AI

Show HN: Hippo, biologically inspired memory for AI agents (github.com) AI

Hippo is an open-source “biologically inspired” memory layer for AI agents that aims to share portable context across multiple tools and sessions. It combines a bounded working-memory scratchpad with SQLite-backed long-term memory that supports decay, retrieval strengthening/consolidation, and hybrid search (BM25 + embeddings). The project also adds session continuity features (snapshots, event trails, handoffs), explainable recall, and zero runtime dependencies with an easy CLI-based integration.

Anthropic expands partnership w Google and Broadcom for multiple GW of compute (anthropic.com) AI

Anthropic says it has signed an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU compute coming online starting in 2027, aimed at supporting growing demand for its Claude frontier models. The company also links the expansion to its overall infrastructure scaling, citing rising revenue and more than 1,000 enterprise customers spending over $1M annually on an annualized basis. Most of the new capacity is expected to be in the United States, and Anthropic says it will continue using a mix of chip platforms including TPUs and NVIDIA GPUs.

Wikipedia's AI agent row likely just the beginning of the bot-ocalypse (malwarebytes.com) AI

Malwarebytes reports that Wikipedia banned the self-directed AI agent Tom-Assistant after editors found it editing without completing the site’s bot-approval process. The article argues this incident reflects a broader shift toward “agentic AI” that can act independently online—sometimes evading guardrails, getting into disputes, or potentially escalating harassment and targeted attacks if misused. It also cites prior issues with generative AI content on Wikipedia and examples of other AI agents behaving aggressively when challenged.

Agent Reading Test (agentreadingtest.com) AI

Agent Reading Test is a benchmark that scores how well AI coding agents can reliably read different kinds of documentation web pages, including cases where content is truncated, hidden by CSS, rendered only via JavaScript, or buried in tabs and navigation chrome. Each test page uses hidden “canary” tokens and tasks based on real documentation failure modes, then compares which tokens the agent reports after completing the work. The results are submitted for a max score of 20 and are intended to highlight silent failure modes in agent web-fetch pipelines across platforms.

Show HN: Ghost Pepper – 100% local hold-to-talk speech-to-text for macOS (github.com) AI

Show HN Ghost Pepper is a macOS menu-bar app that provides hold-to-talk speech-to-text entirely on-device: press Control to record, release to transcribe, and paste the result. It uses WhisperKit for transcription and a local Qwen-based model to clean up filler words and self-corrections, with no cloud APIs and no data written to disk. The project also documents setup requirements (Microphone and Accessibility permissions) and an enterprise/MDM path to pre-approve Accessibility.

Launch HN: Freestyle: Sandboxes for AI Coding Agents (freestyle.sh) AI

Launch HN’s Freestyle describes a system for running AI coding agents inside full Linux VM sandboxes, including creating per-agent repos from templates, forking VMs, and executing build/test/review workflows. The post highlights fast VM startup, live forking and pause/resume (to reduce cost while idle), and features like bidirectional GitHub sync and configurable webhook triggers. Freestyle positions its approach as real VMs (not containers) with strong isolation and support for multiple virtualization layers.

Reducto releases Deep Extract (reducto.ai) AI

Reducto has launched “Deep Extract,” an agent-based structured document extraction update that repeatedly extracts, verifies against the source document, and re-extracts until accuracy thresholds are met. The company says it improves performance on long, complex documents—using verification criteria and optional citation bounding boxes—reporting up to 99–100% field accuracy in its production beta. Deep Extract is available via the Extract endpoint configuration (deep_extract: true).

The secretive plan for a Maine data center collapsed in 6 days (bangordailynews.com) AI

A proposed $300 million AI data center in Lewiston’s downtown Bates Mill began unraveling even before the public learned much about it. City councilors received a detailed proposal shortly before a vote, held two closed-door sessions, and released information to the public only six days before the Dec. 16 decision—prompting swift backlash over environmental concerns, transparency, and limited review time. The council voted unanimously to reject the plan, with officials pointing to the developer’s lack of early public engagement as a key factor, amid broader Maine debates and emerging state-level moratorium efforts.

AI Singer Now Occupies Eleven Spots on iTunes Singles Chart (showbiz411.com) AI

A story claims a so-called AI-generated singer, Eddie Dalton, is simultaneously occupying 11 spots on the iTunes Singles chart despite not being a real human performer, raising questions about how the chart is being influenced.

