Show HN: I recreated Windows XP as my portfolio

Mitch Ivin's portfolio is designed to resemble a Windows XP operating system, where users can "log in" and explore his work, including his about page, resume, projects, and contact information, all presented in a nostalgic desktop environment. The site features various icons and applications, such as a taskbar, start menu, and desktop icons, that mimic the look and feel of the classic operating system.

Read more

Google's new AI mode is good, actually

Google's new "AI mode" is a significant improvement in search capabilities, feeling similar to GPT-5 search but with faster results, although it is currently not available in the EU. The AI mode provides helpful and detailed responses to complex queries, but lacks transparency in its search process, not revealing the specific searches it runs to generate its results.

Read more

Taco Bell AI Drive-Thru

Taco Bell's attempt to implement voice AI ordering systems at over 500 drive-through locations was met with chaos and customer complaints, as the AI struggled to handle creative and absurd orders, as well as regional accents and speech patterns. The company is now reassessing its AI deployment, considering human intervention during peak periods, after realizing that artificial intelligence may not be as effective in handling the complexities of human interaction at the drive-thru window.

Read more

GPT-5 Thinking in ChatGPT (a.k.a. Research Goblin) is shockingly good at search

GPT-5, also referred to as "Research Goblin", has significantly improved search capabilities, allowing it to provide comprehensive answers to complex and trivial questions by searching the internet. The author has used Research Goblin to answer various questions, such as identifying a building, investigating the availability of cake pops at Starbucks in the UK, and finding the official legal name of the University of Cambridge, with impressive results and detailed thought processes.

Read more

Just How Bad Would an AI Bubble Be?

A recent study found that experienced software developers completed coding tasks 20% slower when using AI tools, contradicting the expectation that AI would significantly increase productivity. This "capability-reliability gap" suggests that while AI can perform impressive tasks, it often requires human review and correction, making it less useful in real-world settings and potentially explaining why AI has failed to deliver tangible results for businesses.

Read more

Tricking LLM-Based NPCs into Spilling Secrets

Researchers are investigating the security risks of using Large Language Models (LLMs) to generate dialogue for game NPCs, specifically whether adversarial prompts can trick them into revealing hidden background secrets. This study examines the potential vulnerability of LLM-based NPCs to adversarial prompt injection, which could compromise game security and confidentiality.

ArXiv

Refrag: Rethinking RAG Based Decoding

Large Language Models (LLMs) used in retrieval-augmented generation (RAG) applications face significant latency and memory issues due to long-context inputs, but most computations over the context can be eliminated without impacting performance. The proposed REFRAG framework exploits the sparsity structure of RAG context to achieve substantial speedup, demonstrating a 30.85% time-to-first-token acceleration and extending context size by 16, all without loss in perplexity or accuracy.

ArXiv

Slime Mould Electronic Oscillators

The slime mould Physarum polycephalum is used to create an electronic oscillator, where a protoplasmic tube connects two electrodes and its changing resistance due to peristaltic activity produces oscillations. The oscillator has an average period of 73 seconds and amplitude of 0.6 MOhm, and can operate under direct current voltage up to 15V, suggesting its potential use as a living component in biological and hybrid circuits.

ArXiv

The entropy formula for the Ricci flow and its geometric applications

A new monotonic expression for the Ricci flow is introduced, which has various geometric implications, including the absence of nontrivial periodic orbits and control over the injectivity radius in regions where singularities form. The results also contribute to Richard Hamilton's program for proving the Thurston geometrization conjecture for closed three-manifolds, and an outline for an eclectic proof of this conjecture is provided.

ArXiv

Zero-Shot Reinforcement Learning

Modern reinforcement learning (RL) systems can excel in simulated domains, but struggle in real-world scenarios where data is limited or expensive to obtain, leading to a misalignment between training and deployment environments. This thesis proposes methods to address this issue, known as zero-shot reinforcement learning, by navigating constraints such as poor data quality, partial observability, and limited data availability, with the goal of developing RL methods that can be effectively deployed to solve real-world problems.

ArXiv

Show HN: Semanticcache – A high-performance semantic caching library for Go

Semantic Cache is a high-performance caching library for Go that uses vector embeddings to find semantically similar content, making it suitable for large language model applications, search systems, and other use cases where semantic similarity is important. The library supports multiple backends, including in-memory and Redis, and provides features like semantic search, batch operations, and customizable similarity functions.

Repo

Show HN: VeritasGraph,An on-premise Graph RAG with verifiable source attribution

VeritasGraph is an enterprise-grade graph Retrieval-Augmented Generation (RAG) framework that enables secure, on-premise AI with verifiable attribution, overcoming the limitations of traditional vector-search-based RAG systems. It leverages a knowledge graph to perform complex, multi-hop reasoning, providing transparent and auditable reasoning paths with full source attribution for every generated claim.

Repo

Nano Banana vs. GPT-4o: side-by-side AI image comparisons

Nano Banana 是谷歌推出的 AI 图像编辑模型，代表了从"AI绘画工具"到"AI创意伙伴"的范式转移，能够深刻理解已有图像的上下文、光影、物理逻辑和创作意图。文中展示了多个案例，对比了 Gemini 和 GPT-4o 两种AI模型在图像生成方面的差异。

Repo

Simulating Human-to-Human Dialogue Using Azure AI

The Voice Live API Sales Coach is a demo application that showcases AI-powered voice training for sales professionals using Azure Voice Live API and Azure AI services. It allows users to practice real-world sales scenarios with AI-powered virtual customers, receive instant feedback on their performance, and improve their sales skills through immersive voice conversations.

Repo

Modern (2025) AI Text Editors – A Guide

This repository provides a comprehensive guide to over 30 AI-powered development tools, including IDEs, CLI tools, and cloud platforms, to help developers, teams, and tech leaders navigate the evolving landscape of AI-assisted coding. The guide features comparison tables, pricing information, security considerations, and getting started guides, and is available under the CC0 1.0 Universal (Public Domain) License.

Repo