| View in browser: The Diff No. 007 |
| The Diff No. 007 Saturday, 10 October 2026 · AI dev news, only what is verified |
 |
The runtime has a one-year clock now. Ryan Dahl says the whole Deno team is joining Cloudflare. Deno keeps monthly bug and security releases for a year, then Deno development ends. Deploy runs for six months. Today’s other thread is cheaper judgment: a developer who stopped caring which model he was talking to, a Mac arrow that points at the button only you can press, and a Microsoft model that scores options instead of writing prose. Cheap, fast and contained beats clever, when the work is routine. |
IN THIS ISSUE | 01 🦕 Deno’s clock just started Same code, new migration date. | | 02 💸 He stopped checking the model name A month of cheap Flash sessions, one critic. |  | 03 ➡️ A big arrow for the button only you can press Agents get a finger, not a click. | | 04 🧭 A model that only answers with options Routing and grading, not prose. |  | 05 🚶 Turn the reading pile into a walk Twenty minutes in, nineteen minutes out. |  | 06 📦 Microsoft’s sandbox has an audit switch Useful policy layer, one dangerous flag. |
|
🔥 VIRAL The loud ones on Hacker News. Numbers are a morning snapshot. | 01 HN: 1,101 points at 06:25 IDT · announced Oct 9 · image: GitHub card · denoland/deno 🦕 Deno’s clock just startedSame code, new migration date. SCENE Ryan Dahl’s post says the entire Deno team is joining Cloudflare. They will keep shipping monthly bug and security releases for another year. After that they will end development of the Deno runtime. Deno stays open source, and he says others are welcome to continue it. Deno Deploy shuts down after six months. THE THING The stated plan: paying Deploy customers get migration support to Cloudflare Workers. JSR keeps running, with its infrastructure moving to Cloudflare. The team will keep supporting rusty_v8 and work toward integrating it into workerd. Dahl says the new direction is celld, built on the Workers programming model, with Durable Objects as the base for agent harnesses. That is his plan, not a shipped product. ✂️ STEAL IT List what actually runs on Deno: servers, scripts, Deploy projects, JSR packages. Mark the deadline that applies to each: six months for Deploy, one year of runtime releases. Put one Node or Workers port of the smallest service on the calendar before you need it. Watch VERDICT This is a dated migration notice, not an emergency. A runtime can stay useful after active development ends, but security-only releases are a different bet from a roadmap. Deploy is the near deadline. Announcement · HN discussion Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew it |
02 HN: 1,087 points at 06:25 IDT · post Oct 7 💸 He stopped checking the model nameA month of cheap Flash sessions, one critic. SCENE A developer who writes at dgt.is says he used DeepSeek 4.1 Flash heavily for a month across a dozen projects. Mid-session, he says, he often could not tell whether he was using it or Opus. He pays $10 a month for OpenCode Go and says he has rarely gone over about $1 of expected cost in a session. For occasional critical work he asks Opus 5.5 for a final review, then has DeepSeek apply the fixes. THE THING The change he describes is behavioral: cheap models make it easy to start scratch tasks, UI monkey testing and tidy-up chores. He calls this subjective experience and links a benchmark chart we did not verify. The post also makes claims about cache size, energy use and training data. We did not check those. HN gave the post a lot of attention; attention is not evidence. ✂️ STEAL IT Pick one low-risk task type: file cleanup, UI click-through or a throwaway prototype. Run it on a cheap model with a clear done test and a spending cap. Escalate only when the test fails or a change touches production data. Try today VERDICT Only for disposable or reviewable work. Not proof that DeepSeek matches Opus on your repository, and the cost figures are the author’s. Source review, not our test. Author’s post · HN discussion Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew it |
03 482 stars · MIT · repo created Oct 8 · HN: 385 points · card: GitHub ➡️ A big arrow for the button only you can pressAgents get a finger, not a click. SCENE Franz Enzenhofer’s README starts with the familiar failure: the agent can refactor a repository, then asks you to click Allow in a terminal you are not watching. His macOS tool draws an arrow and a sign over every window. Clicks pass through, keyboard focus stays put, and the arrow removes itself. THE THING bigarrow is a Swift command-line tool plus a skill for Claude Code and Codex. It points at a named button, such as the Allow button in System Settings. It never clicks it. The README says drawing needs no macOS permission, while finding a button or window can require Accessibility or Screen Recording for the app that starts it. Several demos are staged scenes, labeled that way in the README. Project tests and CI numbers are the author’s. We did not run it. ✂️ STEAL IT Use the idea even if you skip the tool: an agent needing a human action should say where, when and why on the screen the human is actually watching. Keep permission prompts you cannot read out of automatic-yes loops. Try today VERDICT Good for a Mac side project where an agent gets stuck on permissions. Not a permission bypass, not cross-platform, and a screen overlay will not help if the human cannot tell what the button grants. Repo + README · HN discussion Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew itTest it for me |
