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.

Thursday, November 12, 2026

20 items · AWS ships an agent built local AWS emulator; Gemini gets a Decisions API
  1. #1Tools & open source

    AWS Emulator: an official local AWS clone written and tested by agents

    AWS released aws-emulator, a Rust emulator of the AWS APIs for testing on a laptop, a dev server or in CI with no AWS account and no calls to AWS. Werner Vogels writes that it covers more than 200 services, is built from more than 400 Smithy API models, and was written and tested entirely by agents behind a merge gate that includes adversarial security review agents. The project is CC0, while its vendor directory is under Apache 2.0.

    Heat · HN front page, about 830 points and 233 comments on Nov 11; 2.8k GitHub stars and 62 forks
    What people foundCoding agents can now get a free, offline AWS to test against: the README offers an agent skill install with aws agent-toolkit add-skill --skill-name aws-emulator --region us-east-1, which needs AWS CLI 2.35.0 or later. Simon Willison welcomes it but notes it currently needs Docker on macOS and Windows, and hopes for a single downloadable binary.
    Learn it15 min · 5 steps

    Key ideas

    Smithy models
    Smithy is the format AWS uses to describe each service's operations, inputs, outputs and errors, and the emulator was generated against those public models.
    Fidelity tests
    Tests run against real AWS first, so the emulator is judged by whether it behaves the same way the real service does.
    Merge gate
    A set of automated checks every change must pass before it enters the codebase, here including agents that review code adversarially for security.

    Steps

    1. Read Werner Vogels' All Things Distributed post first for how agents wrote and tested the code.
    2. Skim the aws/aws-emulator README for the run modes: Linux packages, a container on macOS and Windows, embedded in Rust tests, or built from source.
    3. Note the limits: it is for building and testing, not production or security testing, and per service fidelity scores sit on the releases page.
    4. Read Simon Willison's short note for the Docker caveat.
    5. Self check: why does correcting fidelity tests against real AWS matter more than the code itself when agents write the emulator?
    Try it30 min · 6 steps
    You need: Linux, or Docker on macOS or Windows; a Rust toolchain with Cargo and a C compiler to build from source; AWS CLI v2 for client calls. No AWS account or API key needed.

    Steps

    1. Clone the aws/aws-emulator repository and build it with cargo build -p aws-emulator-router.
    2. Start it with cargo run -p aws-emulator-router; per the README it serves HTTPS on 127.0.0.1:9297 with a self signed certificate.
    3. Turn telemetry off if wanted by setting EMULATOR_TELEMETRY=off before starting, or with aws-emulator --telemetry off for an installed binary.
    4. Follow the repo's guide on pointing an AWS SDK at the emulator, set the endpoint to the local address, and handle the self signed certificate as the guide describes.
    5. Create a bucket, upload a file and list it, then run an existing test suite of your own against the local endpoint.
    6. Check the releases page fidelity score for each service your code uses and note any gaps you hit.

    Small angles to try

    • Install the agent skill and ask a coding agent to write integration tests against the emulator only.
    • Embed it as a Rust dev dependency and run tests in process with no endpoint setup.
    • Run the same small script against the emulator and a sandbox AWS account you own and diff the responses.
  2. #2Models & launches

    Gemini Decisions API: one call scores up to 1,000 questions

    Google launched the Gemini Decisions API with a new model, gemini-decision-1, that returns answers to fixed option questions instead of generating text. One request can carry up to 1,000 yes or no, multiple choice with 2 to 255 options, or rubric score questions over a shared input of text, images, audio, video and PDFs up to 256K tokens. Every answer includes a confidence score from 0 to 1.

    Heat · HN front page, 417 points and 125 comments on Nov 11, plus a Simon Willison write up
    What people foundGoogle lists $0.10 per 1M input tokens, $0 for output and 150 to 250 ms end to end latency, which makes it a cheap classifier, router or eval grader. Simon Willison's video test returned graded probabilities such as about 0.92 for robot and 0.05 for dragon, where Jev and OpenAI's GPT-6 Luna decision model return exactly 1.0 or 0.0, and he calls Gemini's version fully multimodal.
    Learn it15 min · 5 steps

    Key ideas

    Option scoring
    The model rates a fixed set of answers rather than writing free text, so there is nothing to parse.
    Calibrated confidence
    A probability that tracks how often the answer is right, which lets you set thresholds and send unsure cases to a human.
    Shared input billing
    The common input is billed once per request, plus each question's own instruction tokens, so many questions on one document stay cheap.

