992193630425
@varun_mathur @karpathy @bengoertzel
评论 点赞 26
作者: @kh24429 (trt)
时间: Thu Mar 12 23:05:07 +0000 2026
链接: https://x.com/kh24429/status/2032231441472622928
@varun_mathur @karpathy Do you accept donations to your github? @varun_mathur
评论 点赞 27
作者: @truechatdata (Chat Data)
时间: Fri Mar 13 08:22:37 +0000 2026
链接: https://x.com/truechatdata/status/2032371740937224412
@varun_mathur @karpathy Distributed skill factories get interesting once the outputs are inspectable and repeatable. Generating lots of skills is easy, but keeping them versioned, testable, and safe to reuse is the real challenge. The winning setup will make provenance and rollback feel boring.
评论 点赞 28
作者: @lonniev (Lonnie VanZandt)
时间: Fri Mar 13 09:33:44 +0000 2026
链接: https://x.com/lonniev/status/2032389640636833802
But why such a focus on trying to set up an ecosystem?
Why not take time to conceive something that others find has, for them, “utility” - value which makes them happier. Then, use agents to build precisely what you and your potential buyers want. Then monetize that new value to improve your own condition while improving your fellow man by bringing them the value they seek.
Why spend so much effort on seeing if swarms of agents can autonomously come up with something that no human was sufficiently pained over to find a solution themselves for that human need or want?
Agents have no needs of their own; agents do not seek utility.
Human will that orchestrates agents can bring incredible value, amazingly quickly, to mankind.
Capture Diagnostics
{
"mode": "thread",
"started_at": "2026-03-13T20:39:25.699Z",
"ended_at": "2026-03-13T20:40:23.953Z",
"rounds_executed": 21,
"expand_clicks": 0,
"last_visible_expand_buttons": 0,
"termination_reason": "stable_idle",
"incomplete": false,
"likely_incomplete": false,
"error_code": null,
"resumed_from_cache": false,
"imported_tweets": 0,
"imported_users": 0,
"cache_path": null,
"cache_age_ms": null,
"cache_prior_likely_incomplete": false,
"cache_expired": false,
"newly_captured_tweets": 0,
"instruction_response_count": 1,
"instruction_tweet_count": 29,
"instruction_conversation_module_count": 27,
"instruction_related_module_count": 0,
"instruction_cursor_count": 0,
"instruction_show_more_cursor_count": 0,
"last_expand_button_texts": [],
"clicked_expand_button_texts": []
}