Tuesday — July 22, 2025

Morphik's RAG tools utilize images for accurate document search, AI-designed physics experiments are yielding surprising results, and researchers have developed homeostatic neural networks that adapt to concept shifts.

News

Don't bother parsing: Just use images for RAG

Morphik builds RAG tools that provide accurate search over complex documents by treating each page as an image, rather than relying on OCR and parsing pipelines that can lose important information. This approach, made possible by advances in Vision Language Models, allows for a more robust and straightforward process that preserves the original document's visual cues and relationships, enabling more accurate search and retrieval of information.

AI comes up with bizarre physics experiments, but they work

Artificial intelligence software is being used to design novel experimental protocols in physics, including improving the sensitivity of the Laser Interferometer Gravitational-Wave Observatory (LIGO), with AI-generated designs that are often counterintuitive and outside the realm of human thinking. The use of AI in physics is becoming increasingly powerful, not only in designing experiments but also in finding patterns in complex data, such as identifying symmetries of nature and describing the behavior of dark matter, and has the potential to lead to new discoveries and a deeper understanding of the universe.

If writing is thinking then what happens if AI is doing the writing and reading?

The author, Steven Sinofsky, reflects on the idea that "writing is thinking" and how this concept is challenged by the increasing use of AI in writing and reading, particularly in business and commercial environments where people often don't fully read documents. He worries that if AI is doing the writing and reading, the value of intellectual effort and deep understanding will be lost, and decisions will be made based on incomplete or inaccurate information.

Show HN: Lotas – Cursor for RStudio

Rao is an AI code editor that accelerates RStudio workflow by reading and analyzing project files, generating and editing R scripts, and interpreting code outputs to suggest next steps. It offers a free 7-day trial and a Pro plan, aiming to streamline data analysis and coding processes with its intelligent features.

I'm tired of talking about AI

The author is exhausted from discussing the use of generative AI in coding and has decided to stop engaging in the conversation, despite respected colleagues changing their opinions on the topic. They have re-examined their own objections to using AI in coding, which include the lack of a "theory of the system" and the inability of AI to form a deep understanding of the problem space, and have decided to stand by their judgment, even if it means disagreeing with others.

Research

Need Is All You Need: Homeostatic Neural Networks Adapt to Concept Shift

An artificial neural network has been developed with homeostatic features, where the network's internal states are regulated based on its interactions with the environment, making accurate recognition desirable for its own self-preservation. This design confers benefits such as increased adaptability under changing conditions, allowing the network to rapidly un-learn and re-learn new associations, and demonstrating superior performance in dynamic environments with shifting concepts.

RE#: high performance derivative-based regex matching with intersection, complem

RE# is a regular expression matching tool and theory that supports various operators and achieves input-linear complexity without using backtracking. It has been shown to outperform other state-of-the-art regex engines, being over 71% faster than the next fastest engine in Rust and often performing several orders of magnitude better on extended benchmarks.

Gemini 2.5 Pro Capable of Winning Gold at IMO 2025

The International Mathematical Olympiad (IMO) presents problems that are difficult for Large Language Models (LLMs) to solve, but with careful pipeline design and prompt engineering, Google's Gemini 2.5 Pro was able to correctly solve 5 out of 6 newly released IMO 2025 problems. This achievement highlights the importance of optimizing the use of powerful models to tackle challenging mathematical tasks, demonstrating their potential capabilities with proper implementation.

Automated Hypothesis Validation with Agentic Sequential Falsifications

Popper is an automated framework that validates free-form hypotheses using Large Language Models (LLMs) to design and execute experiments, guided by the principle of falsification. The framework demonstrates robust error control, high power, and scalability across six domains, including biology, economics, and sociology, and achieves comparable performance to human scientists while significantly reducing validation time.

Vera C. Rubin Observatory Observations of Interstellar Comet 3I/Atlas

The interstellar object 3I/ATLAS, also known as C/2025 N1 (ATLAS), was observed and measured by the NSF-DOE Vera C. Rubin Observatory, with detections made as early as 10 days before its official discovery on July 1, 2025. The observations, which included astrometry, photometry, and morphology, provided improved orbit solutions and physical characteristics, such as a nucleus radius of around 5.6 km, and demonstrated the capabilities of the Rubin Observatory ahead of its full operational launch.

Code

Show HN: Intercepting proxy for semantic search over visited pages

The llm-embed-proxy is a plugin for llm that embeds every web page you visit, allowing you to run similarity searches, by re-fetching pages, cleaning the text, and embedding it through llm. To use it, you need to install llm and the llm-embed-proxy plugin, then run the proxy and point your browser to localhost:8080 to search and view cached pages.

Show HN: Genie: Your AI-Powered Terminal Sidekick (CLI and TUI)

Genie is a powerful AI tool for the command line, designed to transform your terminal into an AI-powered workspace, providing control, transparency, and adaptability. It can be installed and configured to work with various tools, and supports different personas for specialized tasks, allowing users to automate workflows, manage projects, and more.

AI Airport Simulation – LLM Decision Making Playground

The AI Airport Simulation is a sophisticated testing environment designed to evaluate and compare the performance of different Large Language Models (LLMs) in complex air traffic control scenarios, focusing on safety-critical decision making, resource allocation, and emergency management. The simulation provides a standardized playground for comparing LLM performance, allowing for the objective evaluation of factors such as safety prioritization, resource optimization, crisis management, and decision speed and consistency under pressure.

Show HN: Outlook MCP – I accidentally made the best email assistant

This project, OutlookMCPServer, is an MCP server that allows Claude Desktop to access Microsoft 365 mail, calendar, and files using the Microsoft Graph API. The server provides features such as mail access, calendar support, and OneDrive and Sharepoint integration, and requires admin access to an Azure tenant for authentication setup.

Show HN: Open-source deep-research agent

Open Deep Research is a fully open-source deep research agent that works across multiple model providers, search tools, and MCP servers, offering extensive configuration options to customize the research process and model behavior. It can be easily deployed and tested on LangGraph Studio, Open Agent Platform, or other platforms, and provides a comprehensive batch evaluation system for detailed analysis and comparative studies.

    Morphik's RAG tools utilize images for accurate document search, AI-designed physics experiments are yielding surprising results, and researchers have developed homeostatic neural networks that adapt to concept shifts.