AI news

Show HN: Real-time AI (audio/video in, voice out) on an M3 Pro with Gemma E2B (github.com) AI

The GitHub project “parlor” showcases an early, on-device system for real-time multimodal AI conversations, using a browser mic/camera input stream and replying with streamed audio. It runs locally via a FastAPI WebSocket server that performs speech and vision understanding with Gemma 4 E2B (LiteRT-LM) and text-to-speech with Kokoro. The demo targets Apple Silicon (e.g., M3 Pro) or Linux with a supported GPU and emphasizes hands-free features like voice activity detection and barge-in (interrupting mid-response).

AI dolls offer companionship to the elderly (ft.com) AI

The Financial Times piece discusses the use of AI-powered dolls intended to provide companionship for elderly people, framing them as a potential support for those who may feel isolated. The article is not available in full in the provided text, so details on results or adoption are not included here.

Make Humans Analog Again (bhave.sh) AI

The opinion piece argues that AI agents can make people more “analog” by boosting hands-on creation, movement, and communication rather than replacing human work. It describes examples of using agents for coding, diagramming, and implementing ideas, and argues that better engineering practices (refactoring, documentation, testing) help agents work faster. The author also frames software development skills like delegation and orchestration as new forms of management and emphasizes that AI’s capabilities have limits that humans must bridge.

My university uses prompt injection to catch cheaters (varun.ch) AI

A first-year computer science course reportedly embeds hidden prompt text in nearly invisible font to detect students who paste cheating instructions into LLMs, and the author suggests a floormate discovered the technique after testing it.

LLMs can't justify their answers–this CLI forces them to (wheat.grainulation.com) AI

The article describes “wheat,” a CLI/framework that helps teams using Claude Code turn technical questions into structured decision briefs. It gathers evidence through research, prototype, and adversarial challenge steps, records findings as typed claims with evidence grades, and uses a multi-pass compiler to catch contradictions and block output until issues are resolved. The output is a shareable, self-contained recommendation with an audit trail, illustrated with an example GraphQL migration decision.

New Copilot for Windows 11 includes a full Microsoft Edge package, uses more RAM (windowslatest.com) AI

A new Copilot update for Windows 11 replaces the native app with a web-based “hybrid” version that ships with its own bundled Microsoft Edge/Chromium components. The installer is distributed via the Microsoft Store, but it downloads an installer rather than the full app directly. In tests, the updated Copilot uses significantly more memory—up to around 500MB in the background and about 1GB during use.

AI agents promise to 'run the business,' but who is liable if things go wrong? (theregister.com) AI

The Register examines how liability remains unclear when AI agents “run the business” and errors cascade through automated decisions like HR, finance, and supply chain processes. UK regulators stress that accountable responsibility still sits with the using firm and its responsible individuals, even if the technology is provided by a vendor. Lawyers and analysts say contracts may shift blame through warranties, testing, monitoring, and explainability—yet non-deterministic agent behavior makes it hard to promise (or assign) predictable outcomes, with negotiations focusing on safeguards and the limits of what vendors will accept.

Copilot is 'for entertainment purposes only', per Microsoft's terms of use (techcrunch.com) AI

Microsoft’s terms of use for Copilot say it’s intended for entertainment only and that users shouldn’t rely on its outputs for important advice, as it can make mistakes. The company said it plans to update older wording, which had been criticized online. The article notes that similar disclaimers are used by other AI providers such as OpenAI and xAI.

Show HN: Gemma Gem – AI model embedded in a browser – no API keys, no cloud (github.com) AI

Gemma Gem is a Chrome extension that runs Google’s Gemma 4 model entirely on-device in the browser using WebGPU. It avoids API keys or cloud calls and can use a simple agent loop to read page content, click and fill forms, run page JavaScript, and answer questions about the site you’re viewing.