    Steps

    1. Read the Google blog post first for the three question types, limits and pricing.
    2. Read Simon Willison's test to see real probability outputs and the comparison with Jev and OpenAI.
    3. Note the question types in the API: noul for yes or no, choice, and score for ordered rubrics.
    4. Think of one place in your stack where an LLM answer is parsed into a label today.
    5. Self check: when would a 0.92 score be more useful to you than a hard 1.0?
    Try it30 min · 5 steps
    You need: Python, the google-genai package, and a Gemini API key from Google AI Studio, which has a no cost tier for prototyping.

    Steps

    1. Create an API key in Google AI Studio.
    2. Install the google-genai Python package and set the client to api_version v1beta as the blog example shows.
    3. Call client.decisions.create with model gemini-decision-1, one short text you own and three yes or no questions.
    4. Label 20 of your own support tickets or commit messages by hand, then run the same labels as one choice question per item and compare.
    5. Plot or list the confidence scores for right and wrong answers and pick a threshold below which items go to manual review.

    Small angles to try

    • Use a score rubric question to grade answers from your own eval set.
    • Send an image or a PDF you own with the same questions to test the multimodal input.
    • Check how latency changes from 1 to 100 questions in one request.
  3. #3AI in the world

    48 agents flag a Maya site in public lidar, but it was already mapped

    Pietro Miotti, who says he is not an archaeologist, ran 48 AI agents over about 1,200 km2 of public lidar strips looking for pyramids, platforms and plazas, then checked the raw point cloud. He reported a hilltop site to Mexico's INAH on 20 October, and INAH replied on 21 October that its records showed no registered site there, all within about 72 hours of the first idea on 18 October. An update on 11 November says the site was already mapped in Estrada-Belli et al. 2025, a paper the agents missed in their literature review.

    Heat · HN front page, 983 points and 358 comments, the top AI story of the day
    What people foundThe correction is the useful part: the agents found real structures fast, but the prior work check failed, so the author now calls it a rediscovery. An official registry saying no site is listed is not proof that nobody has published it, so novelty claims from agents need a separate literature search.
    Learn it15 min · 5 steps

    Key ideas

    Lidar
    Laser scanning from the air that can show ground shapes under forest canopy, such as platforms and plazas.
    Point cloud
    The raw set of 3D laser returns, which is checked to confirm a shape seen in processed images is real.
    Prior work check
    A search of published literature to confirm a finding is new before calling it a discovery.

    Steps

    1. Read Pietro Miotti's post, including the 11 November update at the top.
    2. Follow the timeline from 18 to 21 October and note which steps agents did and which the author did.
    3. Look up Estrada-Belli et al. 2025 on Google Scholar to see what a successful literature check would have found.
    4. List the gap between an official registry and the published literature.
    5. Self check: what single step would have turned this rediscovery into a correct claim before the report went out?
    Try it30 min · 5 steps
    You need: Any chat assistant or agent you already use and a free scholarly search such as Google Scholar. No API key needed.

    Steps

    1. Pick a result you or your team recently called new, such as a bug pattern, an algorithm trick or a dataset finding.
    2. Ask an agent to do a literature and web check for prior work and to list sources with links.
    3. Do your own independent search with Google Scholar or a search engine using different keywords.
    4. Compare: list sources the agent missed and sources it cited that do not say what it claimed.
    5. Write a short checklist for novelty claims that includes at least one human search step.

    Small angles to try

    • Run the same check with two different assistants and compare what each misses.
    • Give the agent the Maya site description and see if it finds the 2025 paper.
    • Time how long the human search takes versus the agent search.
  4. #4Research

    LP and Tree-Ising redo MIT's agent weight edit with plain optimization

    MIT CSAIL's Sparks of Weight-Space Programming paper by Takehi, Teoh, Zhang and Khattab had the GPT-6 Astra agent fix thinking loops in Qwen3.5-0.8B by choosing weight edits over about 8 hours. Nathan Barry's 9 November post recasts the edit as a sparse linear program, and the Tree-Ising paper by Salloum and Ali splits it into one Ising or QUBO problem per output row over BF16 sign and exponent bits, solved on a GPU with simulated annealing.

