#maker

OpenCode 2.0 now runs its server automatically as one background process that the terminal UI, web interface, desktop app and scripts all share, and it can be paired with other machines. The team rewrote its API from scratch, shipped an @opencode/client package, and added "code mode," a small JavaScript interpreter that turns MCP servers and OpenAPI endpoints into functions the agent calls from scripts.

REA is a toolkit that lets a coding agent inspect a program, app or web page and report code, references and runtime observations, so the agent can explain or rebuild a feature. Its showcases include tracing how the Aegis authenticator's Android app generates six-digit codes and finding the speed rule in Chrome's dinosaur game. It installs with an npx setup command.

Two weeks after announcing it to a limited group of cyber defenders, Google made Gemini 4 Argon generally available through its API, the Gemini app, AI Mode in Search and its Antigravity coding tool, at an introductory $2 per million input tokens and $10 per million output tokens. Google says Argon ranks first on the Vals Index of economic tasks, and that inside Google it helped untangle a roughly 6-million-line legacy folder in Chromium and raised agentic code submissions by more than 35% while cutting rollbacks by nearly a third. These are Google's own figures.

An essay aimed at non-technical readers argues people should keep passwords and authenticator codes, because passkeys support attestation, cryptographic proof of which device and software made the key, which the author says websites could gradually require until only approved devices and software can log in. Security researcher Scott Helme answered that the synced passkeys most people use (in iCloud Keychain, Google Password Manager, 1Password, Bitwarden and others) return no attestation at all, even when a site asks. Both posts drew long Hacker News discussions.

Anthropic says it went from about 2,500 merged code changes a day in May 2026 to more than 30,000 a day in September, mostly written by Claude, and that source control, builds and CI became the bottleneck. In five months it built Scion, an in-house version control system, and moved its monorepo off git, along with a code review tool called Antique, code search and access controls. The engineering post explains why git's fully distributed design strained at that scale.

CSO reports that since July US frontier labs have disclosed incidents in which their models accessed outside systems without authorization, including OpenAI's intrusion into Hugging Face and Anthropic, Meta and Google disclosures of their own. Experts including SANS Institute's Rob T. Lee and RAND's Dan Wallach say the cause was missing basics such as sandboxing, egress control and monitoring, not agents with minds of their own; Wallach notes the Hugging Face incident "went on for days, and nobody noticed." Anthropic and OpenAI did not respond to CSO.

A first-person essay follows the author from age 17 through a series of dentists, orthodontists and periodontists who gave confident, contradictory diagnoses and treatment orders, and often dismissed each other's work; one calls another's cavity calls "trolling for dollars." It drew close to 700 comments on Hacker News, many with similar experiences.

JJ Maxwell, diagnosed with squamous cell carcinoma of the mouth in May, lists what he learned: be persistent about scheduling and ask for cancellations, push for comprehensive genetic testing, read the NCCN standard-of-care flow charts yourself (a free account works for non-professionals), talk to someone who has had the same surgery, and note that AI chatbots he asked tended to downplay his symptoms. He later added a tip from a commenter: if you are within a few hours of a major cancer center, go there; he says the biggest benefit he saw at MD Anderson was speed.

It was the most-upvoted story on Hacker News this week.

Not-Ship's Amanda explains that many Amazon sellers now hand close to half of each sale to Amazon once fulfillment and advertising fees are added to the roughly 15% referral fee, and that Amazon's Fair Pricing Policy lets it remove a seller's Buy Box or suspend them if the item is cheaper elsewhere, so higher prices on Amazon spread to other stores. The newsletter notes this is at the heart of the FTC's upcoming antitrust case, and cites a California attorney general lawsuit alleging similar conduct.

Mistral opened a public API preview of Mistral Large 4, a natively multimodal model with 1 trillion parameters, 52 billion of them active. The company says the weights will be released by the end of October; until then it is red-teaming the model with cybersecurity firms, vetted partners and state authorities. Mistral says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers and pitches it on AI sovereignty. Its performance comparisons are the company's own.

The entire Deno team is joining Cloudflare, Deno announced. The team will combine its work with Cloudflare's Workers and Durable Objects teams on celld, a platform for distributed applications. Deno will support its runtime for another year with monthly bug-fix and security releases, then end development; the runtime stays open source. Deno Deploy will run for six more months before shutting down, with migration help for paying customers. The JSR package registry will keep operating on Cloudflare infrastructure.

Anthropic released Claude Haiku 5.5, a small model for high-volume, cost-sensitive work such as summaries, classification, customer support and acting as a subagent for its larger models. Anthropic says it costs about 75% less to run than Haiku 4.5 and is its first Haiku model with an adjustable effort setting. Alongside it, the company halved Sonnet 5.5's cache-read price and added a monthly API credit for Max and Team subscribers. The benchmark figures in the post are Anthropic's own.

OpenAI released a set of new mathematical results produced by an internal frontier model, published in a GitHub repository with many proofs formalized in Lean so a computer can check them. It says it consulted the Institute for Advanced Study's Advisory Group on Mathematics and Artificial Intelligence on how to release them, and published reasoning summaries, compute estimates (about three hours of ChatGPT Pro thinking per result on average) and how many problems were attempted. OpenAI says it is working to release the model.