Iran's IRGC Publishes Satellite Imagery of OpenAI's $30B Stargate Datacenter (newclawtimes.com) AI

Iran’s IRGC released satellite imagery and a video targeting OpenAI’s planned $30B Stargate AI datacenter in Abu Dhabi, threatening “complete and utter annihilation.” The article frames this as an escalation from earlier, broader IRGC warnings toward specific identification of the facility, citing prior regional attacks affecting Oracle and AWS-related infrastructure. It argues the main risk for AI “agent builders” is disruption to the compute layer behind OpenAI APIs, increasing the importance of multi-provider resiliency.

Show HN: Modo – I built an open-source alternative to Kiro, Cursor, and Windsurf (github.com) AI

Modo is an open-source, MIT-licensed desktop AI IDE that aims to turn prompts into structured development plans before generating code. Built on top of a Void/VS Code fork, it adds spec-driven workflows (requirements/design/tasks persisted on disk), task run UI, project “steering” files for consistent context, configurable agent hooks, and an Autopilot vs Supervised mode. The project also supports multiple chat sessions, subagents, installable “powers” for common stacks, and a companion UI, with setup instructions and a full repository structure provided on GitHub.

Apex Protocol – An open MCP-based standard for AI agent trading (apexstandard.org) AI

Apex Protocol (APEX) proposes an open, MCP-based standard that lets AI trading agents connect directly to brokers/execution venues using a shared set of tools, real-time state, and deterministic safety controls. It specifies canonical instrument IDs (to avoid per-broker symbol mapping), event-driven notifications over HTTP/SSE, session replay for reconnection, and a conformance-tested protocol surface for multiple languages. The standard is CC-BY 4.0 with reference implementations and governance via a technical advisory committee and an open RFC process.

Show HN: I built a tiny LLM to demystify how language models work (github.com) AI

The Show HN post and GitHub repository introduce “GuppyLM,” a simple ~9M-parameter language model trained from scratch on synthetic fish-themed conversations. It walks through the full pipeline—dataset generation, tokenizer training, a vanilla transformer architecture, a basic training loop, and inference—aiming to make LLM internals less of a black box. The project highlights design tradeoffs (single-turn chats, no system prompt, limited context) and provides notebooks and code for reproducing training and running the model.

Show HN: Mdarena – Benchmark your Claude.md against your own PRs (github.com) AI

mdarena is an open-source tool that benchmarks Claude.md instructions by mining real merged PRs from your codebase, running the generated patches against the repo’s actual test suites, and comparing the results to the gold diffs. It reports test pass/fail, patch overlap, and token/cost-related metrics, using history-isolated checkouts to avoid information leakage. The project also includes a SWE-bench-compatible workflow and notes mixed results when consolidating guidance versus using per-directory instructions.

Recall – local multimodal semantic search for your files (github.com) AI

Recall is an open-source tool that enables local multimodal semantic search over your files by embedding images, audio, video, PDFs, and text into a locally stored vector database (ChromaDB). It matches natural-language queries across file types without requiring tagging or renaming, and includes an animated setup wizard plus a Raycast extension for quick visual results. Embeddings are generated using Google’s Gemini Embedding 2 API, while the vector index and files remain on your machine.

'Cognitive Surrender' Is a New and Useful Term for How AI Melts Brains (gizmodo.com) AI

The article highlights a new term, “cognitive surrender,” used to describe how people may increasingly defer their thinking to AI chatbots—even when the AI is wrong. It summarizes a Wharton study where participants used an AI during a math-style reasoning test and were more likely to accept incorrect answers, with higher reported confidence when using the chatbot. The author notes the work may fit into broader concerns about reduced critical thinking and also flags that psychology findings can be hard to replicate.

Spath and Splan (sumato.ai) AI

The post argues that AI coding agents should interact with code using semantic “narratives” rather than filesystem rituals. It introduces Spath (a symbol-addressing format) and Splan (a minimal grammar for batched code-change intentions), claiming they reduce filesystem operations and improve agent efficiency and reliability via transactional edits. Sumato AI says it is open-sourcing the Spath and Splan grammars and provides an example Spath dialect for Go.

