Code415 Radar

by Ziyu Guo · @Code415zg

What people are paying attention to in AI, LLMs and CS today. Every item comes with a short path to learn it and a small test to try it yourself. Updated every evening, Japan time.

Wednesday, November 4, 2026

20 items · Postman slips plugins into coding agents; Astra gets pushback
  1. #1Tools & open source

    Postman Desktop quietly added plugins and hooks to Claude Code, Codex

    A GitHub issue opened Oct 25 reports that Postman Desktop 12.33.4 enabled Postman plugins in Codex and Claude Code, added a SessionStart hook telling the agent to load a Postman API skill, configured the remote Postman MCP server and dropped a CLI into ~/.local/bin. Later commenters saw the same on builds up to 12.35.1 and in Cursor and GitHub Copilot CLI configs, and the reporter says the installer appears to record trust for its own hook. What is new on Nov 3 is a Postman staff reply promising a clearer opt in, plus a jump to the HN front page.

    Heat · HN front page on Nov 3, with point counts near or above a thousand depending on the snapshot; the issue body has 209 thumbs up and the Postman staff reply has 25 thumbs down
    What people foundPostman staff say the CLI and agent plugins are bundled with the desktop app, can be uninstalled inside the app, and that the team will make the opt in clearer for existing users and review hook approval and session start behavior. One commenter found the in app uninstall did not remove the plugin from Copilot CLI, so check each agent config by hand.
    Learn it15 min · 5 steps

    Key ideas

    SessionStart hook
    A command an agent tool runs automatically when a session starts, whose output is fed to the model as context.
    Agent plugin
    A bundle of skills, hooks and MCP server settings that a coding agent such as Claude Code or Codex loads from its config.
    Hook trust record
    A stored hash in the agent config marking a hook as approved, so it runs without asking the user again.

    Steps

    1. Read GitHub issue #14370 on postmanlabs/postman-app-support from the top, noting each file path the reporter lists.
    2. Read the follow up comments that mention Cursor and GitHub Copilot CLI to see how far the change reached.
    3. Read the Postman staff reply dated Nov 3 and list what it promises and what it leaves open.
    4. Skim the HN thread for alternatives people suggest and how they audit their own machines.
    5. Self check: which three places on a developer laptop could a desktop app write to in order to steer a coding agent?
    Try it30 min · 5 steps
    You need: Your own laptop with any of Codex, Claude Code, Cursor or Copilot CLI installed; a shell and a text editor. No API key needed.

    Steps

    1. Back up your agent config folders before touching anything, for example by copying ~/.codex and ~/.copilot to a dated folder.
    2. Search the Codex config for Postman entries with grep -n postman ~/.codex/config.toml and look for a marketplace entry, an enabled plugin and any trusted_hash line.
    3. Open ~/.copilot/settings.json and the Claude Code plugin list and look for plugins you never enabled yourself.
    4. List ~/.local/bin and check for executables you do not recognize, then note which app installed each one.
    5. Remove or disable anything unexpected, start a fresh agent session and confirm no unexpected startup text appears in context.

    Small angles to try

    • Repeat the audit for every MCP server entry and ask whether each one should be there.
    • Turn the checks into a small script that diffs your agent configs against a saved known good copy each week.
    • Compare how Codex, Claude Code and Cursor ask for approval before a new hook runs.
    Sources:GitHub issue #14370HN discussion
  2. #2Models & launches

    Astra pushback: Honnibal calls it misleading, Willison reruns it locally

    Following the Astra post on LLMs reading and writing neural network weights, spaCy creator Matthew Honnibal argues only three experiments, on networks of 17K to 172K weights, read raw weights with no tools, while the Qwen3.5-0.8B MATH-500 fix used Python and ran the model on up to 50 batches of practice problems. He also notes the Gemma 3 sycophancy detection task, experiment 7a, and the attempt to apply the Qwen style maths fix to Gemma 3, experiment 8b, both failed. Separately Simon Willison ran the no tools game guessing task on a local Qwen 3.8 27B, which named Pac-Man, Pong and Breakout correctly after 46 to 65 minutes each.

    Heat · HN: the original Astra post about 2,928 points; Honnibal's critique 969 points and about 200 comments; Willison's post about 611 points and 180 comments
    What people foundHonnibal points out the hand written MinAtar networks beat PPO only on Breakout, 517 vs 25, and lost on Seaquest, 53 vs 89, and Asterix, 53 vs 55. Willison calls the paper impressive but says he is "not ready to declare that the age of source code is over," and thinks a skilled person with pencil and paper could likely reach the same answers.
    Learn it15 min · 5 steps

    Key ideas

    Raw weight reading
    Giving a model only the numbers of another network's weights and asking what that network does, with no code execution and no running the network.
    Policy network
    A small neural network that maps a game observation to an action, here trained with PPO on simple Atari style games.
    Tool assisted weight editing
    Changing weights while the model may write Python and run the target network to test each change, which is a different skill from reading weights alone.

    Steps

    1. Read Honnibal's post first and copy his table of experiments, sizes and whether each used tools.
    2. Open the original Astra blog post and match each headline claim to an experiment number from that table.
    3. Read Willison's post for how a 27B local model handled the game guessing prompt and how long it took.
    4. Read Sean Goedecke's short post for the argument that a skilled researcher could do the same tasks.
    5. Self check: which Astra results would still stand if the model were forbidden from running any code?
    Try it30 min · 5 steps
    You need: A laptop or workstation with about 20GB free for a 17GB Q4_K_M GGUF, LM Studio, the llm CLI with the llm-lmstudio plugin, Python. No API key needed.

    Steps

    1. In LM Studio, download the Qwen 3.8 27B Q4_K_M build from the lmstudio-community repo and load it with a context window well above the 8192 default.
    2. Install the llm CLI and the llm-lmstudio plugin, then confirm the local model is listed.
    3. Fetch the Pac-Man prompt.txt from the exp1_which_game runs folder in the Astra paper repo, as linked from Willison's post.
    4. Run it with a long timeout, as Willison did with LMSTUDIO_TIMEOUT=900 and llm -f <prompt url> -m lmstudio/qwen/qwen3.8-27b.
    5. Record wall clock time, the guess and the stated confidence, and compare with Willison's 46 minutes and 72% for Pac-Man.