    Heat · HN 273 points and 38 comments for the LP post on Nov 11; Tree-Ising was the top paper of the day on HF Papers with 78 upvotes
    What people foundBarry's LP changes about 8 times fewer weights than Astra, 16,291 versus 130,188, and matches it on Qwen3.5-0.8B, GSM8K 69.6 versus 68.2 and MATH-500 51.7 versus 50.9, with his custom GPU solver taking about a minute per solve where HiGHS took over 20 minutes. Tree-Ising reports 54.6% on MATH-500 versus 51.7% for the agent edit in its own runs, using 4 forward pass requests instead of 50 and about $0.90 of GPU time. Barry also shows a plain steering vector nearly matches the fix without touching weights.
    Learn it15 min · 5 steps

    Key ideas

    Rank 1 edit
    A weight change built from two vectors, here a loop direction and a vector choosing which inputs to push against it.
    L1 sparse linear program
    An optimization that minimizes the sum of absolute values, which drives most entries to zero so only a few weights change.
    Ising or QUBO problem
    A problem over binary choices, here which bits to flip, that annealing style solvers handle well on GPUs.

    Steps

    1. Read Nathan Barry's post first, since it explains the loop problem and the LP in plain terms.
    2. Note where the edit sits: two MLP down_proj matrices in layers 15 and 19.
    3. Read the Tree-Ising abstract on HF Papers and compare its bit level edits with the LP's real valued edit.
    4. Note that each work measures the agent edit itself, so baseline numbers differ slightly between them.
    5. Self check: why does an L1 objective give sparse edits while an L2 objective would not?
    Try it30 min · 5 steps
    You need: Python 3.11 for a dry run on a laptop; a Linux NVIDIA CUDA GPU for real repairs, the README reports an A100 40 GB. No API key needed. The README states no license yet.

    Steps

    1. Clone the Tree-Ising repository and create a Python 3.11 virtual environment as the README shows.
    2. Install the base package with python -m pip install -e ..
    3. List the experiments with run_treeising --config configs/qwen35_0.8b.yaml --list and read the config to see the models and revisions it pins.
    4. On a GPU machine, install the reproduce extras with python -m pip install -e ".[reproduce]" -c requirements-reproduce.txt and run run_treeising --config configs/qwen35_0.8b.yaml --only models,data,splits,check_data.
    5. Run the annealing repair with run_treeising --config configs/qwen35_0.8b.yaml --only practice_answers,statistics,layer_scan,treeising_sa and then --only benchmarks,heldout400 to compare with the agent checkpoint.

    Small angles to try

    • Try the simulated bifurcation repair with --only treeising_sb.
    • Read Barry's weight-space-lp code and compare the edited columns with Tree-Ising's flipped bits.
    • Test Barry's steering vector idea at 0.5x, 1x and 2x strength and watch for loops returning.
  5. #5Trends × tech

    Gmail and Drive × chatbots: one user says Gemini may expose them

    A GitHub user, thisisme6598, published tests claiming that Gemini on a personal account with Gmail and Drive connected sometimes searches them during ordinary chats with no @Gmail or @Drive tag, returning anything from nothing to a summary, a short quote or the full email or file text. The write-up argues such content may then reach Google's human reviewers, citing Google's Help page warning not to connect apps holding confidential information you would not want a reviewer to see, and says this reverses a 2023 promise that personal Gmail, Docs and Drive content would not be seen by human reviewers. GIGAZINE relayed the claims on 11 November; neither GIGAZINE nor the repository reports any response from Google.

    Heat · HN front page, 736 points and 148 comments on Nov 9 for the GitHub write up; GIGAZINE follow up on Nov 11
    What people foundThese are one user's tests, with no sample sizes, and the same prompt gave different outcomes across runs; human review exposure is described as possible, not shown. GIGAZINE says the repository notes the tests cover personal accounts and that commercial accounts exclude human review and training by default, though that note is not in the README. Before linking work files to any assistant, check which account type the connector runs under and Google's own reviewer warning.
    Learn it15 min · 5 steps

    Key ideas

    Connected apps
    A setting that lets a chatbot search your email and files to answer questions, not only when you name them.
    Human review
    People at the provider reading samples of conversations to improve the product, which can include any data pulled into a chat.
    Nondeterministic retrieval
    The model decides each time whether to fetch data, so one clean test run proves little.

