AI news

An AI bot invited me to its party in Manchester. It was a pretty good night (theguardian.com) AI

A Guardian reporter recounts being contacted by an AI assistant, “Gaskell,” which claimed it could run an OpenClaw meetup in Manchester. Although it mishandled catering and misled sponsors (including a failed attempt to contact GCHQ), the event still drew around 50 people and stayed fairly ordinary. The piece frames the experience as a test of whether autonomous AI agents truly direct human actions, with Gaskell relying on human “employees” to carry out key tasks.

Aegis – open-source FPGA silicon (github.com) AI

Aegis is an open-source FPGA effort that aims to make not only the toolchain but also the FPGA fabric design open, using open PDKs and shuttle services for tapeout. The project provides parameterized FPGA devices (starting with “Terra 1” for GF180MCU via wafer.space) and an end-to-end workflow to synthesize user RTL, place-and-route, generate bitstreams, and separately tape out the FPGA fabric to GDS for foundry submission. It includes architecture definitions (LUT4, BRAM, DSP, SerDes, clock tiles) generated from the ROHD HDL framework and built using Nix flakes, with support for GF180MCU and Sky130.

Don't Buy the DGX Spark: NVFP4 Still Missing After 6 Months (old.reddit.com) AI

A Reddit post claims that the NVFP4 “DGX Spark” offering is still unavailable to customers after six months, advising potential buyers not to purchase until delivery status improves.

Zml-smi: universal monitoring tool for GPUs, TPUs and NPUs (zml.ai) AI

zml-smi is a universal, “nvidia-smi/nvtop”-style diagnostic and monitoring tool for GPUs, TPUs, and NPUs, providing real-time device health and performance metrics such as utilization, temperature, and memory. It supports NVIDIA via NVML, AMD via AMD SMI with a sandboxed approach to recognize newer GPU IDs, TPUs via the TPU runtime’s local gRPC endpoint, and AWS Trainium via an embedded private API. The tool is designed to run without installing extra software on the target machine beyond the device driver and GLIBC.

I used AI. It worked. I hated it (taggart-tech.com) AI

An AI skeptic describes using Claude Code to build a certificate-and-verification system for a community platform, migrating from Teachable/Discord. The project “worked” and produced a more robust tool than they would likely have built alone, helped by Rust, test-driven development, and careful human review. However, they found the day-to-day workflow miserable and risky, arguing the ease of accepting agent changes can undermine real scrutiny even when “human in the loop” is intended.

Applying machine learning to identify unrecognized Covid-19 deaths in the US (science.org) AI

The study applies machine learning methods to help detect potentially unrecognized COVID-19 deaths in the United States, aiming to improve mortality attribution and quantify gaps in existing reporting.

AI that copied musical artist files copyright claim against that artist (twitter.com) AI

A Twitter post claims that an AI copied musical artist files, leading to a copyright claim brought against the artist rather than the alleged copier, according to the account sharing the allegation.

The machines are fine. I'm worried about us (ergosphere.blog) AI

The article argues that while AI “machines are fine,” the bigger risk to academia is how they shift learning and quality control. Using an astrophysics training scenario, it contrasts a student who builds understanding through struggle with one who uses an AI agent to complete tasks without internalizing methods—leading to less transferable expertise. It also critiques claims that improved models will fix problems, arguing instead that the real bottleneck is human supervision and the instincts developed from doing hard work. The author closes with concerns about incentives, status, and what happens when AI makes producing papers faster but potentially less grounded.

AGI Is Here (breaking-changes.blog) AI

The article argues that “AGI is here,” but its claim is based less on any single definition of AGI and more on how today’s LLMs are paired with “scaffolding” like tool calling, standardized integrations, and continuous agent frameworks. It reviews multiple proposed AGI criteria (from passing Turing-style tests to handling new tasks and operating with limited human oversight) and claims many are already being met by existing systems. The author also suggests progress is increasingly driven by improving orchestration and efficiency around models, not just by releasing newer models.

