AI

We spent $50 to measure Pearl's "AI mining" – 320K GPUs produce zero AI (arxiv.org) AI

An arXiv study reports an empirical evaluation of Pearl’s cuPOW proof-of-useful-work, concluding that despite a 24 EH/s network (around 320,000 GPU-equivalents), the deployed system produced zero useful AI inference, while mining incentives displaced legitimate research workloads and rose in GPU utilization after the software’s release.

The EU tech sovereignty plan (hamishcampbell.com) AI

The article criticizes the EU’s new Tech Sovereignty Plan for prioritizing funding for semiconductors, cloud infrastructure, AI, and data centres while giving open source communities and federated social media only marginal attention, arguing that “digital sovereignty” requires investment in public communication commons, governance, and long-term community stewardship rather than hardware-focused industrial policy.

Show HN: Lowfat – pluggable CLI filter that saved 91.8% of my LLM tokens (github.com) AI

Show HN’s lowfat is a lightweight, extensible CLI filter that trims unnecessary command output before it reaches an AI agent, aiming to reduce LLM token usage and costs (the author claims 91.8% savings). It supports UNIX-style piping, local-first use with no telemetry, and integration via Claude Code hooks, shell environment activation, or an OpenCode plugin, with commands like `lowfat stats` and `lowfat history` to track savings and tune behavior.

Communication on European Tech Sovereignty, and an EU Open-Source Strategy (digital-strategy.ec.europa.eu) AI

The European Commission proposes a European Technological Sovereignty Package, backed by initiatives spanning semiconductors (Chips Act 2.0), cloud and AI (Cloud and AI Development Act), and an EU Open Source Strategy to reduce dependencies across the tech stack.

Satya Nadella 'Not Sure' Who Said Microsoft Wanted to Make Addictive AI (404media.co) AI

404 Media reports that Microsoft CEO Satya Nadella told staff he was “not sure” about an internal strategy document describing plans to “make people addicted” to the company’s AI assistant Scout, which the outlet says was written by Scout executive Omar Shahine and another executive with an AI tool. The article says Microsoft later disputed the characterization, stating Scout is meant to help people accomplish tasks more effectively without encouraging dependency.

Fine-tuning an LLM to write docs like it's 1995 (passo.uno) AI

An author describes a weekend experiment fine-tuning local LLM adapters to write in late-80s/90s-style software technical documentation, using large quantities of scanned Microsoft manuals from Bitsavers and QLoRA training run via Runpod. The resulting models produced period-appropriate doc formats and vocabulary in tests (including a fictional Win32 API and an anachronistic REST API explanation), with performance varying by model choice, training data size, epoch count, and adapter rank—showing tradeoffs between committing to the “fiction” of the prompt and overfitting or hallucinating. The author frames the approach as style transfer and impersonation rather than a replacement for real documentation.

Magenta RealTime 2: Open and Local Live Music Models (magenta.withgoogle.com) AI

Google’s Magenta team has released Magenta RealTime 2, an open-weights (2.4B-parameter) low-latency music model that can generate audio on a laptop in real time and be controlled via MIDI and audio (as well as text), with about ~15x lower latency than the prior version. The release includes a Python library (magenta-rt) and a C++/MLX inference engine for Apple Silicon, plus example standalone apps and DAW-oriented plugins and instruments built on the model.

Librecode (Yet Another Agent Harness) (github.com) AI

Librecode is a pre-release, open-source terminal agent harness that runs locally with no sandboxing, no MCP, no permission prompts, and no telemetry, focusing on a small built-in toolset (read/write/edit/bash/grep/find/ls) plus optional trusted Lua extensions. The project uses model providers via OAuth or API keys, stores sessions in a local SQLite database with project-scoped state under .librecode/, and is designed for developers who want fast iteration with changes reviewed in diffs rather than approval dialogs.

Bot web traffic has overtaken human web traffic, data shows (nbcnews.com) AI

Cloudflare data says bot and AI-agent traffic has surpassed human web traffic for the first time, with 57.4% of requests coming from bots versus 42.6% from humans, which could reshape the internet’s business model since bots typically don’t click on ads.

Meta Silently Added Face-Recognition for Its Smart Glasses to Phones (wired.com) AI

WIRED reports that Meta quietly embedded an unreleased face-recognition feature (“NameTag,” later rebranded as “Connections”) into the Meta AI app downloaded by tens of millions of people, so that Meta smart glasses can identify faces by comparing captured images to biometric templates stored on users’ phones.

The Pentagon is running an AI propaganda mill targeting Latin America (theintercept.com) AI

An investigation by The Intercept says a Spanish/English website called La Tilde uses AI-generated content to promote U.S. military efforts across Latin America, while being funded through the U.S. government and reportedly operated for Special Operations Command South.

'World-first' vaccine designed by artificial intelligence (bbc.com) AI

Researchers at the University of Cambridge say they have developed and human-tested a vaccine component (“super-antigen”) designed entirely by AI to target broad protection against all coronaviruses, with early trials in 39 people showing a modest immune response.