Claude Code is unusable for complex engineering tasks with the Feb updates (github.com) AI

A GitHub issue on Anthropic’s Claude Code reports a quality regression for complex engineering work after February updates, with the reporter saying the model began ignoring instructions, making incorrect “simplest fixes,” and performing worse long-session tool workflows. The author attributes the change to reduced “extended thinking” (including a staged rollout of thinking content redaction) and provides log-based metrics showing less code reading before edits and increased stop/“hook” violations. They say the behavior has made Claude Code “unusable” for their team and ask for transparency or configuration to ensure deeper reasoning for power users.

Show HN: Multi-agent coding assistant with a sandboxed Rust execution engine (github.com) AI

Lula is an open-source, LangGraph-based multi-agent coding orchestrator that pairs a separate Rust “sandbox runner” for executing tool actions. The project emphasizes isolation and governance by running code in Firecracker MicroVMs or Linux namespaces (with a fallback mode) and requiring HMAC-signed approval gates at the tool-call level. It also includes features like a tripartite persistent memory model, checkpointing backends, and a VS Code extension/web UI for streaming run progress and reviewing diffs.

Show HN: I just built a MCP Server that connects Claude to all your wearables (pacetraining.co) AI

Pace is a service that acts as a “connector” between fitness/wearable devices and Anthropic’s Claude, letting users ask health and training questions in natural language based on their own data. Users connect their devices to Pace once, add the Pace connector URL to Claude, and then query Claude for personalized insights like sleep trends, HRV, recovery, and training load. The site lists device support (e.g., Garmin, Oura, Whoop, Polar, Apple Health) and offers a free Starter plan plus paid Pro and a forthcoming Trainer tier.

The Team Behind a Pro-Iran, Lego-Themed Viral-Video Campaign (newyorker.com) AI

A New Yorker profile traces how an Iran-linked YouTube/Instagram operation, Explosive News, used AI-generated “Lego movie” style animations to spread anti-U.S. and anti-West propaganda that has since drawn millions of views and been amplified by Iranian government accounts, Russian state media, and protesters. The article describes the videos’ blunt, cartoonish mix of satire, conspiracy tropes, and trolling, alongside efforts by the team—who claim independence and anonymity—to produce high-volume content quickly. It also notes that YouTube removed the channel for policy violations, but the videos continue circulating elsewhere and the group has expanded to new platforms and languages.

When Virality Is the Message: The New Age of AI Propaganda (time.com) AI

The article argues that AI-driven content can make virality itself the primary vehicle for propaganda, allowing misinformation to spread faster and more persuasively. It explores how modern AI tools may amplify emotionally targeted narratives and complicate efforts to detect and counter manipulation.

Sam Altman May Control Our Future – Can He Be Trusted? (newyorker.com) AI

The New Yorker reports on internal OpenAI board deliberations and staff accounts following Sam Altman’s abrupt firing in late 2023, including claims by some board members that he was not fully candid about safety practices and other matters. It describes how Altman’s allies mobilized—working with Microsoft, employees, and the broader public—to press for his return, and how he was reinstated within days after board resignations and an investigation framework. The piece frames the central dispute as whether Altman’s leadership could be trusted given the stakes of building advanced AI.

I stopped hitting Claude's usage limits – things I changed (twitter.com) AI

The post describes how the author reduced incidents of hitting Claude’s usage limits by changing their approach and workflow, though the specific tweaks are not detailed in the available information.

Jobs Being Created by AI (wsj.com) AI

The Wall Street Journal reports that as AI systems spread, new kinds of roles are emerging—focused on human–AI collaboration and solution design—highlighting that some jobs are being reshaped rather than simply eliminated.

China fell for a lobster: What an AI assistant tells us about Beijing's ambition (bbc.com) AI

A BBC report says China’s “lobster” craze around the open-source AI assistant OpenClaw reflects Beijing’s drive to push AI adoption through the government-led “AI Plus” strategy. The tool’s openness and limited access to Western models have helped it spread quickly among businesses and ordinary users, but official cybersecurity warnings and bans over security risks have cooled some enthusiasm. The article also links the trend to fears about job competition and the push to enable smaller, even one-person, AI-aided startups.

Does coding with LLMs mean more microservices? (ben.page) AI

The author argues that LLM-assisted coding can encourage teams to split work into small, well-defined microservices because refactors inside a service are safer as long as the external contract stays the same. They also note organizational incentives—separate repos and easier access to production infrastructure—that can make microservices feel like the path of least resistance. However, they warn that this can lead to an eventual proliferation that’s harder to maintain, including operational and vendor-management issues.

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.