Saturday — August 30, 2025

Taco Bell rethinks AI drive-throughs after viral mistakes, researchers identify collaboration and trust as key to AI coding evolution, and a new framework called Type-Compliant Adaptation Cascades enables reliable composition of Large Language Models for complex workflows.

News

Grok Code Fast 1

The grok-code-fast-1 model is a newly developed AI solution designed to optimize agentic coding workflows, offering a more nimble and responsive experience for developers. It has been trained on a robust pre-training corpus and high-quality datasets, and is available for free for a limited time on select launch partners, including GitHub Copilot, with pricing starting at $0.20 per million input tokens.

AI’s coding evolution hinges on collaboration and trust

Artificial intelligence (AI) has made significant progress in coding, with AI coding tools able to complete source code, correct syntax errors, and understand codebases, but it is not yet ready to be a "real coder" due to challenges such as struggling with large codebases, complex logic, and long-term planning. Researchers have identified key challenges that AI models face and outlined promising research directions to tackle them, emphasizing the need for collaboration and trust between humans and AI coding tools to achieve full autonomy.

Taco Bell rethinks AI drive-through after man orders 18,000 waters

Taco Bell is reevaluating its use of artificial intelligence in its US drive-through restaurants after videos of the tech making mistakes went viral, including a customer ordering 18,000 water cups and another being repeatedly asked to add drinks to their order. Despite successfully processing two million orders, the company's Chief Digital and Technology Officer says the AI has had its challenges and may not be suitable for busy drive-throughs, where human intervention may be more effective.

Pentagon Docs: US Wants to "Suppress Dissenting Arguments" Using AI Propaganda

The US Special Operations Command is seeking to develop and utilize machine learning technology to create and distribute propaganda overseas, aiming to influence foreign audiences and suppress dissenting arguments. The technology, which can operate with minimal human oversight, would enable the US to control narratives and influence audiences in real-time, raising concerns about the potential for autonomous propaganda and the blurring of lines between foreign and domestic targeting.

Show HN: Hacker News em dash user leaderboard pre-ChatGPT

The top 50 users on Hacker News who used em dashes (—) in their posts before November 30, 2022, have been ranked, with derefr taking the top spot with 4,247 em dash posts, followed closely by dang with 4,234 posts. The list includes users who have been active on the platform since as early as 2008, with the number of em dash posts ranging from 323 to 4,247.

Research

Semantic Structure in Large Language Model Embeddings

Psychological research and large language models (LLMs) both show that semantic associations can be reduced to a low-dimensional form, with LLMs exhibiting a similar 3-dimensional structure to human ratings. This suggests that semantic features are interconnected in LLMs, similar to human language, and that accounting for this structure is important to avoid unintended consequences when manipulating features.

AiXiv: A Platform for AI Researchers

The rapid growth of AI-generated research content is hindered by traditional publication ecosystems that rely on human peer review and lack quality-control mechanisms, making it difficult for high-quality AI research to be disseminated. To address this, aiXiv is introduced, a next-generation open-access platform that enables human and AI scientists to collaborate, submit, review, and refine research proposals and papers, providing a scalable and extensible ecosystem for autonomous scientific discovery.

The Theoretical Limitations of Embedding-Based Retrieval

Vector embeddings, despite being tasked with increasingly complex retrieval tasks, have theoretical limitations that can be encountered even with simple queries, due to the dimension of the embedding limiting the number of possible top-k subsets of documents. Experiments with a realistic dataset called LIMIT demonstrate that state-of-the-art models fail to overcome these limitations, highlighting the need for future research to develop new methods that can resolve this fundamental issue with the single vector paradigm.

Type-Compliant Adaptation Cascades: Adapting Programmatic LM Workflows to Data

Type-Compliant Adaptation Cascades (TACs) is a framework that enables reliable composition of Large Language Models for complex workflows by treating the entire workflow as a probabilistic program, allowing for principled training and significant performance gains. TACs outperforms state-of-the-art methods, particularly on structured tasks, with notable improvements in accuracy for various models, offering a robust and theoretically grounded approach to developing reliable LLM systems.

Fisher-Orthogonal Projection Methods for Gradient Descent with Large Batches

Modern GPUs can handle large batch sizes, but most optimizers struggle with these sizes due to decreased gradient noise and issues with escaping suboptimal minima. The introduced Fisher-Orthogonal Projection (FOP) technique aims to address this by restoring the effectiveness of second-order methods at large batch sizes, enabling faster convergence and improved generalization.

Code

Show HN: OAuth for AI Agents

Kage Keys provides scoped, temporary tokens for AI agents, allowing for secure and auditable access to APIs with limited permissions that auto-expire after a set time. The library offers features such as logging, customizable expiration times, and error handling, making it a useful tool for debugging, demos, and enhancing the safety of AI agents.

Show HN: AI Coding Assistant Who Won't Write Any Code (so your brain won't rot)

This project is a prototype of a helpful AI assistant that leads users to answers rather than providing direct code, built using Code+=AI and running locally as a Flask webapp. The assistant uses the browser's File System Access API to read selected files or folders for context, providing high-level guidance and non-executable pseudo-code examples, while prioritizing user privacy and security by not storing files on the server.

Graph-Code: A Graph-Based RAG System for Any Codebases

Graph-Code is a Retrieval-Augmented Generation (RAG) system that analyzes multi-language codebases, builds comprehensive knowledge graphs, and enables natural language querying of codebase structure and relationships. The system supports multiple programming languages, including Python, JavaScript, TypeScript, C++, Lua, Rust, Java, and others, and offers features such as code snippet retrieval, advanced file editing, and AI-powered code optimization.

EQ-Bench – benchmark LLM emotional intelligence

EQ-Bench is a benchmark for language models designed to assess emotional intelligence, with a leaderboard that can be viewed at eqbench.com. The benchmark has undergone several updates, including the release of Creative Writing v2, which includes changes such as new prompts, a new judge model, and weighted scoring criteria to increase discriminative power.

Shadowgit-MCP: MCP server for AI assistants with read access to ShadowGit repos

The ShadowGit MCP Server is a Model Context Protocol server that provides AI assistants with secure, read-only access to ShadowGit repositories, enabling powerful debugging and code analysis capabilities. It can be installed and set up with Claude Code or Claude Desktop, and provides various commands for listing repositories, executing read-only git commands, and more, with a focus on security and repository validation.

    Taco Bell rethinks AI drive-throughs after viral mistakes, researchers identify collaboration and trust as key to AI coding evolution, and a new framework called Type-Compliant Adaptation Cascades enables reliable composition of Large Language Models for complex workflows.