Midjourney Medical

Midjourney announced a medical imaging initiative exploring low-cost, AI-enhanced tomography using phased array beamforming techniques, aiming to reduce radiation exposure compared to conventional CT scans. Commenters with medical imaging experience cautioned that nanometer deflection sensitivity doesn’t translate to practical image resolution, and raised concerns about over-detection creating unnecessary follow-ups. Others questioned how Midjourney’s image-generation expertise applies to a fundamentally different imaging modality.

Local Qwen isn’t a worse Opus, it’s a different tool

Alex Ellis shares his experience running local Qwen models (27B, 35-A3B) for a small software business, finding real but caveated value — the RTX 6000 Pro hardware paid for itself within 2-3 months through reduced API costs. Commenters confirmed local models require different prompting techniques than cloud models, comparing it to playing a different instrument. Several noted local models remain expensive when factoring in hardware depreciation and electricity, though the cost calculus shifts for teams with sustained high-volume usage.

US holds off blacklisting DeepSeek, more than 100 firms deemed security risks

Reuters reports the US paused blacklisting DeepSeek while designating over 100 other Chinese firms as security risks, reflecting complex trade-offs between national security and existing export restrictions. Commenters questioned how US-enforced AI restrictions could work technically, drawing comparisons to a “Great Firewall of America.” Some noted Z.ai (maker of GLM 5.2) has been on the Entity List since January 2025 yet continues operations with US exports restricted, highlighting enforcement challenges.

Taxonomy of the Occlupanida (parasitoids on bread bag tags)

The Holotypic Occlupanid Research Group classifies bread bag clips (“occlupanids”) using synthetic taxonomy as organisms in the phylum Plasticae, with no genetic data, fossils, or reproductive biology to work from. Commenters found the elaborate biological framing hilarious after initial confusion about what occlupanids actually are. Users noted this is one of the most frequently resubmitted pages in HN history, with one person reporting a physical collection of 100+ specimens and another confirming they’d bought a lab coat for the joke.

x86 AI Compute Extensions (ACE) Specification

The x86 Ecosystem Advisory Group published a specification for AI Compute Extensions (ACE) adding matrix multiplication primitives for ML workloads, defining new tile and block scale register state with data processing operations on AVX inputs. Commenters asked how ACE differs from existing SSE/AVX instructions available in most x64 machines. One user speculated ACE would initially be limited to server CPUs for years, similar to the slow rollout of AVX-512. Several expressed relief that the specification avoided patent-encumbered approaches.

Storied Colors — a catalogue of named colors

Storied Colors is an ongoing index cataloguing named pigments with their provenance, chemistry, and historical context — one color per day. Each entry documents where a pigment was first ground, whose canvas it dried on, when it went extinct or fell out of fashion. Commenters noted Rebecca Purple, a CSS named color added in tribute to Eric Meyer’s daughter, as the web’s own storied color with a tragic backstory. One color expert felt the displayed VanDyke Brown looked more like Burnt Sienna and recommended monitor calibration.

Clojure Hosted on Go

Glojure is a Clojure interpreter hosted on Go with extensible interop support, enabling Clojure code to call Go libraries directly and bringing Clojure-style functional programming to the Go ecosystem. Commenters noted this is the most promising Clojure-on-Go implementation due to its proper interop support compared to earlier efforts. One user raised questions about REPL implementation — whether it compiles to Go then executes or ships with a full VM. Performance comparisons to mainstream JVM Clojure were a recurring discussion point.

Show HN: Spin Lab

Spin Lab is a browser-based interactive explainer that visualizes table-tennis spin mechanics including topspin, backspin, spin rate, trajectory, and bounce behavior with scrubbable animations. Mobile users reported the first window’s text cannot be scrolled, making the interface unusable on phones. Some felt the UI looked AI-generated without sufficient manual polish before sharing for public feedback. Several commenters expressed interest in the physics simulation but wanted an open-source version.

Show HN: We built an 8-bit CPU as 2nd year EE students

Second-year EE students built an 8-bit Harvard CPU from individual logic gates using Logisim-Evolution simulation, implementing a hardwired control unit rather than relying on microcode or pre-built ALU components. Commenters noted the project builds on Ben Eater’s SAP design from “Digital Computer Electronics” by Malvino and Brown. Some with FPGA experience remarked that Verilog if/else statements create MUXes without fanout issues, while hardwired implementations in simulation expose deeper architectural understanding.

Loreline — Tools for writing interactive fiction

Loreline is an open-source language and toolset for writing interactive fiction, video game dialogue, and branching narratives, integrating with game engines, web apps, and standalone projects while keeping stories portable. Commenters drew comparisons to Inky/Ink from inkle studios (creators of Sorcery! and 80 Days). Others praised Inform 7 as the best language for interactive fiction, noting how programmers react to reading Inform 7 code. One user commented favorably on Loreline’s translation and localization support.

