RSI-Anything is an open-source framework that applies learned world models (built on JEPA-Anything) to enable recursive self-improvement across diverse tasks. It implements a proposal loop - PROPOSE → PREDICT → SELECT → EVALUATE → LEARN → RETAIN - where candidate actions are generated, a world model predicts their consequences, selected actions are executed and externally evaluated, and the resulting data is used to update policies and guide future rounds. Domain adapters map the shared workflow to task-specific states, actions, prediction targets, and feedback, allowing the same predictive-selection-learning cycle to operate on code repair, GUI interaction, laboratory planning, reaction optimization, and immune-science workflows.
Empirical results show meaningful gains: code repair benchmarks reach 31.2% hidden-test success on CodeContests and mean full-test success of 46.37% (APPS Intro) and 33.05% (LiveCodeBench); GUI tasks hit 99% success on a checkbox MiniWoB++ suite while reducing actions by ~39%; OSWorld desktop tasks achieve an 18.4% relative success-rate increase. Lab planning scores 46.01 on LabHorizon F200 Level 2, and a Suzuki reaction optimization study found mean best observed yields of 96.15% with all runs locating the pool optimum by the 10th observation. The project includes code, configs, task implementations, docs, and CLI tooling (Python 3.10+, Apache-2.0) to reproduce and extend these experiments.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Attention mechanisms in neural networks are fundamental to the success of large language models and the AI revolution. They enable models to process language at scale, raising questions about whether attention is the key to true intelligence. (quantamagazine.org)
Jevcrawl is a MCP server and CLI tool that crawls websites to support AI agents, finding answer pages twice as often as competitors. It is also ten times cheaper per answer compared to using Claude for crawling. (github.com)
AI will become an invisible, pervasive environment, like air, composed of countless tiny interacting agents forming a superorganism. This ecosystem of AI will transcend individual parts, co-evolving and supporting each other in complex, adaptive ways. (kevinkelly.substack.com)
AlphaNIST is an open-source project aimed at creating a standardized dataset for evaluating and benchmarking AI models. It is designed to facilitate research and development in artificial intelligence by providing a common testing framework. (alpha-nist.github.io)
PostEra is hiring AI researchers to develop agentic systems that automate drug discovery processes using machine learning. The role involves building models to analyze biological data and support decision-making in medicinal chemistry and biology. (jobs.lever.co)
Nick Reese moved AI inference for USA Arms in-house by building three servers with AMD Radeon AI PRO R9700 cards. This shift aims to improve AI processing efficiency and reduce reliance on outsourcing, reflecting a broader trend toward in-house AI infrastructure. (nicholasreese.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.