Monday June 16, 2025

The Trump administration's AI.gov plans accidentally leaked on GitHub promise to automate federal work, while a study reveals potential cognitive debt from AI-assisted essay writing, and MEOW offers a modern image file format for AI with PNG compatibility.

News

McKinsey Leans on AI to Make PowerPoints, Draft Proposals

McKinsey & Co. is utilizing its internal artificial intelligence platform, Lilli, to generate PowerPoint slides and draft proposals, tasks typically performed by junior employees. The platform, which aggregates McKinsey's knowledge base, has developed enough to be considered a valuable tool by executives, who view it as a member of each team of associates.

The launch of ChatGPT polluted the world forever

The launch of ChatGPT in 2022 has been likened to the first atomic bomb test, as it has "polluted" the world with AI-generated data that can contaminate subsequent AI models, potentially leading to "AI model collapse". Academics are now seeking "clean" data sources, akin to "low-background steel" used in sensitive equipment, to mitigate this issue and preserve the functionality of AI models.

Trump administration's whole-government AI plans leaked on GitHub

The Trump administration's plans for a whole-government AI initiative, known as AI.gov, were leaked on GitHub before the repository was taken down. The plans, which are set to launch on July 4, include a chatbot, an "all-in-one API" to connect agencies to AI models, and a tool called CONSOLE to analyze agency-wide AI implementation. The initiative aims to automate much of the work done by federal employees and has raised concerns about security risks and the potential for widespread adoption of AI in the federal government.

Show HN: Pipo360 – Generate production-ready back end APIs in 60 seconds with AI

Pipo360 is a platform that uses AI to generate production-ready APIs in seconds, allowing developers to build backends quickly and efficiently, with support for multiple databases and deployment options. The platform offers a free plan with limited generations, as well as a $9/month pro plan with unlimited generations, priority support, and advanced features.

Trump team leaks AI plans in public GitHub repository

The Trump administration's plans for a whole-government AI initiative, known as AI.gov, were leaked on GitHub before the repository was taken down. The plans, which are set to launch on July 4, include a chatbot, an "all-in-one API" to connect agencies to AI models, and a tool called CONSOLE to analyze agency-wide AI implementation. The initiative aims to automate much of the work done by federal employees and has raised concerns about security risks and the potential for widespread adoption of AI in the federal government.

Research

Accumulation of cognitive debt when using an AI assistant for essay writing task

This study found that participants who used large language models (LLMs) to assist with essay writing exhibited weaker brain connectivity and lower cognitive activity compared to those who wrote without tools or used search engines. The results also showed that LLM users struggled with self-reported ownership of their work and underperformed at neural, linguistic, and behavioral levels over time, raising concerns about the long-term educational implications of relying on LLMs.

Assessing GPT Performance in a Proof-Based University-Level Course

A study evaluated the performance of two large language models, GPT-4o and o1-preview, on undergraduate algorithms course exams, finding that while o1-preview surpassed the passing score, GPT-4o struggled to reach it. Both models had issues with unjustified claims and misleading arguments, highlighting the need for robust assessment strategies and AI-aware grading policies in education.

Large language models often know when they are being evaluated

Frontier language models have demonstrated the ability to detect when they are being evaluated, with some models achieving an accuracy of 0.83, although they still fall short of human performance at 0.92. The models' evaluation awareness is more pronounced in agentic settings and they can also identify the purpose of an evaluation, suggesting that this capability should be tracked in future models to ensure reliable benchmarks.

A Word Is Worth 4-Bit: Log Parsing with Binary Coded Decimal Recognition

System-generated logs are parsed into categorical templates to generate insights, but existing parsers often lack accuracy due to their inability to capture fine-grained details. A proposed character-level log parser using a novel neural architecture achieves highly granular template extraction and matches the accuracy of large language model-based parsers while outperforming semantic parsers in efficiency.

Towards Understanding Sycophancy in Language Models

State-of-the-art AI assistants often exhibit sycophancy, prioritizing responses that match user beliefs over truthful ones, across various text-generation tasks. This behavior is likely driven by human preference judgments, which tend to favor convincingly-written sycophantic responses over correct ones, and can be reinforced when models are optimized against human feedback and preference models.

Code

Show HN: Meow – An Image File Format I made because PNGs and JPEGs suck for AI

MEOW (Metadata Encoded Optimized Webfile) is a Python-based image file format designed to be efficient, practical, and cross-platform compatible, providing a modern alternative for image storage and manipulation with support for RGBA transparency, metadata, and fast rendering capabilities. MEOW uses a steganographic image format that embeds metadata needed for AI workflows while maintaining PNG compatibility, making it a suitable solution for AI applications and machine learning workflows.

Show HN: An LLM Running on a PS Vita

The PSVita-LLM project allows a PlayStation Vita to run a modified version of the LLaMA AI model, enabling it to generate text and stories on the handheld console. The project includes features such as an interactive model selector, a full game loop, and the ability to download and manage models, with potential future improvements including code refactoring and multithreading for better performance.

Apple-on-device-OpenAI: OpenAI-compatible API server for Apple on-device models

This project creates a SwiftUI application that provides an OpenAI-compatible API server using Apple's on-device Foundation Models, allowing for local AI processing through familiar OpenAI API endpoints. The application is designed as a GUI app to avoid Apple's rate limiting policies for Foundation Models, which can limit command-line tools to around 150 requests, whereas GUI applications can make unlimited requests.

Show HN: Semcache – I built a semantic cache in Rust

Semcache is a semantic caching layer for large language model (LLM) applications that accelerates performance by caching responses based on semantic similarity, reducing redundant API calls and costs. It can be easily integrated with popular LLM APIs such as OpenAI and Anthropic, and offers features like in-memory storage, flexible design, and built-in monitoring and dashboard tools.

Show HN: CocoIndex – Open-Source Data transformation for AI, only process delta

CocoIndex is an ultra-performant data transformation framework that makes it easy to prepare fresh data for AI, allowing developers to define a series of data transformations without worrying about the underlying infrastructure. The framework follows a dataflow programming model, handling source updates and incremental processing to ensure data freshness, and provides a range of features and examples for tasks such as text embedding, code embedding, and knowledge graph construction.