How we run Firecracker VMs inside EC2 and start browsers in less than 1s

Browser Use Cloud rebuilt its infrastructure to run Firecracker micro-VMs inside EC2, cutting browser session costs from $0.06 to $0.02 per hour by optimizing Chromium memory allocation and boot sequence to start browsers in under a second. Several commenters questioned the ethics of building undetectable browser infrastructure used to evade anti-bot measures. A user noted nested virtualization on regular EC2 was only possible since February 2026, previously requiring metal instances. One praised the engineering but worried about surveillance capitalism applications.

How Madrid built its metro cheaply (2024)

Madrid tripled its metro system from 71 miles to 197 miles in just 12 years (1995-2007), faster and cheaper than almost any other city globally, relying on a stable team of well-paid in-house engineers rather than external consultants. Commenters contrasted Madrid’s approach with the Bay Area’s fragmented transit systems across dozens of agencies despite a $30T regional market cap. Several noted US construction wages in NYC and SF being 2-3x those in Madrid as a structural cost difference. Others pointed to Spain’s permissive expropriation laws as an enabling factor.

Biological evolution and information acquisition

Brian Potter examines how biological evolution acquires information through random variation and selection, drawing parallels to technological evolution — evolution builds complex organisms by combining existing useful components as modules, scaling up through combination rather than pure innovation. Some commenters noted the article’s assumption that mutations are “entirely random” is imprecise — transposons have preferential hotspots and individual nucleotide positions vary in mutability. One reader connected the concept to how scientific knowledge is built from smaller composable ideas.

Tesco moving 40k server workloads off VMware amid Broadcom’s abusive conduct

Tesco is moving 40,000 server workloads off VMware, citing Broadcom’s abusive conduct in UK court filings including a ~175% price hike and breach of contract allegations. Commenters noted Broadcom’s well-known business model: acquire declining but profitable enterprise software, slash support and R&D, and raise prices aggressively. Several pointed to Proxmox as the biggest beneficiary of Broadcom’s VMware pricing strategy. Some expressed skepticism about the feasibility of migrating 40,000 workloads while maintaining retail operations during a cost-of-living crisis.

Nim Conf 2026 (Online, Sat June 20)

NimConf 2026 is an online conference on Saturday June 20 at 11:00 UTC, streamed for free with no travel required. For the first time, the conference features two parallel tracks — one for community projects and one for broader social and workflow talks, with separate YouTube premiere playlists for each track.

Scaling opencomputer from 1 VM to 1 million sandboxes

OpenComputer scaled from a single Azure VM to supporting potentially one million sandbox instances across four cloud providers. Growth hit an Azure compute quota ceiling in one region, forcing a redesign into cells with a global registry at the edge for sandbox discovery. The piece details practical lessons in multi-cloud resource management and the architecture behind sub-second sandbox provisioning. This story had no HN comments at time of collection.

U.S. science is in chaos

Scientific American reports that America’s compact between science and politics is broken, with NASA’s AXIS x-ray telescope mission among projects facing cancellation. Researchers describe grant funding drying up, foreign students unable to obtain visas, and a senior exodus at NSF and other agencies. One commenter’s wife, operating an optical trap microscope among roughly 2,000 people worldwide skilled at that instrument, has been crying frequently over the state of science research and plans to leave the country. Academics reported being unable to recruit international graduate students for Fall 2026 cohorts.

The Return of Rigorous Full-System Timing Simulation

SIGARCH discusses why cycle-level full-system timing simulation is returning as a critical tool for modern computer architecture. Modern platforms combining many-core CPUs, deep memory hierarchies, accelerators, and complex I/O have made accurate timing simulation essential for performance analysis and architectural exploration. This story had no HN comments at time of collection.

SteamOS Linux 3.8 released as stable

Valve released SteamOS 3.8 as stable for the Steam Deck, updating the base from KDE Plasma 6.2.5 to 6.4.3 and making Wayland the default display server. The update lands roughly one year after KDE 6.4 was released, with KDE 6.7 now available upstream. Wayland adoption was welcomed as a baseline improvement, with one user planning to try the Niri compositor on Steam Deck.

The founder’s playbook: Building an AI-native startup

Anthropic published a playbook on using Claude across coding, product, marketing, and operations to build a startup, focusing on automating what the authors describe as “2019-style app building.” Several commenters said the playbook is really about using Anthropic’s tools rather than how to build an AI-native startup — a “sell the shovels” approach. Some found the tone overly casual, making startup founding seem like something you decide to do over coffee. Others criticized the lack of concrete case studies or founder interviews.

Show HN: Local personal data redaction for any AI tools

PII-GUI is an open-source local tool that redacts personally identifiable information before sending data to any AI tool, running entirely on the user’s machine with pattern-based PII scanning. The tool aims to address privacy concerns when using cloud AI APIs by stripping sensitive data client-side. This story had no HN comments at time of collection.

