@AYi_AInotes: Two Hong Kong students achieved a 5x speedup on Karpathy's automated research framework. They didn't switch to a stronger model, add more compute, or even change much code. They just added another loop on top of the original loop. This might be the most useful paper of the year for ordinary Agent developers, bar none. Here's the breakdown:
Summary
Two Hong Kong students achieved a 5x speedup by adding another loop outside the original automated research framework, without needing a better model or more compute. It is considered one of the most useful papers for ordinary Agent developers.
View Cached Full Text
Cached at: 07/10/26, 10:09 AM
Two Hong Kong students achieved 5x speed on Karpathy’s automatic research framework.
They didn’t use a stronger model, add more compute, or even change much code.
They just added another loop on top of the original loop.
This might be the most useful paper for ordinary agent developers this year, bar none. Here’s the breakdown: https://t.co/bIUoEFZI06
Similar Articles
@smallnest: I ported @karpathy's autoresearch to automated software development, and after various optimizations, the results are phenomenal.
A developer adapted Karpathy's autoresearch framework for automated software engineering, implementing multiple optimizations that yielded remarkable results.
@gyro_ai: The most painful part of reproducing a machine learning paper is that the paper is vague, key parameters are hidden in the appendix or even not written at all, and you spend most of your time playing detective instead of writing code. paper2code is an Agent skill: give it an arxiv link, and it generates a runnable implementation code. 1308 stars htt…
Paper2code is an AI Agent skill that generates runnable implementation code with citation anchors from an arxiv paper link, automatically audits ambiguities in the paper and marks unspecified parts, helping researchers and engineers efficiently reproduce machine learning papers.
@vintcessun: I always thought AI agents could only write ordinary code. Turns out MIT HAN Lab is directly using an agent workflow to design and optimize CUDA kernels. Hand-tuning is time-consuming and easy to miss solutions. They came up with a workflow of "task contract + agent loop + small-step verification", letting the agent research, implement, verify...
MIT HAN Lab proposes a method to automatically design and optimize CUDA kernels using an AI agent workflow. Through a process of task contracts, agent loops, and small-step verification, the agent can autonomously iterate and optimize within a specialized toolchain, replacing manual tuning.
@sitinme: Saw Karpathy open-sourced a very interesting project autoresearch, which gives a real but small-scale LLM training task to an AI Agent, letting it do research, modify code, run experiments, look at results, and then decide whether to keep or discard the changes. The project is based on a single NVIDIA…
Karpathy open-sourced an experimental project, autoresearch, that lets an AI Agent automatically complete the research loop for small-scale LLM training: modify code, run experiments, evaluate results, and iterate. Humans only need to write the research plan and constraints.
@jakevin7: Sharing something interesting Maka is currently working on: letting agents automatically optimize their own system prompt, fully closed-loop, without any human intervention. Karpathy's autoresearch, AEGIS, etc. have explored similar directions—a goal-driven self-reinforcement learning system.
Maka is a local-first desktop AI workbench whose new feature allows agents to automatically optimize their own system prompts by generating variants, using Harbor container evaluation, and an acceptance policy for iterative improvement, all without human intervention.