    Small angles to try

    • Try the Breakout prompt and compare against Willison's 52 minutes and 88% confidence.
    • Shuffle or relabel the input weights to see whether the guess relies on labels rather than the numbers.
    • Swap in a smaller local model and see where the answers stop being right.
    Sources:Honnibal: the Astra post is misleadingSimon Willison: Qwen 3.8 27B can interpret weightsSean Goedecke: LLMs are not superhuman at weights
  3. #3Models & launches

    Revolut shifts 2,000 engineers to self hosted GLM, The Register reports

    The Register reports that Revolut ops and infrastructure lead David Abitbol told the AI Engineer conference in New York that the fintech cut AI coding inference costs about 90% in under a year by making self hosted models the default. According to The Register, Revolut started with GLM-5.1 and now serves GLM-5.3 on 56 Nvidia B200 GPUs for roughly 2,000 engineers, with self hosted GLM 5.1 costing about $0.30 per million tokens, around a tenth of Codex or Claude Code. Revolut also claims a 60% improvement in coding time and 70% in lead time.

    Heat · HN "Revolut moves to self hosted GLM" 189 points and 160 comments on Nov 3
    What people foundAbitbol said the gains came from tuning rather than more GPUs: NVFP4 quantization, disaggregated prefill and decode, hyperparameter tuning and context length tuning, aiming for under 3 seconds to first token at around 256K context and about 200 tokens per second. He called long context "a scam" if used carelessly, and a router sends vision queries to Gemini or GPT models because GLM is text only.
    Learn it15 min · 4 steps

    Key ideas

    NVFP4
    A 4 bit floating point format on recent Nvidia GPUs that shrinks model weights so more requests fit on the same hardware.
    Disaggregated prefill and decode
    Running prompt processing and token generation on separate GPU pools so each can be sized and tuned for its own load.
    Time to first token
    How long a user waits before the first output token appears, which dominates how fast an interactive coding agent feels.

    Steps

    1. Read The Register article first and note every number: GPUs, cost per million tokens, latency, throughput and context size.
    2. Skim the HN thread for pushback on hardware ownership, depreciation and whether open models are good enough.
    3. Look up what disaggregated prefill and decode means in your preferred inference server's documentation.
    4. Self check: at $0.30 per million tokens, how many tokens a month would your team need before 56 B200 GPUs pay off?
    Try it30 min · 5 steps
    You need: A spreadsheet, your team's recent AI coding bill or token usage, optionally a local GPU with any open weight model. No API key needed for the cost model.

    Steps

    1. Export a month of token usage from your current coding assistant or estimate it from seats and typical daily use.
    2. Build a sheet comparing your current cost with the reported $0.30 per million tokens, adding a line for GPU rental or depreciation.
    3. If you have a local GPU, serve a small open model and measure time to first token at short and long prompts.
    4. Note at which prompt length time to first token passes 3 seconds on your setup.
    5. Write down which of your workloads need vision and would still need a commercial model through a router.

    Small angles to try

    • Rerun the latency test with and without a quantized build to see the speed difference.
    • Cap context length at a few sizes and check whether answer quality on your own repo actually drops.
    • Estimate off peak throughput, since Revolut reports much faster speeds after 5 pm.
    Sources:The RegisterHN discussion
  4. #4Research

    CUA-Master-27B: open computer use agent hits 80.2 on OSWorld Verified

    CUA-Master trains computer use agents with two nested loops: an outer loop that adds open world data aimed at the model's observed weaknesses, and an inner loop that mines workflows from instructional videos and turns them into verifiable tasks. Trajectories are gathered in real environments with a feedback guided snapshot rollback retry. The released 27B model, built on Qwen-3.5-27B, reports 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench and 25.0 on Agents' Last Exam.

    Heat · Hugging Face Papers, about 97 upvotes on Nov 3
    What people foundCode and BF16 weights are public, but the model card states no license and the README gives no serving command, so treat it as research code for now; the card itself advises running it in a sandbox because actions can modify files or run commands.
    Learn it15 min · 4 steps

    Key ideas

    Computer use agent
    A model that operates a desktop through screenshots plus mouse, keyboard or shell actions to finish tasks.
    OSWorld
    A benchmark of real desktop tasks in virtual machines where success is checked by scripts, not by a judge model.
    Snapshot rollback retry
    Saving the environment state before a step so a failed attempt can be undone and retried instead of restarting the whole task.

    Steps

    1. Read the abstract on the Hugging Face Papers page first, then the figure that shows the two loops.
    2. Open the GitHub README and map each pipeline stage to the loop it belongs to.
    3. Read the model card for the base model and the sandboxing advice.
    4. Self check: why does mining tasks from tutorial videos still need a verifier before the tasks are useful for training?
    Try it30 min · 5 steps
    You need: Python, a disposable VM or container for the desktop environment, your own model endpoint or API keys as the README's .env asks, and large GPU memory if you serve the 27B BF16 weights yourself.

    Steps

    1. Clone the CUA-Master repo into a fresh virtual environment.
    2. Install it with pip install -e ".[seeding,annotator,collection,dev]" as the README shows.
    3. Copy the example env file with cp .env.example .env and fill in only endpoints for sandboxes you own.
    4. Read the library snippet that uses CanonicalEpisodeRunner and OpenAICompatibleModel and point it at your own OpenAI compatible endpoint.
    5. Run one simple episode inside the disposable VM and review every action log before giving it more access.

    Small angles to try

    • Compare the agent with a general model on the same three tasks in your VM.
    • Read the task synthesis code and write one verifiable task of your own.
    • Check how the snapshot rollback logic decides when to retry.
    Sources:HF Papers 2611.03126Code on GitHubModel card
  5. #5Trends × tech

    Event and meeting transcripts × open Japanese ASR: Haas-ja-v1.0 trends

    Japanese AI company Third Intelligence released Haas-ja-v1.0, a 615.8M parameter speech recognition model for Japanese, with weights and inference code under Apache 2.0. The company says it is the most accurate open weight model it evaluated and that it beat the commercial Gemini 3.5 transcribe on the benchmark GIGAZINE cites. In GIGAZINE's own test, the free web demo transcribed a 4 minute 19 second stage clip 0.32 seconds after upload, though some names came out as similar sounding words.