OpenAI's fall from grace as investors race to Anthropic (latimes.com) AI

The article says OpenAI’s shares are becoming hard to sell on secondary markets as institutional investors shift toward Anthropic, which is seeing record demand and higher bids. It attributes the pivot to perceived risk-reward, including Anthropic’s focus on profitable enterprise customers versus OpenAI’s heavier infrastructure spending. The piece also notes OpenAI’s recent, large fundraising round and highlights regulatory and security setbacks affecting Anthropic, even as investors remain eager to buy its equity.

Show HN: TermHub – Open-source terminal control gateway built for AI Agents (github.com) AI

TermHub is an open-source “AI-native” CLI/SDK that provides a native control gateway for iTerm2 and Windows Terminal, letting LLMs or AI agents open tabs/windows, target sessions, send text/keystrokes, and capture terminal output programmatically. The project includes a machine-readable spec/handles for AI handoff, plus a send-to-capture “delta” checkpoint mode so agents can retrieve only the new output produced after a command. It’s distributed via npm/Homebrew (macOS) and GitHub releases (binaries), with an SDK preview for JS/TypeScript.

Wavelets on Graphs via Spectral Graph Theory (arxiv.org) AI

The paper presents a way to build wavelet transforms for functions on the vertices of a finite weighted graph using the graph Laplacian’s spectral decomposition. It defines scaled wavelet operators via a kernel g(tL) and forms graph wavelets by localizing these operators, with an admissibility condition ensuring the transform is invertible. The authors also study localization behavior at fine scales and provide an efficient Chebyshev-polynomial method to compute the transform without diagonalizing the Laplacian.

In Japan, the robot isn't coming for your job; it's filling the one nobody wants (techcrunch.com) AI

Japan is accelerating “physical AI” not to replace jobs broadly, but to keep factories, warehouses, and other critical operations running as labor shortages worsen. Backed by government targets and investment, companies are moving from pilots to customer-funded deployments using more autonomous robotics software, orchestration, and integration across existing hardware. Industry sources say Japan’s strength in high-precision robotics components and control systems is a key advantage, with a hybrid ecosystem where incumbents scale while startups build perception and workflow capabilities.

Iran threatens 'complete and utter annihilation' of OpenAI's $30B Stargate (tomshardware.com) AI

Iran’s Islamic Revolutionary Guard Corps has issued a video warning that any attacks on Iranian power infrastructure would be met with “complete and utter annihilation,” naming U.S. and Israeli facilities in the region. The threat specifically targets OpenAI’s reported $30B Stargate AI data center in Abu Dhabi, showing satellite imagery of a 1GW site. The warning follows recent reports of rocket strikes disrupting some AWS data centers and comes amid broader threats from Iran toward major U.S. tech companies.

Introducing GEN-1 [video] (youtube.com) AI

The video titled “Introducing GEN-1” presents an overview of the GEN-1 offering, explaining what it is and how it works.

Policy on adding AI generated content to my software projects (joeyh.name) AI

The author describes a tongue-in-cheek policy for accepting AI-generated code into their projects: bypassing normal code review if the submission compiles, is clearly labeled as “(AI generated),” and includes a signed Developer Certificate of Origin. They note they may still make small changes for QA purposes and will keep the contributor credited as the author, but warn that unlabeled AI code may crowd out human code reviews. The post is framed humorously with examples, including gating a change to leap days.

3 New world class MAI models, available in Foundry (microsoft.ai) AI

Microsoft announced three new MAI models—MAI-Transcribe-1 for speech-to-text, MAI-Voice-1 for voice generation (including custom voice creation), and MAI-Image-2 for image generation—now available in Microsoft Foundry and MAI Playground. The company says MAI-Transcribe-1 targets fast, accurate transcription for the most-used languages, MAI-Voice-1 can generate 60 seconds of audio per second of compute and preserve speaker identity, and MAI-Image-2 delivers faster image generation with similar quality. Microsoft also lists starting prices for each model and notes enterprise controls and red-teaming for safer deployment.