AI

via RustTrending (@RustTrending@yuri.garden): magnitudedev / magnitude: Open source inference engine that tunes its kernels on your hardware to find the fastest way to run each model. 2x faster than Ollama. Works on Apple Silicon, NVIDIA, AMD, Intel, or just a CPU. ★7588 https://github.com/magnitudedev/magnitude

Magnitude is an open-source local inference engine that auto-tunes its kernels to whatever hardware you’re running — Apple Silicon, NVIDIA, AMD, Intel, or plain CPU — and claims roughly 2x the throughput of Ollama. If you run models locally, it’s worth benchmarking against your current setup.

Rachel Laycock argues that experienced developers’ hesitancy about coding agents isn’t nostalgia for writing code or fear for their jobs — it’s hard-won knowledge that “built” and “works in production” are very different things. The questions she lists (Are these the right boundaries? Will we be able to evolve it? What happens when something breaks?) are exactly the ones that agentic workflows tend to gloss over.

via Assaf 🌴 (@assaf@mas.to): Attached: 1 image

Survival LLM bundles a local language model with an offline emergency library — FEMA and CDC guidance, medical and repair references, and about 930,000 Wikipedia articles — so you can ask how to purify water or treat a burn with no internet connection. Every answer cites the source articles it read, with the relevant passage highlighted, which is a smart hedge against small-model hallucination. Free, open source, and available for Apple Silicon Macs and Windows.

via Assaf 🌴 (@assaf@mas.to): > This is where our industry is headed: Owning a software system will depend a lot less on having written it (or similar systems) by hand; instead, what will become most important is knowing just enough to know when something is wrong, and being willing and able to dive as deep as needed, when needed.

Adrian Quiroga uses a SonarQube complexity failure on agent-written code as a jumping-off point for a bigger question: what do you still need to know when agents write most of the code? His answer — fundamentals and breadth to recognize when something’s wrong, plus the ability to dive deep on demand — feels like the right frame for how ownership of software is shifting.

via Assaf 🌴 (@assaf@mas.to): Attached: 1 image

magpie is an open-source macOS menu bar app that runs a local gateway translating between OpenAI Responses, Anthropic Messages, and Gemini APIs, so any coding agent can talk to any model provider. Point Codex at DeepSeek or Claude Code at Kimi with one click, and it rewrites just the config key you changed. It also masks secrets from .env files and logs before requests leave your machine.

via Seths.blog (@thisissethsblog@c.im): Create a handoff doc

Seth Godin suggests that the handoff doc — the thing nobody ever has time to write — is exactly the kind of work an AI assistant should be doing continuously in the background. He includes a ready-to-paste prompt that asks the AI to maintain a structured document covering current state, where things live, decisions and their rejected alternatives, unverified assumptions, open questions, and how to unwind the whole thing. The insistence on recording the “why” and marking inferences separately from stated facts is what makes it more than a status report.

via Matt Weagle (@mweagle@hachyderm.io): “Some of these predictions are just observations, they’ve long been true at companies like Amp. Others will take some time to play out.”

Thorsten Ball plants a flag on a set of deliberately provocative predictions: code review and unit tests are already dead, the craft of writing code disappears while the craft of building software matters more, open source in its current form stops making sense, and the terminal gets washed away by tokens. Whether or not you agree, it’s a useful snapshot of what an AI-forward developer believed in late 2026 — and a good list to argue with.

via El Perro Negro (@elperronegro@shmg.online): For 9 months I have been using RapidRaw as part of my RAW image workflow. So much so that Rawtherapee and Lightroom have fallen by the wayside.

RapidRAW is a non-destructive, GPU-accelerated RAW photo editor that ships in about 20 MB — smaller than many of the images it edits. Built with Rust and Tauri by a solo developer, it offers AI masking, presets, batch edits, and optional generative tools that can run locally, via your own ComfyUI server, or in the cloud.

via Barney Dellar (he/him) (@BarneyDellar@mastodon.scot): “Do you still read the code?” This is a very interesting article on the state of software development, and what it means to write software. https://zanlib.dev/blog/do-you-still-read-the-code/

A familiar state of affairs: teams now share codebases without agreeing on whether anyone is still expected to read the code. Zanlib draws a useful distinction between “accelerators,” who use AI to translate their own understanding into code and stay responsible for it, and “vibe coders,“ who delegate implementation and its revision entirely. The costs of either choice won’t be visible until the code has to survive changes in requirements, tooling, and people.