    Heat · Number 9 in GIGAZINE's 24 hour popular ranking shown in its Nov 3 headline roundup
    What people foundIt is small enough for a single NVIDIA GPU and the company's cost chart ranks it cheapest at comparable accuracy, but GIGAZINE's test shows proper names still need a human pass. It runs on Linux with an NVIDIA GPU only; the model card says CPU and macOS are not supported.
    Learn it15 min · 4 steps

    Key ideas

    ASR
    Automatic speech recognition, turning recorded speech into text.
    Transducer model
    A speech model with an encoder, a prediction network and a joint network that emits text as it moves through audio, the layout this model card describes.
    Longform chunking
    Cutting long audio at quiet points into pieces the model was trained on, here up to 30 seconds, then joining the text.

    Steps

    1. Read the Hugging Face model card first for size, license and requirements.
    2. Read the GIGAZINE article to see how the demo handled a real stage recording and where it failed.
    3. Try the free longform demo linked from the article with a short clip of your own voice.
    4. Self check: why does the library need a separate longform function for audio longer than 30 seconds?
    Try it30 min · 5 steps
    You need: Linux with an NVIDIA GPU, Python 3.11 or newer, uv or pip, NeMo 2.7 as listed, and a WAV, FLAC, MP3 or OGG recording you own. No API key needed.

    Steps

    1. Clone the inference code with git clone https://github.com/third-intelligence/Haas-ASR.git and enter the folder.
    2. Install the locked dependencies with uv sync.
    3. Transcribe a short clip with uv run python examples/quickstart.py --model third-intelligence/Haas-ja-v1.0 example.wav, using your own file name.
    4. For a recording over 30 seconds, load it with haas_asr.load_audio and call haas_asr.transcribe_longform(model, audio, batch_size=64) in Python.
    5. Compare the text with what was actually said and count errors on names and technical terms.

    Small angles to try

    • Run the same meeting recording through a cloud transcription service and compare names and timing.
    • Try the speaker label example, which needs Python 3.12 and its own environment.
    • Change batch_size on a long file and measure GPU memory and speed.
    Sources:Hugging Face model cardGIGAZINE: Haas-ja-v1.0GIGAZINE headline roundup with popular ranking
  6. #6AI in the world

    Vesuvius Challenge: two ink models agree and read PHerc. 306B

    Liam Warner, a freelance software engineer, and Nicola Bodill, a classics student, won the $50K First Letters Prize for reading Greek letters inside PHerc. 306B, a Herculaneum scroll buried by Vesuvius in AD 79 and never unrolled. Their reading, τὸ δι᾽ αὑτὸ αἱρετόν or "that which is in itself desirable", came from multiplying the per pixel ink probabilities of the public TRACE v3B model and a custom fine tune called A-hn. The organisers reviewed 25 submissions for this scroll, and rival entries reached 7, 8 or 9 of the 10 readable characters the prize required.

    Heat · Hacker News, 128 points and 18 comments for the Scroll Prize post on Nov 2
    What people foundThe useful trick for any noisy segmentation job: a pixel counts as ink only where both models agree, and the fine tune added an extra loss penalty for marking blank papyrus as ink, which cut false positives enough to settle the disputed letters. The winning team published code, a Docker image, model weights and data, so the whole claim can be rerun.
    Learn it15 min · 5 steps

    Key ideas

    Ink detection
    A model looks at a flattened slice of a CT scan and predicts, for each pixel, how likely it is that carbon ink sits there.
    Model product ensemble
    Multiplying two models' probabilities keeps a pixel bright only when both are confident, which trades some recall for far fewer false positives.
    Pseudo labels
    Labels made by averaging the predictions of existing models, used to train a new model where human labels are scarce.

    Steps

    1. Read the Scroll Prize post "Writing seen in PHerc. 306B" first, focusing on the method section about segmentation and ink detection.
    2. Open the TRACE write up on GitHub to see how the base ink model was fine tuned and how it was scored by AUC on a held out scroll.
    3. Skim the README of the winning repo SirProdigle/pherc0306b-first-letters to see how the claim image, mesh and checks are packaged.
    4. Note how the A-hn training data mixed half human labels and half machine labels averaged from four public models.
    5. Self check: why does multiplying two probability maps reduce false positives more than averaging them would?
    Try it30 min · 5 steps
    You need: Python 3 with NumPy and Matplotlib on your own laptop; Docker only for the optional rerun. No API key needed.

    Steps

    1. Create a small synthetic image yourself: a few letter shapes drawn as ink on a noisy background.
    2. Make two fake model outputs by adding different random noise and a few different false bright spots to the true ink mask, then clip both to the 0 to 1 range.
    3. Compute the average and the elementwise product of the two maps and threshold each at 0.5.
    4. Count true positives and false positives for model A, model B, the average and the product, and plot all four side by side.
    5. Optional: read the winning repo README and its docs, and if you have the disk space, follow its documented Docker commands to redraw the claim image on your own machine.

    Small angles to try

    • Raise the product threshold and watch where letters start to break apart.
    • Make the two fake models share the same false spots and see how the product loses its advantage when errors are correlated.
    Sources:Scroll Prize: Writing seen in PHerc. 306BWinning submission codeTRACE ink detection write up
  7. #7Tools & open source

    Craft Anything trends: a Rust harness for Claude Code app rebuilds

    Craft Anything, an open source Rust project created on Oct 31, studies a website, app or game from the outside and then drives your own signed in claude program to build a new version and compare it with the original. Its default clean room mode reads nothing inside the original, not its code, files or page source, and works only from what a visitor sees, pictures you supply or a window screenshot you allow. A deep mode lifts that limit for private use and can call engines such as Ghidra.

    Heat · Trending on GitHub this week, about 5.1k stars and 835 forks at the time of checking
    What people foundClean room here means rebuilding from observed behavior, not copying internals, but it does not make every rebuild legal: the README says to use it only on software you own or are allowed to examine, it never bypasses DRM, logins or anti cheat, and the agent is told not to copy logos, photos or text without permission. The project calls itself early alpha; Claude Code is the main builder, the Codex adapter has not been verified on a real install, and of the other ACP agents only Maki has been run live.
    Learn it15 min · 5 steps

    Key ideas

    Clean room rebuild
    Recreating software only from its visible behavior and outputs, without reading its source or files, which lowers but does not remove copyright and licence risk.
    Agent harness
    A program that feeds a coding agent the task, runs its output and checks the result, round after round.
    MCP tool
    A tool exposed through the Model Context Protocol so any compatible agent can call it.