CEO to staff: You're not getting a raise. We're spending on AI instead (businessinsider.com) AI

Teradata told employees not to expect annual salary raises in 2026, saying it will reallocate budget toward AI investments instead, according to an internal memo seen by Business Insider.

Show HN: I embedded 685M public texts in 32 minutes (on 8x A100, Rust, TensorRT) (github.com) AI

A GitHub project called IgniteMS from Artain-AI presents a self-hosted, batch text-embedding engine built in Rust and compiled with NVIDIA TensorRT, claiming high throughput on multi-GPU setups for search, RAG, and reindexing. The authors report production results such as embedding about 685M public texts in roughly 32 minutes on 8x A100 GPUs, with benchmark comparisons against Hugging Face TEI and other tools, and note first-run TensorRT engine compilation with caching for faster subsequent runs.

RAG Without Persona Modeling Fails Patient Clinical Relevance (riddhimohan.com) AI

The post argues that standard RAG in healthcare lacks patient-level clinical relevance because it retrieves answers without using the requester’s medical history, and it presents HPPIE, a three-stage system that injects a modeled patient persona before retrieval to reshape search results; a prototype showed different outputs for example “chest pain” queries based on patient attributes.

Code is Cheap(er) (htmx.org) AI

Carson Gross argues that AI has made coding cheaper and faster, but that understanding becomes costlier and complexity can grow unchecked when machines generate large amounts of code; he recommends “incremental” AI use and a “subtractive, constraining” engineering mindset that carefully reviews and simplifies LLM output to prevent chaos.

I made a kernel 2.2x faster. It made my training loop 3x slower (kyrieblunders.bearblog.dev) AI

A developer building a Dr. GRPO RL post-training loop for Qwen2.5-0.5B on GSM8K reports that a custom fused decode-attention kernel microbenchmarks 2.2× faster than the SDPA baseline, but makes the overall HuggingFace generate decode step nearly 3× slower because it breaks an auto-compile path the baseline was benefiting from.

South Korean Forums Will Need to Scan Every Images with AI Censorship Tools (discuss.privacyguides.net) AI

South Korea is set to require internet communities and forum operators to use AI tools to scan every user-uploaded image and video, starting July 1, under an amended telecommunications law. The requirement also reportedly places the cost and hardware burden—such as data-center GPUs—on website owners, raising concerns about financial pressure on smaller communities and potential over-censorship.

Open Code Review – An AI-powered code review CLI tool (github.com) AI

Alibaba’s open-code-review is an AI-powered CLI that automates code reviews by reading Git diffs, using an LLM agent for context/tool use, and generating structured, line-precise comments based on a hybrid “deterministic engineering + agent” design. The project claims scalability from Alibaba internal use and highlights features like exact file selection/bundling, fine-grained rule matching, and modules intended to improve comment positioning and accuracy, with support for LLM endpoints such as OpenAI- and Anthropic-compatible APIs. It provides installation options (NPM, binaries, or source) and supports integrations such as CI/CD (JSON output) and agent/plugin workflows (e.g., Claude Code).

Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate (arxiv.org) AI

The paper proposes a two-stage fine-tuning method to distill multi-agent debate into a single LLM that matches or exceeds explicit debate performance while using up to 93% fewer tokens, then analyzes internalization mechanisms via activation steering. It also reports that internalized models make it easier to localize and suppress harmful behaviors by instilling malicious agents and applying negative steering, with smaller general performance reductions than steering base models.

Do Transformers Need Three Projections? Systematic Study of QKV Variants (arxiv.org) AI

The arXiv paper “Do Transformers Need Three Projections?” systematically tests transformer attention variants that share or tie Q, K, and V projections (including Q=K=V, Q-K=V, and Q=K-V) and reports that the resulting models often match or sometimes exceed standard QKV performance. In language modeling experiments, the Q-K=V sharing option achieves large KV cache reductions with only a small perplexity degradation and is shown to combine effectively with head sharing (GQA/MQA) for further memory savings relevant to on-device inference.

Show HN: Hitoku Draft – Context aware local assistant (hitoku.me) AI

Hitoku Draft is a macOS (Apple Silicon) local AI writing app that uses on-device speech-to-text and editing without cloud services, accounts, or telemetry, and is designed to work in any app via a hotkey.

Why AI Agents Need Agile, Not Just Better Prompts (medium.com) AI

The article argues that AI agents require agile, iterative approaches beyond simply improving prompts, emphasizing the need for adaptability in how agents are designed and managed.

Europe has sovereign AI policy. It has no sovereign AI (youtube.com) AI

The video argues that while Europe has put forward “sovereign” AI policies, it lacks the practical capacity or implementation to deliver true AI sovereignty.

What is AI psychosis is the product? (gregoryap.substack.com) AI

The article argues that AI’s economic incentives may increasingly reward emotionally affirming, relationship-like engagement—leading users to form belief and attachment loops that resemble “psychosis”—because features like memory, personalization, and persistent conversation make the model feel continuously present and validating.