40.9K
_DGh: github.com/karpathy/autoresearch
Really cool project that both shows the trend for agents and what the core loop for them is, although graphs can still be deceiving
Always love a Karpathy post!
@_dbrogle
#Autoresearch
世界中の人々によるAutoresearchに関する件のリール動画を視聴。
ログインせずに匿名で視聴。
関連検索
30andrej karpathy autoresearchkarpathy autoresearchpi autoresearchautoresearch andrej karpathykarpathy autoresearch projectkarpathy autoresearch github what is autoresearchkarpathy autoresearch github programmdkarpathy autoresearch release date or announcementtobi lutke autoresearchandre karpathy autoresearch
トレンドリール
(12)
37.2K
RAThis AI system called autoresearch can replace marketers and runs experiments for you while you sleep.
Ive been playing around with it for trading bots and its super cool that it can improve on its own #ai #tech
@raycfu
2.9K
TOA PhD thesis that would normally take 7 years. Done in 30 minutes. On a desktop computer 🧪
How long before traditional research timelines become irrelevant? 👇
📌 Follow @todayin_ai for breakthroughs happening in real time
#autoresearch #genomics #ainews #todayinai
@todayin_ai
203.2K
DIAI startups are now crossing a line they weren’t supposed to (just yet)
1. Autoresearch project built by Andrej Karpathy
Basically, bring your GPU and train your own AI model just using AI
Potential applications: custom fine tuning models for specialized tasks without ML/technical expertise
2. AI Agency - instead of one big AI trying to do everything, structure agents more like a company with specialized skills, clear responsibilities and scope of work
3. Paperclip - AI company running a business autonomously
These are all in early stages so don’t expect them to change the world just but just glimpses of what’s to come (if you can afford to pay for all those tokens)
#aitools #aiupdates #digitalsamaritan
@digitalsamaritan
41.2K
ARAndrej Karpathy just released AutoResearch, a minimal 630-line framework that hands a real GPU training loop to an AI coding agent and lets it iterate autonomously. 12,000 stars and 1,600 forks in four days.
The setup takes under five minutes. Clone the repo, run the data prep once, open the folder in Claude or Cursor, and prompt the agent to start. It reads your strategy from a Markdown file called program.md, creates a git branch, edits the training script, runs a five-minute experiment, and commits if the validation loss improved. Typical throughput is about 12 experiments per hour.
What the video couldn’t cover:
The evaluation metric is val_bpb, validation bits-per-byte, which is vocabulary-size independent. This means architectural changes, optimizer swaps, and hyperparameter tweaks are all directly comparable without normalization hacks.
The base repo is NVIDIA-only and tested on H100, but community forks already exist for Mac via MLX and Windows RTX setups. Results are non-portable across hardware because the five-minute wall-clock budget is the control variable.
Karpathy’s vision for the next iteration is asynchronous multi-agent swarms. Agents would post mini-papers in GitHub Discussions, read each other’s findings via CLI, and explore parallel research directions without forcing merges. He described the experience as “post-AGI feels” and “early singularity fun.”
Compared to Sakana AI’s AI Scientist, which attempts the full research pipeline from idea generation to LaTeX paper writing, AutoResearch deliberately stays narrow. One training script, one objective metric, one tight loop. Many in the community note that this tight loop would be a natural low-level engine inside a broader system like Sakana’s.
The repo is at github.com/karpathy/autoresearch. Requires an NVIDIA GPU, Python 3.10+, and uv for package management. If you have the hardware, an overnight run is genuinely worth trying.
#ai #artificialintelligence #technews #chatgpt #algorithm
@arnitly
3.6K
PRApplying an autoresearch agent to improve @phiaco recommendation quality (as a PM with no coding experience) 🤓👚
#phia #fashion #agenticai #productmanager #claude
@prerana.talks
148.4K
MAAndrej Karpathy open-sourced a 630-line Python tool that lets an AI agent run hundreds of ML experiments overnight — and it found a bug he’d missed after years of manual tuning. The catch: it only works in tightly constrained problems with clear metrics. Not replacing researchers, just the grunt work.
Source: https://venturebeat.com/technology/andrej-karpathys-new-open-source-autoresearch-lets-you-run-hundreds-of-ai
#AIResearch #Karpathy #MachineLearning #Autoresearch #OpenSource
@mansispeaks_
2.3K
NOKarpathy’s AutoResearch method: remove yourself from the loop.
Just give the agent the objective, metric, and boundaries. And go. It works.
Full episode with @saranormous on No Priors. Listen now on YouTube or your favorite podcast player.
@nopriorspod
28.9K
IHNon ML Researcher shows us AGI / Singularity! (AutoResearch)
@iharnoor
5.4K
TAAndrej Karpathy just built something that does AI research while you sleep.
