AI news

We replaced RAG with a virtual filesystem for our AI documentation assistant (mintlify.com) AI

Mintlify says it replaced RAG-based retrieval in its AI documentation assistant with a “virtual filesystem” that maps docs pages and sections to an in-memory directory tree and files. The assistant’s shell-like commands (e.g., ls, cd, cat, grep) are intercepted and translated into queries against the existing Chroma index, with page reassembly from chunks, caching, and RBAC-based pruning of inaccessible paths. By avoiding per-session sandbox startup and reusing the already-running Chroma database, the team reports cutting session boot time from about 46 seconds to ~100 milliseconds and reducing marginal compute cost.

Understanding young news audiences at a time of rapid change (reutersinstitute.politics.ox.ac.uk) AI

The Reuters Institute report synthesizes more than a decade of research on how 18–24-year-olds access and think about news amid major media and technology change. It finds young audiences have shifted from news websites to social and video platforms, pay more attention to individual creators than news brands, and consume news less frequently and with less interest—often saying it is irrelevant or hard to understand. The study also highlights greater openness to AI for news, alongside continued concerns about fairness and perceived impartiality, and it concludes publishers need to rethink both distribution and news relevance for younger people.

Cursor 3 (cursor.com) AI

Cursor has released Cursor 3, a redesigned, agent-first workspace intended to make it easier to manage work across multiple repositories and both local and cloud agents. The update adds a unified agents sidebar (including agents started from tools like GitHub and Slack), faster switching between local and cloud sessions, and improved PR workflows with a new diffs view. It also brings deeper code navigation (via full LSPs), an integrated browser, and support for installing plugins from the Cursor Marketplace.

Google releases Gemma 4 open models (deepmind.google) AI

Google DeepMind has released Gemma 4, a set of open models intended for building AI applications. The page highlights capabilities such as agentic workflows, multimodal (audio/vision) reasoning, multilingual support, and options for fine-tuning. It also describes efficiency-focused variants for edge devices and local use, along with safety and security measures and links to download the model weights via multiple platforms.

Show HN: TurboQuant for vector search – 2-4 bit compression (github.com) AI

Show HN spotlights py-turboquant (turbovec), an unofficial implementation of Google’s TurboQuant vector-search method that compresses high-dimensional embeddings to 2–4 bits per coordinate using a data-oblivious random rotation and math-derived Lloyd-Max quantization. The project is implemented in Rust with Python bindings via PyO3 and emphasizes zero training and fast indexing. Benchmarks on Apple Silicon and x86 compare favorably to FAISS (especially at 4-bit) in speed while achieving comparable or better recall, with much smaller index sizes than FP32.

ESP32-S31: Dual-Core RISC-V SoC with Wi-Fi 6, Bluetooth 5.4, and Advanced HMI (espressif.com) AI

Espressif announced the upcoming ESP32-S31, a dual-core 32-bit RISC-V SoC combining Wi‑Fi 6, Bluetooth 5.4 (including LE Audio and mesh), and IEEE 802.15.4 for Thread/Zigbee, plus a 1Gbps Ethernet MAC. The chip targets next-generation IoT devices with a 320MHz core, multimedia-oriented HMI features (camera/display/touch and graphics acceleration), security hardware (secure boot, encryption, side-channel and glitch protections, and TEE), and support for ESP-IDF and Matter-related frameworks.

Show HN: Apfel – The free AI already on your Mac (apfel.franzai.com) AI

Show HN project Apfel presents a free, on-device AI for macOS Apple Silicon that exposes Apple’s built-in language model as a terminal CLI, an OpenAI-compatible local HTTP server, and an interactive chat. The tool is designed to run inference locally with no API keys or network calls, and it supports features like streaming and JSON output for use with existing OpenAI client libraries. The post also highlights related companion tools in the “apfel family,” such as a GUI and clipboard-based actions.

A Recipe for Steganogravy (theo.lol) AI

The article describes a Python CLI concept for “steganogravy,” using neural linguistic steganography to hide a small payload in the introduction text of AI-generated recipe blog posts. It explains the basic arithmetic-coding approach, the need for encoder/decoder to match model settings and prompts, and practical limitations like inefficiency and tokenization divergence. The author also notes a filtering method to prevent decoding failures and illustrates recovery of a hidden message from the generated text.