    Steps

    1. Read the storytold/craft-anything README first, especially the sections on clean room mode, deep mode and the legal notes.
    2. List which agents are verified and which are only partly working, so expectations match an early alpha.
    3. Look at the CLI examples to see how a rebuild is started from a URL or from a plain description.
    4. Decide for one app you use whether you actually have the rights to study and rebuild it.
    5. Self check: what does clean room mode refuse to do, and why does that still not give permission to rebuild someone else's product?
    Try it30 min · 5 steps
    You need: Rust 1.97 or newer, Git, a signed in Claude Code install on your own machine; Linux desktop builds need the extra GTK packages listed in the README. No separate API key needed when Claude Code is already signed in.

    Steps

    1. Clone the repo and install the CLI with cargo install --path apps/craft-anything-cli from the repo root.
    2. Run craft-anything-cli doctor and fix anything it reports about the install, coding agents or engines.
    3. Start with a target that needs no rights at all: craft-anything-cli rebuild - "A snake game in Rust" --rounds 1.
    4. Then point a rebuild at a site you own, or a locally served open source web app whose licence allows it, and stay in the default clean room mode.
    5. Compare the rebuilt version with the original side by side and note which parts the agent got right from visible behavior alone.

    Small angles to try

    • Run craft-anything-cli mcp and call the same tools from another agent you already use.
    • Rebuild your own small site twice with different prompts and compare the generated code quality.
    Sources:GitHub: storytold/craft-anything
  8. #8Tools & open source

    stack-diagram: an agent skill that draws your stack with real logos

    stack-diagram is an MIT licensed agent skill: describe your stack in a sentence and the agent draws an architecture diagram using official logos fetched from svgl, devicon or gilbarbara/logos, never recolored or redrawn. The default output is a self contained HTML file, with PNG, a dependency free React component or a Claude Artifact as options. It works with Claude Code, Codex, Cursor and other agents that support skills.

    Heat · A new repo created on Nov 1 that drew attention among agent skill users; star counts differed widely between checks, so no figure is given
    What people foundIt installs with one command and needs only Python 3.9 or newer plus network access for logos, with Playwright and Chromium added only for PNG export. The README's example prompts show the right level of detail: name each service and what connects to what, and the skill handles the layout.
    Learn it15 min · 4 steps

    Key ideas

    Agent skill
    A packaged folder of instructions and scripts that a coding agent loads when a task matches it.
    Logo sources
    Public icon sets such as svgl and devicon that host official brand marks as SVG files.

    Steps

    1. Read the caru-ini/stack-diagram README first, including the example prompts and output formats.
    2. Look at agentskills.io, linked from the README, to see how skills are discovered by different agents.
    3. Note which steps need the network, such as downloading logos, and which run locally.
    4. Self check: why does the skill fetch logos at build time instead of bundling them?
    Try it30 min · 5 steps
    You need: Node for npx, Python 3.9 or newer, a coding agent that supports skills such as Claude Code; Playwright with Chromium only for PNG. No extra API key needed.

    Steps

    1. Install the skill with npx skills add caru-ini/stack-diagram.
    2. In your agent, describe the stack of one of your own projects in one or two sentences, naming each service and how they connect.
    3. Open the generated HTML file and check both the light and dark theme.
    4. Fix one wrong connection by asking the agent to change it, then regenerate.
    5. Commit the diagram next to your project README and compare it with any older hand drawn diagram.

    Small angles to try

    • Ask for the React component output and drop it into your docs site.
    • Install Playwright with Chromium and export a PNG for slides.
    Sources:GitHub: caru-ini/stack-diagram
  9. #9Tools & open source

    Cicada: open source Mac app for live captions and AI meeting notes

    Cicada, or 知了, is an Apache 2.0 macOS app that captures system audio, so it gives live bilingual subtitles, meeting notes and interpreting for any call or stream without adding a bot to the meeting. Speech recognition, language detection and recording run on the Mac, while translation goes to Google Translate and notes, translation and speech go to Ling and Ming models when you add a Ling API key. Holding the right Option key lets you speak and have translated speech played into the call.

    Heat · A new repo created on Nov 1, with 83 stars on GitHub at the time of checking
    What people foundRead the network list before using it at work: the README says recognized text goes to Google Translate, to Ling when a key is set, voice clips go out only when interpreting with a key, the BlackHole virtual microphone is downloaded from its official source when you enable interpreting, and the app checks GitHub for updates daily. There is no analytics or telemetry, and the notes window is hidden from screen sharing by default.
    Learn it15 min · 4 steps

    Key ideas

    System audio capture
    Recording what the Mac plays, such as the other side of a call, instead of joining the call as a separate participant.
    Virtual microphone
    A software audio device, here BlackHole, that lets an app feed generated speech into a call as if it came from a mic.
    On device speech recognition
    Turning speech into text with Apple's local speech APIs so raw audio does not leave the machine.

    Steps

    1. Read the wquguru/cicada README first, especially the privacy and network sections.
    2. List every outbound destination the README names and what data each one receives.
    3. Look at the install steps to see why the first launch is blocked, since the app is signed locally but not notarized.
    4. Self check: in a call with no Ling key set, which data can still leave the Mac?
    Try it30 min · 5 steps
    You need: An Apple silicon Mac on macOS 26 or later; a Ling API key only for notes, interpreting and voice features. No API key needed for subtitles with the default translation.

    Steps

    1. Download the DMG from the repo's Releases page and verify it with shasum -a 256 Cicada-*.dmg against checksums.txt.
    2. Drag the app into Applications, allow it under Privacy and Security, and grant audio capture during onboarding.
    3. Play the sample meeting from onboarding, or a public talk video, and watch the live subtitles.
    4. Export the session to Markdown or SRT and check the transcript for errors on technical terms.
    5. Optional: watch outbound connections with your own firewall tool to confirm they match the README's list.

    Small angles to try

    • Add a Ling key and compare its notes and translation with the Google Translate fallback.
    • Remap push to talk to a Control or Command combination so Input Monitoring permission is not needed.
    Sources:GitHub: wquguru/cicada
  10. #10Tools & open source

    Agent Autoloop: background Claude Code and Codex jobs, approval cards

    Agent Autoloop is a small self hosted hub that routes events from cron, GitHub webhooks, HTTP or issue polling to a command line agent through a rules.json file; the agent runs one round and exits. When the agent is unsure, such as before an action you cannot undo, it asks a question that arrives as a desktop notification or Feishu card, and the same session resumes after you answer. Built in, tested support covers Claude Code, Codex and ACP, and it is MIT licensed.

