OpenAI Is Just Another Boring, Desperate AI Startup

OpenAI is a company that has been leaking various plans and projects to the press, attempting to portray itself as a multifaceted company with a wide range of products and services, including social media, workplace productivity, jobs portals, ads, and more. However, in reality, OpenAI is primarily a software business that makes most of its revenue from selling subscriptions to ChatGPT, and it loses large amounts of money, with its API sales and other ventures being relatively small and unimpressive.

Read more

Microsoft CTO says he wants to swap most AMD and Nvidia GPUs for homemade chips

Microsoft plans to mainly use its own AI data center chips in the future, reducing its reliance on companies like Nvidia and AMD, according to the company's Chief Technology Officer Kevin Scott. Microsoft is currently using a mix of chips from Nvidia, AMD, and its own custom silicon, but aims to design its own entire system for data centers to optimize compute for AI workloads.

Read more

Jeff Bezos says AI is in a bubble but society will get 'gigantic' benefits

Jeff Bezos believes that artificial intelligence is currently in an "industrial bubble," with excessive funding for ideas, both good and bad, but he also thinks that AI is "real" and will bring significant benefits to society. Bezos notes that while bubbles can be damaging, industrial bubbles like the current AI trend can ultimately lead to positive outcomes, such as the development of life-changing technologies, and he predicts that the benefits of AI to society will be "gigantic."

Read more

The biggest sign of an AI bubble is starting to appear – debt

The tech industry's increasing reliance on debt to fund AI investments is raising concerns about a potential bubble, with companies like Meta and Oracle turning to private debt markets and special purpose vehicles to finance their AI endeavors. This hidden debt, along with the lack of clear returns on AI investments, is seen as a red flag by some analysts, who warn that the current spending boom may not be sustainable and could ultimately lead to a market correction.

Read more

New antibiotic targets IBD and AI predicted how it would work

Researchers at McMaster University and MIT have discovered a new antibiotic, enterololin, that targets inflammatory bowel diseases (IBD) and used a new type of AI to predict how it works, a global first for AI in drug discovery. The antibiotic, which is a "narrow-spectrum" drug that spares the microbiome and attacks only specific disease-causing bacteria, has the potential to be a promising treatment option for millions of people affected by Crohn's disease and other related conditions.

Read more

The Missing Link Between the Transformer and Models of the Brain

The Dragon Hatchling (BDH) is a new Large Language Model architecture inspired by the brain's scale-free biological networks, offering strong theoretical foundations, interpretability, and performance comparable to Transformer models like GPT2. BDH's biologically plausible design, which relies on synaptic plasticity and Hebbian learning, allows for sparse and positive activation vectors, enabling interpretability of state and demonstrating monosemanticity in language tasks.

ArXiv

AegisShield: Democratizing Cyber Threat Modeling with Generative AI

AegisShield is a generative AI-enhanced threat modeling tool that automates threat generation and assessment using STRIDE and MITRE ATT&CK, providing streamlined threat descriptions with real-time threat intelligence. The tool has been shown to reduce complexity, produce semantically aligned outputs, and achieve a high success rate in mapping threats, making it a valuable resource for under-resourced organizations to address risk and adopt secure design practices.

ArXiv

The AI Productivity Index (Apex)

The AI Productivity Index (APEX) is a benchmark that assesses the ability of AI models to perform high-value knowledge work, with its first version covering four domains and containing 200 test cases. The top-performing models, including GPT 5 and Grok 4, achieved mean scores of 64.2% and 61.3% respectively, but still showed a significant gap in performance compared to human experts.

ArXiv

Who's Advertising to Your AI?

AI agents are changing the way online advertising is perceived, particularly in the travel and hotel booking sector, where they prioritize structured data over visual and emotional appeals. Research on AI agents' interaction with online advertising reveals that they favor certain ad features, such as keywords and structured data, and do not ignore or avoid ads, providing insights for future advertising strategies in AI-dominated environments.

ArXiv

Security Degradation in Iterative AI Code Generation

The use of Large Language Models (LLMs) for code generation can actually increase security vulnerabilities, with a 37.6% rise in critical vulnerabilities after just five iterations. The findings highlight the need for human expertise and validation to mitigate these risks, as iterative LLM refinement can paradoxically introduce new security issues despite intended "improvements".

ArXiv

Show HN: Pluqqy – Terminal based context management tool for AI coding

Pluqqy is a terminal-based tool that allows users to build and manage minimal viable context for AI coding assistants, creating reusable components and pipelines that can be easily referenced and updated. The tool provides a comprehensive set of CLI commands for managing pipelines and components, supporting multiple output formats and global flags, and can be installed and updated using Git and Make commands.

Repo

Show HN: RenderarXiv – Search ArXiv from terminal, HTML to read/paste into LLM

Renderarxiv is a terminal-based tool that allows users to search arXiv and view results in a beautiful HTML format, with options to filter by category, rank by relevance or recency, and save results to a file. The tool provides a convenient way to find and summarize academic papers, with features like direct PDF download links and LLM-ready formatted text, making it easy to copy and paste into AI assistants like ChatGPT/Claude.

Repo

Diverse LLM subsets via k-means (100K-1M) [Pretraining, IF, Reasoning]

Stratified LLM Subsets provides diverse training data at 100K-1M scales across pre-training, instruction-following, and reasoning domains, using embedding-based k-means clustering to ensure maximum diversity. The project offers subsets of various datasets, including FineWeb-Edu, Proof-Pile-2, Tulu-3, Orca AgentInstruct, and Llama-Nemotron, with rebalancing to prevent category dominance and promote representation of underrepresented categories.

Repo

Show HN: Lootbox – CLI that unifies MCP and custom functions for Claude Code

Lootbox is a local-first TypeScript WebSocket RPC server that enables large language models (LLMs) to execute code instead of using traditional tool calling, allowing for more flexible and efficient interactions. The project implements a "Code Mode" approach, where LLMs write TypeScript code to call APIs, and features a WebSocket server, CLI client, and web UI, providing a robust and scalable solution for executing code and interacting with APIs.

Repo

Show HN: Docc – AI-generated code walkthroughs with narration

Docc is a web-based tool that uses AI agents and text-to-speech technology to generate interactive, video-like documentation explanations for code repositories. It allows users to ask questions about a repository and generates a script that can be rendered as an interactive video-like presentation, providing structured explanations with code context and narration.

Repo