April 2026 TLDR Setup for Ollama and Gemma 4 26B on a Mac mini (gist.github.com) AI

The gist provides a step-by-step guide for running Ollama on an Apple Silicon Mac mini, pulling the Gemma 4 12B model, and configuring it to start automatically with the model preloaded and kept alive. It includes commands to verify GPU/CPU usage, create a launch agent to periodically “warm” the model, and set OLLAMA_KEEP_ALIVE to prevent unloading due to inactivity. It also notes relevant Ollama updates such as the MLX backend and summarizes key memory considerations for a 24GB system.

Salomi, a research repo on extreme low-bit transformer quantization (github.com) AI

Salomi is a GitHub research repo exploring extreme low-bit (near-binary) transformer quantization and inference for GPT-2–class models, with code, experiments, and evaluation tooling. It specifically tests whether strict 1.00 bpp post-hoc binary quantization can match or beat higher quantization baselines and concludes it does not hold up under rigorous evaluation. The repo instead reports more credible results around ~1.2–1.35 bpp using methods such as Hessian-guided vector quantization, mixed precision, and magnitude-recovery, and directs readers to curated assessment and validation documents over older drafts.

Show HN: Mkdnsite – Markdown-native web server for humans (HTML) and agents (md) (github.com) AI

Mkdnsite is an open-source “Markdown-native” web server that serves a directory or GitHub repo of .md files without a static-site build step. It renders HTML for browsers and uses HTTP content negotiation to return raw Markdown for AI agents (e.g., via Accept: text/markdown), along with an auto-generated /llms.txt and an optional MCP endpoint. The project supports Bun/Node/Deno, runtime editing without redeploy, and includes features like search, theming, math (KaTeX), Mermaid, and syntax highlighting.

Show HN: Semantic atlas of 188 constitutions in 3D (30k articles, embeddings) (constitutionalmap.ai) AI

Constitutional Map AI is a web tool that builds a 3D semantic atlas of constitutional law by embedding thousands of constitutional articles from 188 constitutions. It clusters the text into thematic “neighborhoods” and lets users compare countries on a shared semantic space using keyword or semantic search, with metrics like coverage and entropy. The site’s data is sourced from the Constitute Project and the code is open source, with a note that AI clustering or segmentation errors are possible.

The Anti-Intellectualism of the Silicon Valley Elite (thenation.com) AI

The article argues that Silicon Valley’s top figures—citing figures like Peter Thiel and Marc Andreessen—promote an anti-intellectual worldview that treats deep learning as unnecessary, even while profiting from it. It links this stance to attacks on higher education and the humanities, skepticism toward inquiry that could challenge the managerial class, and a broader desire for insulation from accountability. The piece also criticizes how AI and tech “shortcuts” can be used to replace thinking, while the same elite dismisses the people and disciplines that make that knowledge possible.

AbodeLLM – An offline AI assistant for Android devices, based on open models (github.com) AI

AbodeLLM is an Android app that runs an offline AI assistant using open-source models such as LLaMA and DeepSeek, with chat processed entirely on-device and no internet required. It supports optional multimodal inputs (vision and audio depending on models), context retention, and an “Expert Mode” for tuning generation and cache/token limits. The project includes installation steps and a list of supported model variants along with minimum hardware requirements.

The Claude Code Leak (build.ms) AI

An article argues that the alleged leak of Claude Code’s source code matters less than the broader lessons it highlights: product-market fit and seamless model-to-agent integration outweigh the quality or even the cleanliness of the underlying code. The writer also discusses how the code appears to be “bad” yet still supports a valuable product, why observability and automation may be more important than implementation details, and how the ensuing DMCA and clean-room rewrites reflect ongoing copyright tensions in AI development.

Trinity Large Thinking (openrouter.ai) AI

OpenRouter lists Arcee AI’s open-source “Trinity Large Thinking” model and its pricing on the platform, including per-token input/output costs and usage statistics. The page explains how OpenRouter routes requests to multiple providers with fallbacks to improve uptime, and how to enable reasoning output via a request parameter and the returned reasoning_details.

Perplexity Says MCP Sucks (suthakamal.substack.com) AI

The author argues that Perplexity’s critique of MCP’s token overhead is directionally right but misses the bigger issue: MCP doesn’t provide trust-aware controls for where sensitive data goes after authorization, so different kinds of regulated data are treated identically. They propose adding sensitivity metadata to tool responses, a shared trust-tier registry for inference providers, and runtime enforcement (including redaction/blocking or attestation) to prevent unsafe routing. The piece also notes similar trust gaps in WebMCP and frames MCP’s performance debate as secondary to missing data-governance primitives.