    Heat · A very new repo created on Nov 3 with only a handful of stars at the time of checking; included for the design, not for popularity
    What people foundWhile waiting for your answer no agent process runs and no tokens are spent; only the lightweight hub keeps running and the wait state sits in its database. The kernel is about 1,650 lines with zero runtime dependencies, state comes from an append only event log, and chained agent events stop at 8 levels, after which the hub only records the event and sends you a card.
    Learn it15 min · 4 steps

    Key ideas

    Event driven agent
    An agent that starts only when a trigger such as a cron tick or webhook fires, does one round of work and exits.
    Append only event log
    A record that is only ever added to, so current state can be rebuilt by replaying it after a crash.
    Loop depth cap
    A limit on how many times agent outputs can trigger further agent work, which stops runaway chains.

    Steps

    1. Read the KKKKhazix/agent-autoloop README first, focusing on the architecture, rules.json and the depth cap.
    2. Compare which agents are built in and tested against the ones listed only as untested command examples.
    3. Look at how a question to the human ends the agent process and how the session is resumed later.
    4. Self check: what is still running on your machine while a question waits three days for an answer?
    Try it30 min · 5 steps
    You need: macOS or Linux, Node 24, Git, curl and a command line agent you are already signed in to, such as Claude Code or Codex. No GitHub account needed.

    Steps

    1. Read the install script in the repo first, then install with the README's one line curl command on your own machine.
    2. Run agent-autoloop try to see the built in demo.
    3. Edit ~/.agent-autoloop/rules.json to add one cron rule that asks your agent to summarise a notes folder you own, and validate it with agent-autoloop check.
    4. Let the rule fire, answer any question card that appears, and confirm the session resumes.
    5. Run agent-autoloop why 1 to see why that round ran and what it produced.

    Small angles to try

    • Point a GitHub webhook from your own test repo at the hub and trigger a round on a new issue.
    • Write a rule that makes the agent emit events back to itself and watch the depth cap stop it.
    Sources:GitHub: KKKKhazix/agent-autoloop
  11. #11Tools & open source

    Kairos: open source fraud decisioning that coding agents tune over MCP

    Finic released Kairos under Apache 2.0, a real time decision service for sign ups, logins, orders, payments and payouts that returns approve, step up, review or block with a risk score. It pairs classic models with Finic's Grayson decision models, which take plain language questions such as whether an email looks disposable. An MCP server lets coding agents read decisions, build workflows and run experiments that train and compare model variants on your own data.

    Heat · Show HN on November 3; the vote count differed between HN pages checked on November 4, so no single figure is given. The GitHub repo is new and shows only a handful of stars.
    What people foundThe interesting design is the agent loop: Claude Code, Cursor or Codex proposes variants, Kairos trains and tests them, and the agent ships the best one, with every action logged under the key it used. HN commenters mainly flagged the name clash with several other projects called Kairos.
    Learn it15 min · 4 steps

    Key ideas

    Risk decisioning
    A service that scores each user event and returns an action such as approve, review or block.
    Step up
    Asking for extra verification, such as a one time code, instead of blocking outright.
    MCP server
    A Model Context Protocol endpoint that exposes a product's actions as tools a coding agent can call.

    Steps

    1. Read the Kairos README on GitHub first, focusing on the decision outcomes and the MCP section.
    2. Read the Show HN post to see the two problems the authors target: scarce labeled data and fraudsters who adapt.
    3. Compare a rules engine, a gradient boosted model and a plain language question model, and note what each needs as input.
    4. Self check: why would an agent that proposes model variants still need a held out test on your own data before shipping?
    Try it30 min · 6 steps
    You need: Node.js with pnpm, or Docker; a Grayson API key from portal.finic.ai for self hosting; synthetic test events only.

    Steps

    1. Clone the repo with git clone https://github.com/finic-ai/kairos and run pnpm install in it.
    2. Run cp .env.example .env and set GRAYSON_API_KEY in the new file.
    3. Start it with pnpm kairos dev; the README says the API runs on port 8787 and the console on port 5273.
    4. Send a few made up sign up events with obviously fake and normal emails, and compare the decisions and reasons.
    5. In the console under Settings, enable MCP, create a key, and connect your coding agent following the README.
    6. Ask the agent to propose one model variant and check what experiment Kairos records.

    Small angles to try

    • Use docker compose up --build instead of pnpm for a team style setup.
    • Restrict the MCP key to read only tool groups and confirm the agent cannot change workflows.
    • Feed identical events twice and check whether the decision is stable.
    Sources:GitHub: finic-ai/kairosShow HN: Kairos
  12. #12Engineering & CS

    Carlini: LLMs port Stockfish to a 1996 Pentium Pro, rated above Kasparov

    Nicholas Carlini ran chess engines on a 200 MHz Pentium Pro with 64 MB of RAM, booted from a floppy. A Stockfish 17 port, retuned for that CPU, rated about 2846 at one second per move and an estimated 3300 at classical time controls, above Kasparov's rating of just under 2800. Against original 1996 Crafty on modern hardware it scored 78 wins, 17 draws and 5 losses over 100 games.

    Heat · HN front page, 467 points and 230 comments in the HN search index on November 4
    What people foundCarlini says the code was written by LLMs, naming Fable 5.1, Opus 5.5 and GPT-6, while the essay text is his own. The port alone was not enough: retuning for the old CPU lifted Stockfish from about 2496 to 2846 at one second per move, and compiler and alignment changes alone added roughly 27 more.
    Learn it15 min · 4 steps

    Key ideas

    Elo rating
    A number where a 400 point gap means the stronger side is expected to score about 10 to 1.
    NNUE
    A small neural network evaluation designed to run fast on a CPU inside a chess search.
    SPSA tuning
    A method that nudges many engine parameters at once using match results to find better values.

    Steps

    1. Read Carlini's essay first, especially the tables of ratings per engine and time control.
    2. Note which gains came from hardware and which from software; he credits about 720 of 1335 Elo to software.
    3. Look at the list of tuning steps and see how each later step gave a smaller gain.
    4. Self check: why do wide confidence intervals matter when comparing two ports that differ by 30 Elo?
    Try it30 min · 5 steps
    You need: A modern computer, Stockfish from its official site, and any chess GUI or match runner that plays engine matches. No API key needed.