    Steps

    1. Read the gemini_files README first, including its four outcome categories and test method.
    2. Read the GIGAZINE article to see which lines it adds, such as the account type note.
    3. Find Google's warning about reviewers and connected apps, cited in both sources, and read it in full.
    4. List what data your own assistant connectors can reach.
    5. Self check: why does a single test that returns no personal data not disprove the claim?
    Try it30 min · 5 steps
    You need: A personal Google account you own, ideally a fresh test account with only dummy data, and the Gemini app. No API key needed.

    Steps

    1. In a test account, create a dummy email and a Drive file with a made up trip booking and a unique fake marker word; use no real personal data, since content may reach reviewers.
    2. Turn on Connected Apps for Google Workspace in Gemini settings.
    3. In a new chat, mention a trip in passing without any @Gmail or @Drive tag and note whether the marker or booking details appear.
    4. Delete the conversation and repeat at least five times, recording each result as no retrieval, summary, partial quote or full text.
    5. Turn Connected Apps off, delete the test chats, and write down what you would keep out of connected accounts.

    Small angles to try

    • Repeat with a different model setting and compare outcome counts.
    • Run the same trial with another assistant's email or file connector on a test account.
    • If you have a work account, read your admin's data policy for it instead of testing it.
  6. #6Tools & open source

    Yogurt: MIT licensed, fully local meeting notes, no bot joins the call

    FrigadeHQ released Yogurt, a free desktop app that records meetings in the background without adding a bot, then transcribes, summarizes and answers questions with on device models. In the Show HN post the author says it runs on whisper.cpp, Apple's SpeechAnalyzer where available, and llama.cpp, is built with Electron, and can share notes with cloud models through a local MCP server. Installers exist for macOS 14.4 or later on Apple Silicon, Windows 10 x64 and Linux x64 as an AppImage.

    Heat · Show HN, 233 points and 94 comments on Nov 11; three releases on Nov 11, latest v0.3.5; 74 GitHub stars at review time
    What people foundPlan for 8 GB of memory. The useful design idea is the local MCP server: transcripts stay on disk by default, and an agent only sees them when the user points it there.
    Learn it15 min · 5 steps

    Key ideas

    On device transcription
    Speech is turned into text by a model running on the local machine, so the audio never goes to a server.
    llama.cpp
    An open source C and C++ runtime that runs quantized language models on ordinary CPUs and GPUs.
    Local MCP server
    A small server on the same machine that exposes data, here meeting notes, as tools an AI client can call through the Model Context Protocol.

    Steps

    1. Read the Show HN post by the Yogurt author first, since it explains the pipeline and the reasons for each part.
    2. Skim the Yogurt README for platforms, memory needs and the download names on the releases page.
    3. Open docs/README.md in the repo to see how recording, transcription and summarization connect.
    4. Look up what the Model Context Protocol is and how a client connects to a local server.
    5. Self check: which steps of a Yogurt meeting run locally, and what is the only path by which a note can reach a cloud model?
    Try it30 min · 6 steps
    You need: A Mac with Apple Silicon on macOS 14.4 or later, Windows 10 x64, or Ubuntu 22.04 x64; 8 GB of memory; a microphone. No API key needed.

    Steps

    1. Download the installer for your platform from the latest release, for example Yogurt-mac-arm64.dmg, Yogurt-windows-x64.exe or Yogurt-linux-x86_64.AppImage.
    2. Install and open the app, then let it fetch the local models it picks for your hardware.
    3. Record a five minute mock meeting with yourself or a teammate who agrees, reading a short agenda aloud.
    4. Read the transcript and summary, and note any wrong names, numbers or action items.
    5. Ask the app two questions about the meeting, one answerable and one not, and check whether it admits the gap.
    6. Optionally connect a cloud client to the local MCP server and check exactly which notes it can read.