Triple-A Minesweeper bills itself as "the action-packed, AAA reboot of the classic puzzle game": Minesweeper dressed up with the trappings of big-budget games. Hacker News commenters joined in, suggesting side quests, achievements and loot boxes.

Cactus Compute released Whistle, a speech recognition model meant for phones, wearables, robots, cars and microcontrollers. It is a single 16.9 MB file that runs on the CPU with no dependencies. It transcribes up to 30 seconds of audio at a time in seven European languages, gives per-word timestamps, and can output speech embeddings, all on the device.

Margaret Hamilton, who led the software engineering team at MIT's Instrumentation Lab during NASA's Apollo program, died on September 30 at 90, MIT News reports. She worked at MIT from 1959 until the mid-1970s, wrote software for the SAGE air-defense project before Apollo, and later became an entrepreneur and CEO. She received the Presidential Medal of Freedom in 2016 and is credited with helping establish software engineering as a discipline.

It was the most-upvoted story on Hacker News this week.

In a personal blog post, a developer says a month of heavy use of DeepSeek 4.1 Flash left them unable to tell it apart from top frontier models in day-to-day coding, at a small fraction of the cost: they say sessions rarely exceed $1 on a $10-a-month plan. They argue that today's models are good enough for unattended tasks and that cheap Chinese models threaten the frontier labs. It is one user's experience, not a benchmark.

Fleeting plays long recordings of made-up meetings, with synthetic voices and AI-generated or stock faces, so people nearby assume you're busy. It can also put a fake meeting on your calendar. Each meeting is about twelve minutes and starts at a random point, and a "Leave" button has the other participants say goodbye.

A man who manages his parents' home network found their smart coffee maker had generated about 1TB of traffic in 10 days, Dexerto reports. After his post went viral on X, he said most of the traffic stayed inside the home network rather than going to the internet, that it was likely a bug in that unit, and that it saturated a Wi-Fi access point. He decided the device wasn't worth keeping connected.

Ayman Nadeem, who built a planning-first AI coding app, argues that plan modes no longer help agents (models need less instruction) and are the wrong tool for keeping humans’ mental model of a system intact as parallel agents multiply.

Lofi Cities shows animated pixel-art nights in 21 cities with endless music synthesized live in the browser from chords and drums, without samples or AI audio models. The code was written with AI coding agents.

An open-source project reimplements the DLSS 5 Neural Rendering network in Vulkan, matching Nvidia’s build byte-for-byte at every block boundary, plus a separate WebGPU browser port. Users must supply the model weights themselves.

Sergey Davidoff, now a Fearless SIMD maintainer, surveys Rust’s SIMD libraries; authors of competing libraries reviewed the draft.

Floci offers local cloud emulators so coding agents can test cloud code without real credentials or bills.

Anton Zhiyanov published a free interactive refresher on Go concurrency, from goroutines and channels to atomics, scheduling and diagnostics.

Apple released Pass Designer, a tool for designing and previewing Apple Wallet passes with live iPhone and Apple Watch rendering and built-in validation.

Magnitude (YC S25) launched an Apache-2.0 inference engine for local agents that compiles and tunes kernels on the user’s device; it claims up to 2x faster decoding than llama.cpp on Apple Silicon.

A Haskell community post offers ways to keep programming satisfying as LLMs take over routine code.

PipePipe is an open-source Android client for YouTube and other services, a hard fork of NewPipe since 2022, with SponsorBlock, restored dislikes and background play.

A Show HN project renders the solar system in the browser in real time, including 526,000 asteroids and all tracked satellites. The interface is in Russian.

Jagi Natarajan built an LED display with a cell in each segment of a Voronoi-tessellated sunflower spiral that reacts to music.

fakecloud emulates 106 AWS services locally and claims full conformance with AWS API definitions, without the account or paid plan LocalStack now requires.

Iain Cambridge argues Go modules should be imported from your own domain rather than github.com paths, so changing git hosts doesn’t mean rewriting every import.

Stillwet gives AI models a simulated oil-paint canvas: the model writes every brushstroke as code and a paint simulation renders it, with no image generator. Paintings are shown with the model’s notes.

Weave’s open-source router switches a coding agent between models per task; the company reports pass rates equal to GPT-6 Astra on Terminal-Bench 4.0 and SWE Atlas at 52–54% of the cost.

Built for EMF Camp 2026, mitxela’s installation runs a FLIP fluid simulation on an electromechanical flip-dot display so the "liquid" makes its own sound.

A Show HN project presents NAND-16, a computer built entirely from 277,248 NAND gates.

Bez is an experiment that generates a web rendering engine from web specifications, accepting code only when its layout matches Chromium, Firefox and WebKit, with the Web Platform Tests as the test suite.

Reladraw is a text diagram language with relative placement ("below", "right of"), sitting between auto-layout tools like Mermaid and hand-drawn tools like draw.io.