    Steps

    1. Install Stockfish and a chess GUI that can run engine against engine matches.
    2. Set up two copies of Stockfish, one with one thread and a tiny hash table to mimic old hardware, one with default settings.
    3. Play a short match of 20 to 40 games at a fixed time per move.
    4. Record wins, draws and losses and estimate the rating gap from the score.
    5. Ask a coding agent to explain which Stockfish options most affect strength on a weak CPU, then check its claims in the Stockfish docs.

    Small angles to try

    • Repeat at a longer time per move and see whether the gap shrinks.
    • Have a coding agent write a tiny engine in a language of your choice and play it against limited Stockfish.
    Sources:Carlini: algorithmic improvements in chessHN discussion
  13. #13Research

    CWE-bench v2: Gemini 4 Argon, Claude Opus 5.5 near 68% on security fixes

    Collinear AI published CWE-bench v2 on the Hugging Face blog on November 3: 100 tasks over 91 weakness types where an agent must fix a security flaw in a real repository without breaking it. Scoring is programmatic and deterministic. Gemini 4 Argon leads at 67.8% pass@1 for $3.92 per rollout, Claude Opus 5.5 scores 67.3% for $1.83, and DeepSeek V4.1 Flash reaches 57.0% for $0.04.

    Heat · Low HN attention; published as a Hugging Face blog post on November 3 with a gated Apache 2.0 dataset of 100 tasks and results for 11 models.
    What people foundCost is the practical lever: about 10 points of accuracy separate the leader from a model nearly 100 times cheaper per rollout. With four attempts the leader reaches 81.0%, and 95 of 100 tasks were solved by at least one model.
    Learn it15 min · 4 steps

    Key ideas

    CWE
    Common Weakness Enumeration, a catalog of software weakness types such as integer underflow.
    pass@1
    The share of tasks solved on a single attempt.
    Harbor
    An open source framework from the Terminal-Bench team for running agents on containerized benchmark tasks.

    Steps

    1. Read the CWE-bench Hugging Face blog post first, including the cost per rollout chart.
    2. Open the dataset card and read the task format and the running section.
    3. Pick one task and read its description and tests before looking at any fix.
    4. Self check: why does a fix count only if the original functionality tests still pass?
    Try it30 min · 5 steps
    You need: Docker, Python, a Hugging Face account that has accepted the dataset conditions, and an API key for the model your agent uses. Runs only inside local containers.

    Steps

    1. Install Harbor with uv tool install harbor or pip install harbor.
    2. Accept the gated dataset terms on the Hub, then run hf auth login --token $HF_TOKEN --add-to-git-credential.
    3. Download the dataset locally and pick one task folder.
    4. Run that single task with harbor run --path plus the task folder, your agent, your model, --env docker and the model API host, as shown on the dataset card.
    5. Read the agent's patch and the test output to see whether the weakness was fixed without breaking tests.

    Small angles to try

    • Run the same task with a cheap model and a frontier model and compare cost and result.
    • Try a task in a language you know well and judge the patch quality yourself.
    Sources:Hugging Face blog: CWE-benchDataset: collinear-ai/cwe-bench
  14. #14Research

    Ai2 StateBench: frontier models fall to 13% tracking swapped variables

    Ai2 released StateBench on November 3, a benchmark with four levels of state tracking, from fixed assignments up to assignment by reference and swaps. Gemini 3.8 Flash, Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna handled the easier levels but fell to 13% or lower on the longest problems at the hardest level. In 275M parameter models trained from scratch, a Gated DeltaNet variant stayed near perfect on the hardest level to about 10 times its training length, while transformers never learned it.

    Heat · Low HN attention; published as a Hugging Face blog post by Ai2 on November 3 with an open Apache 2.0 code repo and a dataset named allenai/state-bench.
    What people foundFor agents that must follow values which change and swap across a long log or codebase, Ai2's results suggest today's frontier models are unreliable once the chain gets long, so keep explicit state outside the model.
    Learn it15 min · 4 steps

    Key ideas

    State tracking
    Keeping the current value of each variable correct as a long sequence of updates arrives.
    Assignment by reference
    Updates like x = y or swaps, where a value depends on another variable's current value.
    Gated DeltaNet
    A recurrent style layer that updates a fixed size memory, used here as an alternative to attention.

    Steps

    1. Read Ai2's StateBench post on the Hugging Face blog first, focusing on the four levels.
    2. Look at the frontier model chart and find where accuracy drops with length.
    3. Read the small model study and note the 256 assignment training limit and 64,000 assignment test limit.
    4. Self check: why is x = x + 5 easier to track than a swap of x and y?
    Try it30 min · 5 steps
    You need: Python and the uv tool; no GPU needed to build data. No API key needed, unless you also query a hosted model.

    Steps

    1. Clone the repo with git clone https://github.com/allenai/state-bench and run uv sync.
    2. Validate the default config with uv run state-bench config check configs/experiments/default.yaml.
    3. Generate data with uv run state-bench build configs/experiments/default.yaml.
    4. Run uv run pytest to confirm the setup works.
    5. Take a few generated problems at each level and paste them into a model you use, then score its answers by hand.

    Small angles to try

    • Lengthen the sequences step by step and find where your model starts failing.
    • Give the model a scratchpad and ask it to write the full state after every update.
    Sources:Hugging Face blog: StateBench by Ai2GitHub: allenai/state-bench
  15. #15Trends × tech

    School AI bans × Google Search: AI off by default, but Search goes too

    Google said on November 3 that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be off by default for K-12 student accounts until administrators turn them on. The New York Times reports that schools blocking AI entirely will lose student Search, since AI Overviews come with it, and that the changes start in early 2027. Google also made Guided Learning in Classroom available worldwide, with a Search pilot in select US schools later this month.

    Heat · HN front page, 612 points and 430 comments on the NYT story in the HN search index, plus a second submission at 45 points; covered by ITmedia and GIGAZINE on November 4.
    What people foundGoogle education VP Chris Phillips told the NYT that Search and AI are deeply integrated. For engineers the lesson is about product design: once an AI layer cannot be switched off inside a core feature, regulated customers who need an AI free mode lose the whole feature.
    Learn it15 min · 4 steps

    Key ideas

    Off by default
    A feature stays disabled until an administrator deliberately turns it on.
    AI Overviews
    AI written summaries that appear at the top of Google Search results.
    Guided Learning
    A Gemini mode for students that gives hints and steps instead of finished answers.