Comment “karpathy” for the github link!
56,000 people starred this project in days. Here is why.
Karpathy, one of the most respected AI researchers alive, built a tool called autoresearch.
The concept is simple and the implications are massive.
You give an AI agent a real machine learning training setup. You go to sleep. The agent runs experiments on its own. It modifies the code. Trains for five minutes. Checks if the result improved. Keeps what works. Discards what fails. Repeats all night.
You wake up to a full log of experiments and a better model.
Here is exactly what I am testing with it.
I run an AI agency. My entire business is built on building systems that get smarter over time for the businesses I serve. So the first thing I am training this on is my own caption and reel scripts. Every piece of content I have written. Every hook. Every framework. Let it find the patterns that made the best ones work and build a model around that.
Then I am taking it to clients. Feed it their cold email sequences. Their sales call transcripts. Their support tickets. Let it run experiments overnight on what combinations produce the best outcomes.
Wake up to a model trained specifically on their business.
That is not a generic AI tool. That is a custom brain built from their own data. And that is a premium deliverable worth serious money.
This is the direction I am building GoToAgents.ai toward. Autonomous systems that improve themselves while you sleep.
@tamnguyen_ai
4.9K
ADComment “Research” and I’ll send you the step-by-step guide.
This guy just open-sourced a way to let AI run its own experiments and improve itself while you sleep.
It’s called autoresearch and here’s exactly what it does.
Training an AI model normally takes a ton of work.
You make a change.
Run a test.
Wait for results.
Decide if it worked.
Then do it all again.
Most researchers can only get through a few experiments a day.
Autoresearch hands that whole process to an AI agent.
Every 5 minutes it makes a change, runs the test, and checks if the model got better.
If it did, it keeps it. If not, it rolls back.
Then it tries something new… and it can be running this all night long.
The results are already wild.
Karpathy the creator ran 126 experiments in a single overnight session.
Tobi Shopify’s CEO set it up before bed.
By morning his smaller model was beating a model twice its size.
Someone ran it on a Mac Mini and the model improved by getting simpler.
Currently the main repo only works with an NVIDIA GPU right now.
But there are Mac and Windows forks if you need them.
@adamstewartmarketing
1.5K
REAutoresearch might actually be the wave to self improve your Claude skills #ai #claude #autoresearch
@reyhanmerekar
✨ #Autoresearch発見ガイド
Instagramには#Autoresearchの下にthousands of件の投稿があり、プラットフォームで最も活気のあるビジュアルエコシステムの1つを作り出しています。
Instagramの膨大な#Autoresearchコレクションには、今日最も魅力的な動画が掲載されています。@digitalsamaritan, @mansispeaks_ and @arnitlyや他のクリエイティブなプロデューサーからのコンテンツは、世界中でthousands of件の投稿に達しました。
#Autoresearchで何がトレンドですか？最も視聴されたReels動画とバイラルコンテンツが上部に掲載されています。
人気カテゴリー
📹 ビデオトレンド: 最新のReelsとバイラル動画を発見
📈 ハッシュタグ戦略: コンテンツのトレンドハッシュタグオプションを探索
🌟 注目のクリエイター: @digitalsamaritan, @mansispeaks_, @arnitlyなどがコミュニティをリード
#Autoresearchについてのよくある質問
Pictameを使用すれば、Instagramにログインせずに#Autoresearchのすべてのリールと動画を閲覧できます。あなたの視聴活動は完全にプライベートです。ハッシュタグを検索して、トレンドコンテンツをすぐに探索開始できます。
パフォーマンス分析
12リールの分析
💡 トップ投稿は平均108.4K回の再生（平均の2.5倍）
週3-5回、活動時間に定期的に投稿
コンテンツ作成のヒントと戦略
🔥 #Autoresearchは高いエンゲージメント可能性を示す - ピーク時に戦略的に投稿
✨ 多くの認証済みクリエイターが活動中（67%） - コンテンツスタイルを研究
✍️ ストーリー性のある詳細なキャプションが効果的 - 平均長594文字
📹 #Autoresearchには高品質な縦型動画（9:16）が最適 - 良い照明とクリアな音声を使用
#Autoresearch に関連する人気検索
🎬動画愛好家向け
Autoresearch ReelsAutoresearch動画を見る
📈戦略探求者向け
Autoresearchトレンドハッシュタグ最高のAutoresearchハッシュタグ
🌟もっと探索
Autoresearchを探索#andrej karpathy autoresearch#karpathy autoresearch#pi autoresearch#autoresearch andrej karpathy#karpathy autoresearch project#karpathy autoresearch github what is autoresearch#karpathy autoresearch github programmd#karpathy autoresearch release date or announcement