Show HN: 65k AI voters predict UK local elections with 75% accuracy (kronaxis.co.uk) AI

Kronaxis reports a forecast for the 7 May 2026 UK local elections using 65,000 synthetic “voters” built from Census 2021 demographics plus a personality and political-history model. After testing the approach against 10 recent English by-elections and applying a calibration correction for consistent bias, the company claims about 75% winner accuracy on that limited validation set. For the first 20 councils in its release, it predicts Reform UK wins 18 of 20, with Labour narrowly holding Manchester and Greens winning Bristol, while predicting Conservatives take no council seats. The post emphasizes that calibration used the same by-elections as evaluation and will need to be validated by the actual election results.

Ukrainian drone holds position for 6 weeks (defenceleaders.com) AI

A Ukrainian remotely operated, machine-gun armed UGV (TW 12.7) reportedly stayed on station at a contested crossroads for over six weeks, moving forward daily and withdrawing to cover at night. The system answered multiple calls for fire, helping suppress Russian activity and support infantry tasks, highlighting growing maturity and reliability of Ukraine’s domestically produced strike ground robots. The article also stresses the need for operator training, protected recovery methods to avoid risking personnel, and manufacturer testing to improve sensors and turrets under realistic conditions.

The revenge of the data scientist (hamel.dev) AI

The post argues that much of “LLM harnessing” and evaluation is still traditional data science, despite claims that the field is declining or that engineering teams can rely on APIs and generic tooling. It highlights common eval pitfalls—such as using generic metrics, unverified LLM judges, weak experimental design, low-quality data/labels, and over-automation—and explains how data scientists would approach each with trace analysis, error breakdowns, proper validation, and domain-expert labeling.

Obfuscation is not security – AI can deobfuscate any minified JavaScript code (afterpack.dev) AI

The AfterPack blog argues the “Claude Code source leak” didn’t expose hidden code: Claude Code’s CLI JavaScript was already publicly accessible on npm, with only a source map accidentally revealing additional internal comments and file structure. It also contends the bundled code is minified rather than truly obfuscated, and that AI/AST parsing can extract large amounts of prompts, tool descriptions, and configuration strings directly from the minified bundle. Anthropic says the issue was a packaging mistake and not a security breach, noting similar source map exposure occurred before.

Show HN: Git bayesect – Bayesian Git bisection for non-deterministic bugs (github.com) AI

Git bayesect is a Python tool that applies Bayesian inference to automate “git bisect” for flaky or non-deterministic failures, estimating which commit most likely introduced a change in failure likelihood. It uses a greedy entropy-minimization strategy and a Beta-Bernoulli approach to handle unknown failure rates, with commands to record pass/fail observations and select the most probable culprit commit. The README also includes examples and a demo that simulates a test whose failure probability shifts over a repo’s history.

Prompt Engineering for Humans (michaelheap.com) AI

The article argues that “prompt engineering” is essentially the same as good management: providing clear context, constraints, success criteria, and validation so people (and AI) don’t have to guess. Using an example with an agent building a Trello CLI feature, the author shows that vague instructions produced a technically correct but incomplete result, while more specific context led to an immediately usable command. The piece concludes that at scale, ambiguity is costly and managers must design requirements carefully rather than simply assign tasks.

Inside the 'self-driving' lab revolution (nature.com) AI

The article reviews how “self-driving” laboratories are using AI, robotics and automated instrumentation to plan and carry out experiments with minimal human input. It highlights systems such as Ross King’s robotic platform Eve/Adam and GPT-4/LLM-driven approaches that can interpret scientific requests, run multi-step procedures, and even adjust based on experimental “eyes.” While the technology is still early and not a full replacement for human expertise, the piece argues it is already improving speed and lowering some research costs, prompting debate about how biology and chemistry may be done in the future.

Show HN: Claude Code rewritten as a bash script (github.com) AI

The GitHub project “claude-sh” ports Claude Code’s functionality to a ~1,500-line bash script, relying only on curl and jq (optional ripgrep/python3). It supports streamed output, tool use (Bash, Read/Edit/Write/Glob/Grep), permission prompts for non-safe commands, CLAUDE.md project instruction loading, git-aware context, session save/resume, and basic rate-limit retry and cost tracking. The README also documents installation, environment variables, and command-line/slash commands like /help, /cost, /commit, and /diff.