    Steps

    1. Read Google's November 3 blog post by Chris Phillips first.
    2. Read the NYT story via the Philadelphia Inquirer to see the Search trade off and the August reversal.
    3. Skim the ITmedia article for the reaction and Google's response.
    4. Self check: what choices does a school that bans AI have for student search after early 2027?
    Try it30 min · 5 steps
    You need: Any chat model you already use; an API key only if you call a hosted model from code.

    Steps

    1. Write a system prompt that tells a model to act as a tutor that gives hints, never final answers.
    2. Pick five homework style questions from your own field, such as small coding exercises.
    3. Ask each question once with the tutor prompt and once without it.
    4. Try to talk the tutor into giving the full answer and note how often it gives in.
    5. Write down what a teacher review log would need to capture to spot misuse.

    Small angles to try

    • Design a feature flag scheme where an AI layer can be disabled without removing the core feature.
    • Compare how your own product would behave for a customer that requires an AI free mode.
    Sources:Google blog: Designing responsible AI for learningNYT via Philadelphia InquirerITmedia NEWS
  16. #16Engineering & CS

    GitBook and WorkOS breaches: vendors point to AI agent tooling

    GitBook's public status page says unauthorized access to parts of its infrastructure was detected on Nov 2, some services were taken offline as a precaution, and users do not need to do anything right now. A GitBook support email quoted in an HN post goes further: it calls the incident a highly targeted attack using AI agents, says Git Sync secrets, API tokens, third party tokens and GitBook Assistant questions may have been exposed, and says all customer API tokens were invalidated. Separately, a WorkOS bulletin dated Nov 2 says an "AI-assisted offensive security framework" automated recon and exploitation against an internal sales app, exposing CRM and support data, with no evidence that WorkOS application data was accessed.

    Heat · Both stories were discussed on Hacker News on Nov 2 and 3; point counts differed between the sources checked today, so no number is given
    What people foundPer the support email quoted on HN, GitBook asks customers to rotate GitBook tokens and connected third party secrets and to authenticate the GitBook MCP server again; the status page itself does not mention AI agents or these steps. WorkOS warns that stolen contact and conversation data can make phishing look legitimate.
    Learn it15 min · 5 steps

    Key ideas

    Token rotation
    Issuing a new secret and revoking the old one so a leaked copy stops working.
    MCP server auth
    The credential an AI client such as an editor agent uses to call a vendor's Model Context Protocol server on your behalf.
    AI assisted intrusion
    An attack where agent tooling automates reconnaissance and exploitation attempts that a human once ran by hand.

    Steps

    1. Read the GitBook status page first and note exactly what it does and does not claim.
    2. Read the support email quoted in the HN post and list each secret type it says may be exposed.
    3. Read the WorkOS bulletin and separate the data it says was accessed from what it says was not.
    4. Map which of those secret types exist in your own repos, CI and agent configs.
    5. Self check: which claims come from an official public page and which only from a quoted email?
    Try it30 min · 5 steps
    You need: Your own accounts and repos, a text editor, an optional local secret scanner. No API key needed

    Steps

    1. List every vendor token your team holds for docs, auth and sync tools, with owner and creation date.
    2. Search your own repositories and CI variables for any of those tokens stored in plain text.
    3. Open each AI agent or MCP client config on your machine and list which servers it authenticates to and with what credential.
    4. For any token tied to a breached vendor, rotate it in that vendor's own dashboard and update the secret store, then confirm the old value no longer works against your own integration.
    5. Write a short runbook entry recording who rotates what, so the next vendor incident is a drill.

    Small angles to try

    • Check whether any agent config grants broader scopes than the task needs and narrow them.
    • Add a calendar reminder to review token age every quarter.
    • Brief your team on phishing that cites real past support threads, as WorkOS warns.
    Sources:GitBook status: security incidentHN post quoting the GitBook support emailWorkOS security bulletin
  17. #17Trends × tech

    Siri AI in Japan × usage caps: Apple meters cloud AI after record quarter

    In Apple's first earnings report under CEO John Ternus, he described Siri AI as the Siri people had always wanted, per ITmedia. Siri AI launched as a US English beta in September and added Japanese, French, Korean, Portuguese and Spanish in October. Siri handles more than 2.5 billion requests a day, and server side Siri AI features and intelligent photo editing come with usage limits.

    Heat · No. 8 on the ITmedia NEWS access ranking on 2026-11-04, the only Apple story in a top 10 led by GMO breach coverage
    What people foundApple guides Oct to Dec gross margin to 45% to 46%, citing memory costs and the cost of expanding Siri AI, and says heavy users may be able to buy expanded access later. Even at Apple scale, cloud inference for consumers is being metered.
    Learn it15 min · 5 steps

    Key ideas

    Server side inference
    Running a model in the data center instead of on the phone, which costs money per request.
    Usage cap
    A limit on how many expensive requests one user can make before a feature stops or asks for payment.
    Gross margin
    Revenue minus the direct cost of what was sold, as a share of revenue.

    Steps

    1. Read the ITmedia NEWS article on Apple's results and Siri AI first.
    2. Note which Siri AI features run on servers and which are capped.
    3. Compare the 2.5 billion daily requests with how often one person might use an assistant.
    4. Think through how a cap changes product design for a feature you build.
    5. Self check: why would a company with on device chips still meter some AI features?
    Try it30 min · 5 steps
    You need: A spreadsheet or Python, your own app's usage logs or made up sample data. No API key needed

    Steps

    1. Pick one AI feature in your own product or a hobby project and estimate requests per active user per day.
    2. Write down a cost per request from your provider's public price page.
    3. Model monthly cost for light, median and heavy users in a spreadsheet.
    4. Choose a cap that keeps the heavy tier within budget and note how many users would hit it.
    5. Draft the message a user sees when they hit the cap.

    Small angles to try

    • Compare a hard cap with a slower fallback model after the limit.
    • Estimate how much moving one step on device would cut cost.
    • Design a paid expanded tier like the one Apple hints at.
    Sources:ITmedia NEWS: Apple results and Siri AIITmedia NEWS access ranking
  18. #18Research

    Harness evolution erases a coding agent's skills: SWE-Bench 94% to 32%

    A new paper by Shangziqi Zhao, Yitong Zhang and Jia Li asks whether coding agents can keep evolving their harness, the scaffolding around the model, across software engineering tasks without catastrophic forgetting. It arranges three task datasets into six evolution directions and tests two harness self evolution methods. With DeepSeek-V4-Flash, Meta-Harness reached 94.0% on SWE-Bench, but after further evolution on FeatureBench that score dropped to 32.0% by the fifth round, a 66% relative fall.