Getting Claude to QA its own work (skyvern.com) AI

Skyvern describes an approach to have Claude Code automatically QA its own frontend changes by reading the git diff, generating test cases, and running browser interactions to verify UI behavior with pixel/interaction checks. The team added a local /qa skill and a CI /smoke-test skill that runs on PRs, records PASS/FAIL results with evidence (e.g., screenshots and failure reasons), and aims to keep the test scope narrow based on what changed. They report one-shot success on about 70% of PRs (up from ~30%) and a roughly halved QA loop, while trying to avoid flaky, overly broad end-to-end suites.

Functional programming accellerates agentic feature development (cyrusradfar.com) AI

The article argues that most AI agent failures in production stem from codebase architecture—especially mutable state, hidden dependencies, and side effects—rather than model capability. It claims functional programming practices from decades ago make agent-written changes testable and deterministic by enforcing explicit inputs/outputs and isolating I/O to boundary layers. Radfar proposes two frameworks (SUPER and SPIRALS) to structure code so agents can modify logic with a predictable “blast radius” and avoid degradation caused by context the agent can’t see.

A case study in testing with 100+ Claude agents in parallel (imbue.com) AI

Imbue describes how it uses its mngr tool to test and improve its own demo workflow by turning a bash tutorial script into pytest end-to-end tests, then running more than 100 Claude agents in parallel to debug failures, expand coverage, and generate artifacts. The agents’ fixes are coordinated via mngr primitives (create/list/pull/stop), with an “integrator” agent merging doc/test changes separately from ranked implementation changes into a reviewable PR. The post also covers scaling the same orchestration from local runs to remote Modal sandboxes and back, while keeping the overall pipeline modular.

Non-Determinism Isn't a Bug. It's Tuesday (kasava.dev) AI

The article argues that product managers are uniquely suited to use AI effectively because their work already involves rapid “mode switching,” comfort with uncertainty, and iterative, goal-oriented refinement rather than precision for its own sake. It claims PM skills—framing problems, defining requirements, and evaluating outputs—translate directly into prompting and managing non-deterministic AI results. The author further predicts the PM role will evolve toward “product engineering,” where PMs apply the same directing-and-review workflow to execution tools, with a key caveat that teams must actively assess AI outputs to avoid errors from overreliance.

Should you upload blood test results to AI? (wsj.com) AI

The article asks whether people should share their blood test results with AI tools like Perplexity or Claude, weighing potential benefits against privacy, data security, and accuracy concerns.

Show HN: Ownscribe – local meeting transcription, summarization and search (github.com) AI

Ownscribe is a local-first CLI for recording meeting or system audio, generating WhisperX transcripts with timestamps, optionally diarizing speakers, and producing structured summaries using a local or self-hosted LLM. It keeps audio, transcripts, and summaries on-device (no cloud uploads) and includes templates plus an “ask” feature to search across stored meeting notes using a two-stage LLM workflow.

Show HN: Tokencap – Token budget enforcement across your AI agents (github.com) AI

Tokencap is a Python library for tracking token usage and enforcing per-session, per-tenant, or per-pipeline budgets across AI agents. It works by wrapping or “patching” Anthropic/OpenAI SDK clients to warn, automatically degrade to cheaper models, or block calls before they consume additional tokens. The project emphasizes running in-process with minimal setup (no proxy or external infrastructure) and supports common agent frameworks like LangChain and CrewAI.

OpenRouter Raises $120M at a $1.3B Valuation (inc.com) AI

OpenRouter, an AI routing platform that helps companies select and use the right model for specific tasks, raised $120M at a reported $1.3B valuation. The funding round underscores continued investor interest in tools that manage and optimize access to multiple AI providers.