    Small angles to try

    • Turn off Wi Fi before recording to confirm the whole pipeline works offline.
    • Compare summary quality on a machine with 8 GB versus 16 GB or more of memory.
    • Record the same clip in a second language and compare transcript accuracy.
  7. #7Models & launches

    Huawei CSL posts MACaQ, a 2 to 4 bit quantizer for GGUF and vLLM

    Huawei CSL published MACaQ, short for Model Agnostic Calibrated Quantization, on GitHub under Apache 2.0. It produces 2, 3 and 4 bit GGUF files for llama.cpp and W4A4 models for vLLM, Diffusers and PyTorch speech, with supported models including Qwen3.8-27B, Qwen3-Coder-30B-A3B, Qwen3.8-Flash-Next, FLUX.2-klein-4B and Wan2.2. The repo has 16 commits and no tagged releases so far.

    Heat · HN, 98 points and 24 comments on Nov 11
    What people foundThe README reports Qwen3.8-27B in vLLM at W4A4 with perplexity 2.8994 against 2.8463 for BF16, and 2.33x prefill and 2.58x decode speed. It also says a 122B model quantizes in 64 minutes with 32.5 GB of device memory. No prebuilt quantized checkpoints are linked, so plan to quantize yourself.
    Learn it15 min · 5 steps

    Key ideas

    Calibrated quantization
    Weights are rounded to fewer bits using a small sample of real text to decide how to round with the least damage.
    W4A4
    Both the weights and the activations are stored in 4 bits, which speeds up the math, not only the memory footprint.
    Perplexity
    A measure of how surprised a model is by held out text; lower is better, and a small rise after quantization means little quality loss.

    Steps

    1. Read the MACaQ main README first, especially the backend list and the Qwen3.8-27B table.
    2. Open the llama.cpp backend README in the repo to see the convert, quantize and run flow.
    3. Read a short primer on GGUF quant types such as IQ4_XS so the format names make sense.
    4. Compare the BF16 and W4A4 rows and work out the relative perplexity change yourself.
    5. Self check: why can W4A4 speed up decode while weight only 4 bit mainly saves memory?
    Try it30 min · 6 steps
    You need: Linux with Python 3.12, CMake, and ideally an NVIDIA GPU with the CUDA toolkit; CPU only works but is slower. Disk space for a small model such as Qwen/Qwen3-4B. No API key needed.

    Steps

    1. Clone the repo, create a Python 3.12 venv, and install requirements/llamacpp.txt, or requirements/llamacpp-cpu.txt for CPU only.
    2. Build the bundled llama.cpp with CMake as the llamacpp README shows, including the llama-quantize and llama-perplexity targets.
    3. Download Qwen/Qwen3-4B and convert it with convert_hf_to_gguf.py --outtype f16 to an FP16 GGUF.
    4. Run llamacpp/macaq.py quantize with --format IQ4_XS on your own calibration text file, adding --context 128 --ubatch 128 --chunks 2 for a quick first pass.
    5. Generate with llamacpp/macaq.py run and a fixed prompt, then compare answers and file sizes against the FP16 GGUF.
    6. Measure perplexity of both files on the same held out text with the built llama-perplexity tool.

    Small angles to try

    • Rerun with --vq-sweeps 4 and check whether perplexity improves at the cost of time.
    • Quantize the same model with stock llama-quantize IQ4_XS and compare against the MACaQ file.
    • On CPU only, use --device cpu and --gpu-layers 0 and time the whole run.
  8. #8Tools & open source

    Jane Street publishes a Socratic tutor skill for coding agents

    Jane Street's new agent_skills repo, MIT licensed, holds one skill, socratic-tutor, that turns an agent into a strict tutor that teaches by asking. It runs a 3 to 5 question diagnostic, builds a 0 to 5 skill map over 4 to 10 subskills, asks one question at a time, and explains directly only after 2 to 3 failed tries. Earlier subskills come back as unannounced reviews, and a level drops if the user fails one.

    Heat · HN, 110 points and 27 comments on Nov 10
    What people foundIt activates on requests like teach me, tutor me, quiz me, drill me or help me master, not on one off explain questions. A level counts only after two independent checks including one unprompted recall, and the save state and resume from commands carry progress between sessions.
    Learn it15 min · 5 steps

    Key ideas

    Agent skill
    A folder with a SKILL.md file of instructions that a coding agent loads when a request matches its description.
    Socratic method
    Teaching by asking narrowing questions so the learner reasons to the answer instead of being told it.
    Spaced review
    Retesting earlier material after a gap, which makes knowledge stick better than one pass.