    Heat · Listed on Hugging Face Daily Papers on Nov 3 with public code; no upvote count was visible today
    What people foundThe abstract says evolving on a later dataset "can substantially undermine" earlier tasks. If you auto tune prompts, tools or agent rules on new work, rerun your old evals before keeping the change.
    Learn it15 min · 5 steps

    Key ideas

    Agent harness
    The prompts, tools, rules and control loop wrapped around a model, separate from its weights.
    Harness self evolution
    Letting an automated process rewrite that harness based on results on a task set.
    Catastrophic forgetting
    Losing performance on earlier tasks after adapting to new ones.

    Steps

    1. Read the abstract on the Hugging Face paper page first.
    2. Skim the GitHub README to see the three datasets and the six transfer directions.
    3. Find the SWE-Bench to FeatureBench result and note the round where the drop appears.
    4. Compare this with how you change your own agent prompts or rules files.
    5. Self check: why can tuning the harness, not the weights, still cause forgetting?
    Try it30 min · 5 steps
    You need: Linux, Python 3.12 and 3.13, uv, git; the dry run needs no API key, while full runs need a DeepSeek or other OpenAI compatible key and an E2B account

    Steps

    1. Clone the repository on your own machine and read the README install section.
    2. Follow the README to create the Harbor environment with uv venv and the pinned lock file using --no-deps.
    3. Run the README's sanity check bash agentic-harness-engineering/scripts/evaluate_mini_swe_workspace_harbor.sh --dry-run, which makes no model or E2B calls.
    4. Open the configs for one transfer direction and note which dataset is evolved first and which is evaluated after.
    5. Apply the idea to your own setup: save your current agent rules, change them for one new task type, then rerun a few old tasks and compare.

    Small angles to try

    • Keep a small frozen regression set of old tasks and run it after every harness change.
    • Compare forgetting when only the prompt changes versus when tools change.
    • With a budget, run one real transfer config end to end.
    Sources:Hugging Face paper page 2611.03219Code on GitHub
  19. #19Trends × tech

    Smart glasses × Gemini: XREAL AURA Android XR opens preorders in Japan

    XREAL opened Japanese preorders on Nov 2 for AURA, its first Android XR glasses, developed with Google and Qualcomm, with shipping from late November. Prices are 229,800 yen for 12GB and 256GB and 269,800 yen for 16GB and 512GB. Gemini can recognize what the wearer sees in the real world or on screen, answer questions, help operate apps and highlight specific spots in a shared view.

    Heat · Covered by Impress Watch on Nov 2; free demo events by online lottery run in Fukuoka on Nov 14, Tokyo on Dec 1 to 2 and Osaka on Dec 19
    What people foundThe glasses stay under 95g because the Snapdragon chip sits in a separate compute pack of about 265g, a trade that keeps a camera fed assistant wearable.
    Learn it15 min · 4 steps

    Key ideas

    Android XR
    Google's Android based platform for headsets and glasses.
    Tethered compute pack
    A separate box with the processor and battery, linked by cable, so the glasses stay light.
    Multimodal assistant
    A model that takes camera images and voice together to answer about what the user sees.

    Steps

    1. Read the Impress Watch article first and list the specs and Gemini features.
    2. Note the split of weight between glasses and compute pack.
    3. Think about which tasks in your own app a user could do by looking and speaking.
    4. Self check: what does a camera fed assistant need from an app that a touch screen does not?
    Try it30 min · 5 steps
    You need: A computer, optionally your own phone camera and any multimodal model you already use. No new hardware needed

    Steps

    1. Pick one screen of your own app and write the three most common things a user does there.
    2. Describe each as a spoken request a glasses user would make while looking at the screen.
    3. Take a screenshot of that screen and ask a multimodal model you already use to find the control for each request.
    4. Note where the model fails because labels are missing or ambiguous.
    5. Fix one label or accessibility description in your app and repeat the check.

    Small angles to try

    • Try the same with a photo of a real world object linked to your product.
    • Measure how many steps a voice flow saves versus touch.
    • List privacy notices needed when a camera sees your app's data.
    Sources:Impress Watch: XREAL AURA in Japan
  20. #20Tools & open source

    One user's report: Codex training opt out switched itself back on

    One Hacker News user says their company turned off data sharing in Codex, Claude Code and similar tools to protect IP, then found the Codex setting turned back on without anyone opting back in. The post points to the "Improve the model for everyone" option in ChatGPT data controls, links an earlier HN thread on the same issue, and mentions a second opt out on OpenAI's privacy portal. This rests on that single user's observation, and no OpenAI response was seen.

    Heat · Hacker News post from Nov 3; point counts differed between the sources checked today, so no number is given
    What people foundThe poster suggests also opting out through OpenAI's privacy portal, not only the in product toggle. Teams with IP rules should treat any training toggle as something to recheck, not set once.
    Learn it15 min · 4 steps

    Key ideas

    Training opt out
    A setting that tells a provider not to use your prompts and outputs to train future models.
    Account level vs workspace level controls
    Some data settings apply per user and some per organization, so one may not cover the other.

    Steps

    1. Read the HN post first and note exactly what the poster observed.
    2. Open your own ChatGPT data controls and find the option named in the post.
    3. Check your organization's agreement or admin console for how training data is handled.
    4. Self check: what evidence would show a setting changed by itself rather than by a teammate?
    Try it30 min · 5 steps
    You need: Your own OpenAI account and any coding agent accounts you use. No API key needed

    Steps

    1. Open the data controls in your own OpenAI account and record the current state of the model improvement option with a dated screenshot.
    2. Do the same for every other coding agent you use, in its own settings.
    3. If your company uses a business or enterprise plan, ask the admin how training is controlled at the organization level.
    4. Set a weekly reminder to recheck the screenshots against the live settings for a month.
    5. If a setting changes without anyone touching it, report it to the vendor's support with your dated records.

    Small angles to try

    • Keep a shared team checklist of every AI tool and its training setting.
    • Compare personal accounts with organization managed accounts.
    • Check whether signing in on a new device resets any setting.
    Sources:HN: report on Codex training setting

Get the Radar by email

One email each evening, Japan time, with that day's items. Opening soon: leave your email and we'll send a confirm link when it starts.