Wednesday — September 10, 2025

A hacker used a "memory mailbox" technique to add a live LLM to Animal Crossing, Anthropic's $1.5B AI copyright settlement was rejected by a federal judge, and researchers introduced R-Zero, a self-evolving LLM framework that generates its own training data.

News

I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

The author of the text created a bridge between the 2001 Nintendo GameCube game Animal Crossing and a cloud-based AI, allowing the game to generate new, dynamic dialogue without modifying the original game code. This was achieved by using a "memory mailbox" technique, where an external Python script writes data directly to the GameCube's RAM, which the game can then read and use to display new dialogue.

Anthropic judge rejects $1.5B AI copyright settlement

A federal judge has expressed concerns over a proposed $1.5 billion copyright settlement between Anthropic PBC and authors, stating that the agreement is "nowhere close to complete" and that he feels "misled" by the parties involved. The judge, William Alsup, has denied the motion to approve the deal without prejudice, seeking further clarification on the claim process for class members before making a decision.

Geoffrey Hinton: 'AI will make a few people much richer and most people poorer'

Computer scientist Geoffrey Hinton warns that AI will exacerbate income inequality, making a few people much richer while leaving most people poorer. To read the full article, a subscription to the Financial Times is required, with options starting at $1 for 4 weeks or $45 per month for standard digital access.

Apple barely talked about AI at its big iPhone 17 event

Apple's recent iPhone 17 event was notable for its lack of discussion about AI, a departure from last year's event where Apple Intelligence was a major focus. The company did mention AI in the context of powering background features, such as improved game performance and live translation, but it did not highlight AI as a major selling point for its new devices, unlike some of its competitors like Google and Samsung.

I don't want AI agents controlling my laptop

Granting unrestricted access to AI tools on personal computers can be risky due to the lack of strong security boundaries between different applications and processes. Potential solutions to this issue include using cloud environments or virtual machines, which provide isolation and reproducibility, or integrating AI tools into web browsers, which can enforce boundaries and allow for more fine-grained control over access to sensitive information.

Research

R-Zero: Self-Evolving Reasoning LLM from Zero Data

Self-evolving Large Language Models (LLMs) can achieve super-intelligence through autonomous learning, but current training methods rely on human-curated data, limiting their potential. The R-Zero framework overcomes this by generating its own training data, using two co-evolving models, a Challenger and a Solver, to create a self-improving curriculum that substantially boosts reasoning capabilities in LLMs.

An AI system to help scientists write expert-level empirical software

An AI system utilizing a Large Language Model and Tree Search creates expert-level scientific software by systematically improving a quality metric, allowing it to explore and integrate complex research ideas. The system has achieved state-of-the-art results in various fields, including bioinformatics, epidemiology, and geospatial analysis, outperforming human-developed methods and accelerating scientific progress.

Set Block Decoding Is a Language Model Inference Accelerator

Set Block Decoding (SBD) is a new paradigm that accelerates language model generation by integrating next token prediction and masked token prediction, allowing for parallel sampling of multiple future tokens. SBD achieves a 3-5x reduction in forward passes required for generation without sacrificing accuracy, and can be implemented by fine-tuning existing models without architectural changes or extra training hyperparameters.

An AI system to help scientists write expert-level empirical software

An AI system combining a Large Language Model and Tree Search creates expert-level scientific software by systematically improving a quality metric and navigating a large solution space. The system achieves state-of-the-art results in various fields, including bioinformatics, epidemiology, and geospatial analysis, outperforming human-developed methods and accelerating scientific progress.

Maestro: Joint Graph and Config Optimization for Reliable AI Agents

Maestro is a holistic optimizer for large language model (LLM) agents that jointly searches over graph structures and node configurations to improve agent quality, outperforming existing prompt optimizers on several benchmarks. By leveraging reflective textual feedback and jointly optimizing graphs and configurations, Maestro achieves significant gains in performance while requiring fewer rollouts, and addresses structural failure modes that cannot be fixed by prompt tuning alone.

Code

gcloud MCP Server: Seamless Integration with GCP Resources for AI agents

The gcloud Model Context Protocol (MCP) server allows AI assistants to interact with the Google Cloud environment using natural language, enabling users to automate complex workflows and simplify cloud management. The server can be set up with various AI clients, including Gemini CLI, Claude Desktop, and Visual Studio Code, and provides a range of tools and permissions management to restrict access to certain gcloud commands.

Show HN: OSS app to find LLMs across multiple LLM providers (Azure, AWS, etc.)

Any-llm is a unified API that allows users to access different large language model (LLM) providers through a single interface, simplifying the process of switching between models and providers. The project offers a range of features, including a simple and unified interface, support for multiple providers, and active maintenance, making it a convenient solution for developers working with LLMs.

Windows-Use: an AI agent that interacts with Windows at GUI layer

Windows-Use is a powerful automation agent that interacts directly with the Windows GUI layer, allowing AI agents to perform tasks such as opening apps and typing without relying on traditional computer vision models. The agent can be installed using Python 3.12 or higher and UV or pip, and can be used to automate various tasks on Windows 7-11.

GEPA: System Optimization Through Reflective Text Evolution

GEPA (Genetic-Pareto) is a framework for optimizing text components in systems, such as AI prompts or code, using reflective text evolution and large language models. The GEPA algorithm iteratively mutates and reflects on system behavior to drive targeted improvements, and can be used to optimize systems with minimal evaluations, achieving significant performance gains as demonstrated in examples such as improving GPT-4.1 Mini's performance on the AIME benchmark by 10%.

Show HN: Project Chimera – Hybrid AI Agent Combining LLM, Symbolic, and Causal

Project Chimera is an advanced AI agent that combines a Large Language Model (LLM) with a symbolic safety net and a causal oracle to make strategic decisions that are intelligent, safe, explainable, and provably profitable. The agent has been shown to outperform a pure LLM in a simulated e-commerce business scenario, achieving significant profit and high brand trust while navigating complex trade-offs.

    A hacker used a "memory mailbox" technique to add a live LLM to Animal Crossing, Anthropic's $1.5B AI copyright settlement was rejected by a federal judge, and researchers introduced R-Zero, a self-evolving LLM framework that generates its own training data.