    Steps

    1. Read the socratic-tutor SKILL.md first; it is short and doubles as an example of skill writing.
    2. Note the trigger phrases and how the skill refuses one off explain questions.
    3. Study the level rules: two independent checks, unprompted recall, and the extra bar for level 5.
    4. Read the repo README for the two ways to use it, as a skill folder or pasted into a chat prompt.
    5. Self check: what does the tutor do after your first wrong answer, and what changes after the third?
    Try it30 min · 5 steps
    You need: A coding agent that supports SKILL.md skills, or any chat model where a system prompt can be pasted. No API key needed beyond your usual agent login.

    Steps

    1. Clone the agent_skills repo and copy the socratic-tutor directory into your agent's skills folder.
    2. Start a session and say: teach me how this repo's build works, pointing the agent at a project of your own.
    3. Answer the diagnostic honestly and check that it shows a skill map before teaching.
    4. Give one deliberately wrong answer and confirm it narrows the question instead of revealing the answer.
    5. Type save state, start a fresh session, and paste the block after resume from: to check that progress returns.

    Small angles to try

    • Paste the SKILL.md body into a plain chatbot and compare its behavior with the agent version.
    • Use just tell me on one question and confirm Socratic mode resumes on the next.
    • Ask for full mastery and see how the final mixed assessment differs from the default target of 4.
  9. #9Tools & open source

    Klad: a paint app your coding agent draws in through a CLI

    Klad is a 2D painting and animation app where a person paints with a pen and a coding agent paints on the same drawing through commands like klad stroke, klad fill and klad undo. There is no image model: strokes, shapes and fills are stored as data, so every element stays editable, and klad skill prints an agent skill file. The app is source available under the PolyForm Internal Use License, while the file format, CLI reference, skills and client crate are MIT.

    Heat · Show HN, 79 points and 31 comments on Nov 11
    What people foundThe license is free for personal use and for companies under 500,000 euros in yearly revenue; larger companies must buy a licence per user. There are no installers yet, so build from source with Rust and just, about ten minutes the first time; macOS is the main target and Linux has mouse input only.
    Learn it15 min · 5 steps

    Key ideas

    Vector strokes as data
    Each brush mark is saved as editable parameters, not pixels, so an agent can read, change or undo it precisely.
    CLI as agent interface
    Giving an agent small, exact commands with readable output instead of asking it to drive a GUI or generate an image.
    Source available license
    The code is public to read and build, but use is limited by terms that are stricter than an open source license.

    Steps

    1. Read the Klad README first, focusing on the CLI command list and license section.
    2. Open docs/cli.md in the repo to see the exact arguments and JSON output of the commands.
    3. Read docs/skill/SKILL.md to see how it tells an agent to look, change and check a drawing.
    4. Compare this design with an image generation approach and list what an agent gains from editable strokes.
    5. Self check: which parts of Klad could you reuse in a commercial product without buying a licence, and why?
    Try it30 min · 5 steps
    You need: macOS preferred, or Linux with pkg-config, libfreetype-dev, libfontconfig1-dev, libxkbcommon-dev and libwayland-dev; Rust via rustup and just; a coding agent. No API key needed for Klad itself.

    Steps

    1. Run git clone https://github.com/klad-app/klad && cd klad and then just compile, which takes about ten minutes the first time.
    2. Start the editor with target/release/klad open, or target/linux/release/klad open on Linux, and put the CLI on your path with target/release/klad install.
    3. Run just samples and open a sample drawing.
    4. Give your coding agent the output of klad skill and ask it to add a simple shape and a fill to the open drawing.
    5. Read what it did with klad strokes, adjust one stroke by hand, then ask the agent to klad undo its last change.

    Small angles to try

    • Ask the agent for a three frame bounce and use klad inbetween and klad arcs to check the motion.
    • Export with klad export --godot and load the result in a Godot project.
    • Feed the agent klad skill drawing as well and compare its drawing quality.
  10. #10Engineering & CS

    Critical Apache httpd HTTP/2 RCE, found by Anthropic and Google

    Apache fixed CVE-2026-119618, a critical race condition in mod_http2's assign_headers that can corrupt heap memory and lead to remote code execution, in httpd 2.4.71 released on Nov 11. Apache credits Anthropic's OSS Scanner and two researchers from Google Cloud Vulnerability Research, who reported it on Oct 26 and Nov 3. Wiz reported no public exploit and no use in the wild at publication.