Why do commercial spaces sit vacant? (2025)

Andrew Burleson explains “extend and pretend” — landlords keep commercial spaces vacant rather than lowering rents because paying tenants would demand rent reductions, crashing the market. A half-rented building at rent X can be more profitable than a fully rented building at significantly lower rents. Commenters questioned why “real” income streams can’t replace theoretical maximum rents when evaluating property values. A user noted that extend and pretend is unraveling as interest rates persist. One commenter shared personal experience running a small business unable to negotiate reasonable rent despite surrounding vacancies.

Pink Cosmo berries a hit in their trial season (2023)

Pink Cosmo berries, a new pink blueberry variety developed through selective breeding, performed well in their trial season. The pink color results from a double recessive gene for pigment, eliminating the blue anthocyanins that give standard blueberries their color. One commenter questioned the genetic description, suggesting “double recessive gene” rather than “dominant recessive gene” was the correct terminology. Others questioned the value of breeding blueberry varieties without the beneficial health pigments and wondered about consumer willingness to pay a premium for novelty.

A robot is sprinting towards you. Do you want it running on Claude or Grok?

OpenRouter dropped 11 LLMs into a 2D battle royale across 30 games to compare models on autonomous decision-making under pressure. Grok 4.1 Fast won 43% of matches, while Claude Sonnet 4.6 never won but kept asking others to team up and revealing its position. Commenters noted frontier-tier models (Opus 4.7, GPT-5.5) would have cost ~$3,000 for the same 30 games, raising viability questions at scale. One user confirmed DeepSeek V4 Flash being cost-effective is unsurprising for coding tasks.

Made a free macOS menu bar app that fixes typing in the wrong keyboard layout

FlicKey is a free macOS menu bar utility that remembers the input language per app, per browser tab, and per chat conversation, automatically switching keyboard layouts on focus change. One commenter said this has been their only serious macOS complaint for 20 years and shared their own open-source solution (InputLanguageKeeper). A Windows user noted the “let me use a different input method for each app window” setting has existed in Windows for years, expressing surprise macOS lacks equivalent native functionality.

ChatGPT’s image generator can be manipulated to produce violent, sexual content

Mindgard research claims ChatGPT’s image generator can be manipulated to produce violent and sexually explicit content through indirect prompt manipulation, suggesting content filters remain insufficient. Several commenters felt the write-up was overly dramatic, questioning why someone afraid of disturbing imagery would voluntarily red-team AI models. Others noted the finding isn’t a vulnerability — ChatGPT is simply responding to a prompt, and there’s no real security boundary being crossed. One compared it to finding that a sharpie can draw disturbing things if you press hard enough.

Show HN: BlitzGraph — Supabase for graphs, built for LLM agents

BlitzGraph is a graph database platform pitched as “Supabase for graphs,” built specifically for LLM agents with typed JSON queries, bidirectional relationships, and MCP server support enabling agents to connect from Claude or Codex. Y Combinator-backed and in public beta, it offers a free tier with a live playground. Commenter Reubend asked the founder to share specific issues encountered with Dgraph, TypeDB, and SurrealDB across 120+ opened issues. Another asked how it’s more suitable for agents compared to Neo4j and TypeDB, specifically inquiring about temporal modeling capabilities.

Using AI to improve a challenging reaction in medicinal chemistry

OpenAI connected GPT-5.4 with Molecule.one’s Maria platform to create a near-autonomous AI chemist that improved Chan-Lam Coupling yields for over 80% of tested substrates by discovering a surprising additive through AI-driven high throughput screening. A former chemist commented this resembles 1990s HTS with robotics, but with an AI engine mapping variable space instead of a trained chemist selecting variables. Others noted well-run autolabs recursively training discipline-specific models are becoming valuable assets, while one commenter wished Sam Altman had written the announcement himself for added credibility.

TREX: An AI code reviewer that runs your code

Greptile built TREX, an AI code reviewer that runs pull request code in a sandbox, generates artifacts, and drives interactive testing sessions to catch bugs before merge. It orchestrates multiple agents, designs multi-modal artifacts, and uses a model-agnostic evaluation harness. One commenter confirmed methods like this substantially improve coding agent output by plugging the feedback loop hole. Commenters raised practical concerns about handling external dependencies (S3 buckets, secret managers, third-party APIs) and potential security exploits via sandboxed code execution.

Language Courses in the Public Domain

The Yojik website hosts a comprehensive archive of public domain language courses from FSI, DLI, and Peace Corps, along with preserved Shtooka project audio collections and tutorials for importing vocabulary lists. Content includes professionally re-recorded Cortina language courses, with French, German, and Russian translations of the site in progress. This story had no HN comments at time of collection.