🔭 UNDER THE RADAR Quiet tools and field reports worth your time. | 04 HN: 157 points · Microsoft launch Oct 9 · official input price shown below 🧭 A model that only answers with optionsRouting and grading, not prose. SCENE Microsoft says Xbox Research used Microsoft-Decision-1 on more than 10,000 open-ended pieces of feedback from surveys, Steam and X. The job was sorting them into themes the researchers had already chosen. Microsoft reports quality competitive with GPT-6 Sol, 14 times faster and 200 times less expensive. THE THING Decision-1 is Microsoft’s post-trained Qwen3.5-9B. Given a fixed list of answers, it returns a calibrated probability for each option. It handles yes or no, multiple choice, ratings and rubric grading through a structured API call. Microsoft says it tested 36 additional public and private benchmarks across nearly 150,000 questions, and that decisions flipped on 1.3% of perturbations. All of those figures are Microsoft’s own. We did not call the model. ✂️ STEAL IT Find one place where your agent asks a large model a small question: route this ticket, accept this answer, pick this tool. Log the options and the model’s choice for a week. That log shows whether a cheaper scorer is worth testing. Watch VERDICT Useful category, vendor evidence. Check calibration on your own examples, the exact current price and availability in your region before routing anything important. Microsoft launch post · Foundry model page Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew it |
05 0 stars · MIT · repo created Oct 7 · image: Nazar Boyko 🚶 Turn the reading pile into a walkTwenty minutes in, nineteen minutes out. SCENE Nazar Boyko built Tabs to Trails for the reading backlog he never got through. He picks a walk length and the app chooses saved articles that fit. His reported test: a ten-minute request produced 10:15 of audio, and a twenty-minute request produced 19:39, on an M5 Max. THE THING A bookmarklet saves pages. Local Gemma 4 condenses text only when it does not fit. A Kokoro voice reads it, and ffmpeg builds one MP3 with a halfway cue. The README says its number guard checks numbers in condensed text against the source. It does not claim to catch a changed meaning, which is exactly where a rewrite can go wrong. Setup needs Node.js 22 or newer, Ollama, a 6.6 GB Gemma pull and a first voice download of about 330 MB. It serves a phone page on your local network. We did not run it. ✂️ STEAL IT Steal the constraint, not the stack: ask an audio tool for a duration, then check what it dropped. Read a condensed summary of an important source before you rely on it. Try today VERDICT A weekend experiment for people comfortable with Ollama. It has no stars or issues yet, so expect rough edges. Do not treat condensed text as the source, and review the local-network server before using it on a shared network. Author’s write-up · Repo + install · Issues Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew itTest it for me |
06 2,442 stars · MIT · repo created Feb 6 · HN: 185 points · card: GitHub 📦 Microsoft’s sandbox has an audit switchUseful policy layer, one dangerous flag. SCENE The MXC README presents a containment SDK for code execution rather than another agent harness. Think of the moment an agent runs a command you have not read. MXC is meant to place that command inside a policy-controlled boundary. THE THING The Rust, .NET and Node SDK uses a Windows process container, Linux bubblewrap or macOS seatbelt by default. The README marks some backends experimental, and network support depends on the backend. Its --audit mode disables sandbox security and is not for untrusted code. Optional Windows telemetry has several opt-in conditions. We read the README and issue list, but did not execute the SDK. ✂️ STEAL IT Write down your command policy first: readable paths, writable paths, network and time limit. Then test one denial on each operating system you care about. A sandbox that only works as a demo is not a boundary. Watch VERDICT Worth tracking if you run untrusted code, with one caveat: an active issue tracker and experimental backends mean this is not boring infrastructure yet. Do not use audit mode as a convenience switch, and do not assume every backend isolates the network equally. Repo + README · Open issues · HN discussion Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew itTest it for me |
📊 THE ATTENTION GAP HN points at 06:25 IDT. Attention, not quality. | Deno | | 1,101 | | DeepSeek post | | 1,087 | | bigarrow | | 385 |
|
🗂️ THE BORING STUFF Oct 9 · Claude Code 2.1.295. The changelog now lists OSC 7501 support and an onFailure:block option for command and HTTP hooks. This updates yesterday’s terminal-status report; it is not a new proposal. Changelog → Oct 9 · Open Intelligent UI. CopilotKit published an MIT-licensed follow-up to No. 005’s closed ChatGPT UI story. The repo is older, created March 12. Its docs describe a sandboxed iframe without same-origin storage and validated sendPrompt and openLink callbacks. The default demo needs a model key, and a connection test can make a paid call. Docs reviewed, not run. Announcement → Official sources checked this morning. No paid model calls or project installations made. |
🧪 ONE THING TO TRY TONIGHT Put the human step on the screen. Pick one agent job that sometimes needs you. Make it say where to look, what the button grants and what happens if you wait. Then check that the action was really taken, not merely announced. |
What this issue could not see. HN, GitHub, official posts, docs and first-person reports were read; projects were not run. Deduped against every published issue. Reddit returned rate-limit errors and X, Product Hunt and Bluesky returned no usable posts, so those platforms are not covered today. LinkedIn was a login wall. Lobsters and dev.to were sampled overnight. Video discovery found no current video that passed review, so no video is recommended rather than showing an older one. |
|
| Images: credited in each report; source screenshots and announcement artwork, not our tests. |
|