# What is Autoresearch? Source: https://www.youtube.com/watch?v=qb90PPbAWz4 Recap page: https://rapidrecap.app/video/qb90PPbAWz4 Generated: 2026-03-11T22:31:43.875+00:00 --- ## Quick Overview Andrej Karpathy's concept of "Autoresearch" involves using an AI agent in a continuous loop to plan, execute, and analyze short training experiments on AI models to achieve a specified goal, automating the tedious process of hyperparameter tuning and model improvement, which can be productized into niche 'agent in a box' products, marketing optimization services, internal productivity labs, or financial auto-quant tools. **Key Points:** - Autoresearch is defined as a super-nerd robot intern that runs science experiments on AI models overnight, iterating on plans, code edits, training, result reading, and deciding the next step. - The core loop involves: plan what to try, edit code, run a short training experiment (e.g., 5 minutes on a GPU), read results, and decide what to change next, repeating until better results are found. - Five potential business models utilizing autoresearch include: Niche 'agent in a box' products (like an Amazon listing experimenter), Print-money A/B testing for marketing, Research-as-a-Service, Power-tool inside your own product, and Agency selling '100x more testing' than competitors. - The Research-as-a-Service model involves a loop of search -> read -> summarize -> compare -> repeat, charging per report or via monthly subscription for dashboards. - The internal 'productivity lab' involves defining KPIs (like response time, close rate) and letting agents iterate on workflows, templates, and routing rules to reduce manual work and focus leadership on high-impact decisions. - The setup process shown requires Python 3.10+, an NVIDIA GPU (or equivalent like Apple M1 Pro with Metal/MPS support), installing UV, cloning Karpathy's GitHub repository, installing dependencies, running prepare.py, and then train.py. - Karpathy's work has inspired AgentHub, an agent-swarm collaboration platform described as the 'GitHub for agents.' ![Screenshot at 00:04: A screenshot of the GitHub repository displaying the session report graph showing performance improvement across 126 experiments, illustrating the iterative nature of the autoresearch process.](https://ss.rapidrecap.app/screens/qb90PPbAWz4/00-00-04.jpg) **Context:** The video discusses Andrej Karpathy's concept of "Autoresearch," which he recently launched as an open-source project, and explores ten potential business models derived from this technology. The presentation explains what autoresearch is—an automated AI agent loop for scientific experimentation—and shows installation steps via a Google Colab environment, emphasizing its potential to revolutionize AI development and various business operations. ## Detailed Analysis The video details Andrej Karpathy's concept of Autoresearch, an AI agent designed to automate the scientific experimentation loop for improving AI models. This involves the agent autonomously planning hyperparameter changes, editing Python code, running short training experiments (around 5 minutes on a GPU), analyzing metrics, and deciding the next steps, repeating this process until improvements stall. The speaker outlines ten business ideas leveraging this technology. Idea 1 is niche 'agent in a box' products, where tiny autoresearch loops are tuned for specific pain points (like Amazon listing optimization) and sold via monthly subscription. Idea 2 is "Print-money A/B testing for marketing," where the agent tests variants of ads and landing pages, focusing on conversion rate optimization (CRO). Idea 3, Research-as-a-Service, involves using the loop for market research, M&A decks, and compliance tracking, charging per report or monthly. Idea 4 is embedding an autoresearch-style agent as an 'Optimize' button within existing SaaS workflows to upsell pro/enterprise tiers. Idea 5 is an agency selling "100x more testing" than competitors, charging a monthly fee plus performance bonuses based on KPI lifts. Idea 6, Auto-quant for trading ideas, uses autoresearch to run fast backtests of trading rules overnight. Idea 7 is an "Always-on lead qualification + Follow-up" system integrated with a CRM to score leads and draft follow-ups. Idea 8 is a Finance Ops autopilot that grinds through invoice matching and expense report generation. Idea 9 proposes an internal 'productivity lab' for an organization, using agents to iterate on workflows and reduce manual grunt work. Finally, Idea 10 is a Done-for-you research/DD shop selling structured briefs and monthly updates to investors and acquirers. The speaker shows the installation steps on Google Colab, noting the need for an NVIDIA GPU, though he successfully ran it on an Apple M1 Pro, and mentions Yuchen Jin's new open-source project, AgentHub, which serves as a GitHub-like platform for agent collaboration. ### What Autoresearch Is - The AI agent plans experiments, edits Python code, runs short training experiments on a GPU (like a powerful gaming card) for about 5 minutes, reads results, decides the next step, and repeats the loop until it finds something better than before. ### Business Idea 1 - Niche 'agent in a box' products - Package tiny autoresearch loops tuned for one painful niche (e.g., Amazon listing experiments, pricing optimizer for SaaS) - Charge a monthly fee; value prop is running experiments 24/7 and showing the winner. ### Business Idea 2 - Print-money A/B testing for marketing - Use the same pattern, but swap training code for ad/landing page experiments - Agent writes variants of headlines, layouts, offers; auto-tests creatives, angles, and audiences; profit by running it for your own products or offering an 'always-on experiment engine' retainer service. ### Business Idea 3 - Research-as-a-service - Recipe is a loop: research search -> read -> summarize -> compare -> repeat (venturebeat+2) - Addresses money problems like constantly updated market reports, fast M&A diligence, and compliance tracking; charge per report or monthly subscription for dashboards. ### Business Idea 4 - Power-tool inside your own product - If you have a SaaS/workflow, embed an autoresearch-style agent users press 'optimize' to run mini-research loops (tune prompts, pick best pricing, rank suppliers) - Charge higher tiers for this feature and use it as a wedge to upsell pro/enterprise plans. ### Business Idea 5 - Agency that sells 'we run more tests than anyone else' - Pitch is 'We do 100x more testing than other shops, for the same or lower fee' - Niche examples include Shopify store conversion lab, B2B SaaS pricing experiment service, and email subject-line optimizer. ### Business Idea 6 - Auto-quant for trading ideas - Use autoresearch to run small, fast backtests of many simple trading rules (LLM-based, factor screens, sentiment filters) on one GPU overnight - Keep promising strategies, then trade your own capital or sell signals/strategy reports. ![Screenshot at 00:00: Andrej Karpathy being introduced as the 'godfather of AI' associated with OpenAI, Tesla, and Eureka Labs.](https://ss.rapidrecap.app/screens/qb90PPbAWz4/00-00-00.jpg) ![Screenshot at 00:04: A screenshot from Karpathy's GitHub showing the progress chart with 126 experiments improving performance over time.](https://ss.rapidrecap.app/screens/qb90PPbAWz4/00-00-04.jpg) ![Screenshot at 00:27: A slide detailing the simple mental model for how to use a 'research bot': 1. Write a clear task; 2. Give the bot access \(code/GPU or internet/documents\); 3. The bot runs a loop; 4. Come back later to see results.](https://ss.rapidrecap.app/screens/qb90PPbAWz4/00-00-27.jpg) ![Screenshot at 08:28: A flowchart illustrating the marketing A/B testing loop: Define goal -\> AI plans experiment -\> Edit code/settings -\> Run short training on GPU -\> Read metrics -\> Test if result is better -\> Save best config or log/discard, then plan next experiment.](https://ss.rapidrecap.app/screens/qb90PPbAWz4/00-08-28.jpg)