Thursday — July 17, 2025

Researchers propose a "day-dreaming loop" to enhance LLM capabilities, a team at ZeroEntropy develops a reranker model using chess Elo scores, and a new MCP server gives LLMs temporal awareness and time calculation abilities.

News

LLM Daydreaming

Large language models (LLMs) have impressive capabilities but have yet to produce a genuine breakthrough, possibly due to their lack of fundamental aspects of human thought, such as the ability to learn from experience and engage in background processing, or "daydreaming". To address this, a proposed solution is the implementation of a "day-dreaming loop" (DDL), a background process that continuously samples pairs of concepts from memory, explores non-obvious links between them, and filters the results for genuinely valuable ideas, potentially leading to innovative breakthroughs.

Show HN: Improving search ranking with chess Elo scores

At ZeroEntropy, a team of mathematicians and competitive programmers developed a new approach to training a reranker model using the concept of chess Elo Scores, which outperforms other models in improving search retrieval accuracy. The team abandoned traditional methods of using human-annotated pairs and instead used pairwise comparisons to create nuanced and exhaustive scores, resulting in a more accurate and context-aware reranker model.

Show HN: A 'Choose Your Own Adventure' written in Emacs Org Mode

The provided text appears to be a table of contents for a sample book, offering links to begin the story, purchase the book, and access various introductory sections. The sections include a preface, dedication, acknowledgments, and copyright information, all of which can be accessed through the given links.

The party trick called LLM

The authors tested 36 chatbots on local government websites in the Netherlands and found that all of them failed, with 100% being inaccessible to people with visual impairments and providing poor quality outputs for basic tasks. The authors argue that the hype around chatbots and AI is based on a "communicational illusion" that tricks people into thinking that machines can think and produce language, when in reality they are just calculating probability and statistics.

Claude Is Back on Windsurf

Windsurf announced that Claude Sonnet 4 is now available with first-party support from AnthropicAI, offering Pro and Teams users 250 requests per month at a limited-time discount of 2x credits per request. The announcement was made on X, a social media platform, and includes an image and a call to action to sign up for the service.

Research

Chain of thought monitorability: A new and fragile opportunity for AI safety

AI systems that process human language can be monitored for potential misbehavior through their chains of thought, offering a promising approach to AI safety, albeit an imperfect one. Further research and investment in this method, known as CoT monitoring, is recommended, with developers also considering how their decisions may impact the effectiveness of this safety measure.

Empirical evidence of LLM's influence on human spoken communication

The release of chatbots like ChatGPT has introduced a new medium that can spread cultural patterns to hundreds of millions of people, raising questions about their impact on human culture. An analysis of over 740,000 hours of human discourse from YouTube and podcasts found a significant increase in the use of words commonly generated by ChatGPT, suggesting that machines can measurably reshape human culture and create a closed cultural feedback loop between humans and machines.

Artificial Finance: How AI Thinks About Money

Large language models (LLMs) exhibit a risk-neutral decision-making pattern, favoring choices based on expected value calculations, but occasionally produce inconsistent responses when evaluating trade-offs between present and future. The LLMs' aggregate responses are most similar to those of participants from Tanzania, suggesting potential cultural and training influences embedded within their outputs, and contributing to the understanding of how LLMs emulate human-like decision behaviors.

Which Economic Tasks Are Performed with AI? Evidence from Claude Conversations

Researchers analyzed over four million conversations to study AI usage patterns across the economy, finding that AI is primarily used in software development and writing tasks, but also extends to about 36% of occupations for at least a quarter of their tasks. The analysis reveals that AI is used for both augmenting human capabilities (57% of usage) and automating tasks (43% of usage), providing insights into AI's evolving role in the economy.

Assessing RAG and HyDE on 1B vs. 4B-Parameter Gemma LLMs for Personal Assistants

This study evaluated two augmentation strategies, Retrieval-Augmented Generation (RAG) and Hypothetical Document Embeddings (HyDE), on compact large language models, finding that RAG reduces latency and eliminates factual hallucinations, while HyDE enhances semantic relevance but increases response time. The results suggest that RAG is a more practical choice for on-device personal assistants using small-scale language models, due to its efficiency and accuracy benefits.

Code

Show HN: An MCP server that gives LLMs temporal awareness and time calculation

The "Passage of Time" Model Context Protocol (MCP) Server is a tool that gives language models temporal awareness and time calculation abilities, allowing them to understand and work with time in a more human-like way. The server provides a range of functions, including current date and time, time differences, and timestamp context, which can be used to enable language models to have more informed and contextually aware conversations about time and schedules.

Metaflow: Build, Manage and Deploy AI/ML Systems

Metaflow is a human-centric framework that helps scientists and engineers build and manage real-life AI and ML systems, streamlining the development lifecycle from prototyping to production deployment. It provides a simple and friendly Pythonic API, supports rapid prototyping, scaling, and deployment, and is used by thousands of AI and ML experiences across various companies, including Netflix, Amazon, and Goldman Sachs.

Show HN: I built a dream interpreter in JavaScript, no AI, no server, just logic

The Starwhale Oracle is an interactive, celestial dream interpreter that transforms drifting thoughts into glowing truths, providing symbolic guidance from the stars through a dream interpretation engine, dream symbol cards, and a personal dream journal. Users can write their dreams, choose the emotional tone, and receive poetic insights, exploring symbolic meaning cards and saving their dreams in a private journal, all without requiring accounts or uploads.

Show HN: Autopilot for Cursor IDE

This MCP server integrates Nautex AI with the Cursor IDE to facilitate effective communication of product and technical requirements to LLM coding agents, streamlining AI-assisted development. The tool-chain uses a step-by-step plan to convey requirements, allowing coding agents to work more efficiently, and includes features such as specification generation, codebase mapping, and task planning to ensure high-quality code output.

Show HN: We made GPT-4.1-mini beat 4.1 at Tic-Tac-Toe using dynamic context

This repository provides examples and guides to help users get started with Opper, and welcomes pull requests. For more in-depth information, users can visit the Opper blog for technical deep dives or the documentation for examples and explanations of key concepts.

    Researchers propose a "day-dreaming loop" to enhance LLM capabilities, a team at ZeroEntropy develops a reranker model using chess Elo scores, and a new MCP server gives LLMs temporal awareness and time calculation abilities.