Crit is a local, browser-based review surface for AI agent output: point at line 47 of the diff, type “this drops the refresh token,” and the agent picks up the comment and fixes it. Beyond code diffs it also handles plans and docs, running dev servers, and static HTML artifacts — each of which terminal scrollback handles badly. Self-hostable and works with git, jj, and sapling.

via Assaf 🌴 (@assaf@mas.to): Throughout a single piece of writing you’ll face two different questions: What context does the reader already have? and, crucially, What context does the reader need? The gap between those two answers defines the piece.

Addy Osmani has distilled eighteen rules for clear writing into an open-source Agent Skill that can interview you, rewrite a draft, or simply critique one without touching your file. There’s also a private browser-based editor that flags AI writing tells, readability problems, and gaps in substance. The underlying essay stands on its own as good writing advice, whether or not you ever install the skill.

Anshu Chimala, who led engineering and design teams at Apple for 12 years, argues that AI models aren’t bad at design — we’re just prompting them into design-by-committee. Because LLMs predict the most likely next token, they default to the blandest possible choices; this piece lays out an end-to-end process for pushing past that toward genuinely distinctive work.

CodeScene combines static analysis with behavioral analysis of your version-control history to surface the hot spots where technical debt actually costs you — and it scores maintainability with a “CodeHealth” metric that’s validated against defect and delivery data.

via John Voorhees (@johnvoorhees@macstories.net): Hybrid compute for Perplexity’s Personal Computer is an exciting development that runs on Apple silicon Macs and blends cloud-based frontier models with local AI. Details on MacStories.

Perplexity’s new “hybrid compute” runs agent tasks in the cloud by default, but a local classifier flags potentially sensitive data and offers to route that portion to a model running on your Apple silicon Mac — and it works for sessions started from an iPhone or iPad, too. Data privacy is the next big blocker for agentic tools in regulated industries, so it’s good to see someone tackling it directly rather than asking customers to trust a cloud endpoint.

via Assaf 🌴 (@assaf@mas.to): Attached: 1 image

omp is a coding agent that runs with an IDE wired directly into it, so the model can use real editor and language-server context instead of guessing from raw text. An interesting take on where agentic coding tools are heading: less chat window bolted onto a terminal, more agent living inside the dev environment.

via Assaf 🌴 (@assaf@mas.to): Ephemeral is fine. We used the sleep app for about four months. Our son sleeps through the night now, so we retired it. If the app had taken me months to build, that might sting. Because it took a week, I’m just glad it worked when we needed it.

Adam Waxman picks up Robin Sloan’s “an app can be a home-cooked meal” thread and shows what it looks like in 2026: a sleep tracker built from his consultant’s PDF, a fitness app that sizes his morning smoothie to that day’s run, a marathon plan derived from his actual race data. With AI assistance, software written for an audience of one (or four) is now a weekend project instead of a career.

This rings very true for me.

via Glenn Fleishman (@glennf@zeppelin.flights): Why pay for bad advice

Tim Ferriss opens his own BookScan numbers to argue that LLMs are eating the how-to nonfiction category alive: adult nonfiction down 9% year-over-year in Q1 2026, with self-help down 26.3%. It’s a rare, concrete look at AI disruption hitting an industry in real time rather than in forecasts — and a preview of what happens to any business built on packaging answers people can now just ask for.

Linear is repositioning itself as “the product development system for teams and agents” — issue tracking, planning, and triage built around workflows shared by humans and AI agents. Worth a look if you’re thinking about how coding agents fit into an existing project management practice (and, as always, the design and performance polish of the product itself is exemplary).

via Ars Technica (@arstechnica@mastodon.social): Attached: 1 image

Someone fed the text of the MIT License into Suno’s AI music generator with the prompt “sad girl with piano” and got back a genuinely affecting piano ballad. The highlight: it nails “WARRANTIES OF MERCHANTABILITY” with an Imogen Heap-style glissando, then immediately mangles “FITNESS” into “fistiff.”

via Miguel de Icaza ᯅ🍉 (@Migueldeicaza@mastodon.social): Attached: 1 image

A short video gag from Ryan Els capturing the daily relationship between a developer and their IDE — including the autocomplete that insists it knows what you meant to type. The replies are just as relatable: “I just want to use the tab button!”

A Siri shortcut that captures highlights from whatever you’re listening to in Overcast, transcribes and summarizes the surrounding audio, and syncs the result to Readwise (and from there to Obsidian, Notion, Roam, etc.) or email. “Hey Siri, highlight” is exactly the interaction I want when a good idea goes by while my hands are busy.

Via https://overcast.fm/+QCTSW_xSc/15:55

Design concept for an icon to easily identify AI-generated concept.

via Ars Technica: A jargon-free explanation of how AI large language models work