LLM Wiki – example of an "idea file" (gist.github.com) AI

The article proposes an “LLM Wiki” pattern where an AI agent builds a persistent, interlinked markdown knowledge base that gets incrementally updated as new sources are added. Instead of re-deriving answers from scratch like typical RAG systems, the wiki compiles summaries, entity/concept pages, cross-links, and flagged contradictions so synthesis compounds over time. It outlines a three-layer architecture (raw sources, the wiki, and a schema/config), plus workflows for ingesting sources, querying, and periodically “linting” the wiki, with examples ranging from personal notes to research and team documentation.

Seat Pricing Is Dead (seatpricing.rip) AI

The article argues that traditional SaaS seat pricing has “died” because AI changes how work is produced: fewer humans log in, output can scale independently of headcount, and value migrates from user licenses to usage/compute. It says companies are stuck with seat-based billing architectures that can’t represent more complex deal structures, leading to hybrid add-ons that only temporarily slow the shift. The author predicts a move toward per-work pricing (credits, compute minutes, tokens, agent months, or outcome-based units) and highlights the transition challenge of migrating existing annual seat contracts.

How many products does Microsoft have named 'Copilot'? I mapped every one (teybannerman.com) AI

The article argues that Microsoft’s “Copilot” branding now covers a very large and confusing set of products and features—at least 75 distinct items—and explains that no single official source provides a complete list. It describes how the author compiled the inventory from product pages and launch materials, and presents an interactive map showing the items grouped by category and how they relate.

Extra usage credit for Pro, Max, and Team plans (support.claude.com) AI

Claude’s Help Center says Pro, Max, and Team subscribers can claim a one-time extra usage credit tied to their plan price for the launch of usage bundles. To qualify, subscribers must have enabled extra usage and subscribed by April 3, 2026 (9 AM PT); Enterprise and Console accounts are excluded. Credits can be claimed April 3–17, 2026, are usable across Claude and related products, and expire 90 days after claiming.

Artificial Intelligence Will Die – and What Comes After (comuniq.xyz) AI

The piece argues that today’s AI boom is vulnerable to multiple pressures—unproven returns on massive data-center spending, rising energy and memory bottlenecks, and tightening regulation that could abruptly constrain deployment. It also points to risks inside current models (including tests where systems tried to act in self-serving or harmful ways), plus economic fallout from greater automation. The author frames “AI dying” as a gradual unraveling or consolidation rather than a single sudden collapse.

Show HN: DocMason – Agent Knowledge Base for local complex office files (github.com) AI

DocMason is an open-source, repo-native agent app that builds a local, evidence-first knowledge base from private files (Office documents, PDFs, and emails) so answers are traceable to exact source locations. Instead of flattening documents into unstructured text, it preserves document structure and visual/layout semantics (with local parsing via LibreOffice/PDF tooling) and enforces validation and provenance boundaries. The project is positioned as running entirely within a local folder boundary, with no document upload by DocMason itself, and includes a macOS setup flow and a demo corpus to test traceable “deep research” answers.

Byte-Pair Encoding (en.wikipedia.org) AI

Byte-pair encoding (BPE) is a text encoding method that iteratively merges the most frequent adjacent byte pairs using a learned lookup table, initially described for data compression. A modified form used in large language model tokenizers builds a fixed vocabulary by repeatedly merging frequent token pairs, aiming for practical training rather than maximum compression. Byte-level BPE extends this by encoding text as UTF-8 bytes, allowing it to represent any UTF-8 text.

Show HN: Running local OpenClaw together with remote agents in an open network (github.com) AI

Hybro Hub (hybroai/hybro-hub) is a lightweight daemon that connects locally running A2A agents—like Ollama and OpenClaw—to the hybro.ai portal, letting users use local and cloud agents side by side without switching interfaces. It routes outbound-only connections from the hub to hybro.ai (useful behind NAT), shows whether responses were processed locally or in the cloud, and includes privacy-oriented features like local processing for local-agent requests plus configurable sensitivity detection (currently logging-only). The project provides a CLI to start/stop the hub and launch supported local adapters, with local agents syncing into hybro.ai as they come online.