    Heat · HN, 179 points and 76 comments
    What people foundAffected: httpd 2.4.55 to 2.4.70 and standalone mod_h2 2.0.8 to 2.0.42; upgrade to 2.4.71. Wiz suggests Protocols http/1.1 to turn off HTTP/2 as a stopgap, and counts vulnerable resources in 64% of cloud environments and publicly exposed instances in 51%, before checking configuration. The Wiz post has a wrong CVE number in one heading and conflicting mod_h2 fix versions, so trust Apache's advisory.
    Learn it15 min · 5 steps

    Key ideas

    Race condition
    Two threads touch the same data at once without proper locking, so the result depends on timing and can corrupt memory.
    mod_http2
    The httpd module that serves HTTP/2; it only matters if it is loaded and h2 or h2c is enabled in the Protocols directive.
    Protocols directive
    The httpd setting that lists which protocols a server or virtual host offers, such as h2, h2c and http/1.1.

    Steps

    1. Read the CVE-2026-119618 entry in Apache's httpd 2.4 vulnerabilities page first.
    2. Note the affected ranges for httpd and for standalone mod_h2, and the fixed release 2.4.71.
    3. Read the Wiz post for context and spot its two inconsistencies against the Apache entry.
    4. Look up the Protocols directive in the httpd core documentation.
    5. Self check: a server runs 2.4.70 with mod_http2 not loaded; is it exposed, and should it still upgrade?
    Try it30 min · 5 steps
    You need: Docker on your own machine and your own httpd config files. No API key needed. Checks only, against local containers and your own installs; no traffic to any other system.

    Steps

    1. Pull the official image and print its version with docker run --rm httpd:2.4 httpd -v; tags such as httpd:2.4.71 are listed on Docker Hub.
    2. List loaded modules with docker run --rm httpd:2.4 httpd -M and look for http2_module.
    3. Copy the default config out with docker run --rm httpd:2.4 cat /usr/local/apache2/conf/httpd.conf > my-httpd.conf and search it for LoadModule http2_module and any Protocols line.
    4. Do the same on your own servers: run httpd -v or the apache2 equivalent your distro uses, then httpd -M, and compare the version with the 2.4.55 to 2.4.70 range.
    5. If you cannot upgrade at once, add Protocols http/1.1 to a copy of your config and validate it with httpd -t before deploying.

    Small angles to try

    • Write a small script that checks the version and module list across all your own containers.
    • Check whether your distro has backported the fix, by reading its security tracker for the CVE.
    • Review the other two CVEs fixed in 2.4.71, in mod_md and mod_vhost_alias, against your module list.
  11. #11Research

    RoboScript: robots learn a task in context from one demo

    Researchers from ArcheOne-AI propose World In-Context Learning, where a frozen model starts from a single demonstration, rehearses the task in a simulator and writes the feedback into its context instead of its weights. Their RoboScript framework freezes the result as a reusable Task Memory, deploys it on new scenes and adapts it to a physical dual-arm robot.

    Heat · HF Papers, third on the Nov 11 daily list with 65 upvotes; no HN thread found
    What people foundPer the authors, RoboScript reaches 83.4% mean success across all 50 RoboTwin 2.0 tasks in both clean and randomized settings, using one demonstration per task and no gradient updates, more than 20 points above the strongest trained action policy. The pattern is the agent harness loop of rehearse, write notes, reuse, applied to robots.
    Learn it15 min · 4 steps

    Key ideas

    World In-Context Learning
    A frozen model improves at a task by practicing in a simulator and keeping what it learned as text in its context, with no retraining.
    Task Memory
    The frozen record of what worked during rehearsal, reused when the robot meets a new scene of the same task.
    RoboTwin 2.0
    A simulated benchmark of 50 dual-arm manipulation tasks with clean and randomized scene settings.

    Steps

    1. Read the abstract on the HF Papers page for World In-Context Learning for Agentic Robot Manipulation.
    2. Open the RoboScript README and note the parts it lists: model gateway, perception with GroundingDINO and SAM 3, cuRobo motion planning and benchmark adapters.
    3. Compare the loop to agent memory files used by coding agents: rehearse, record feedback, freeze, reuse.
    4. Self check: why can a frozen model beat a trained action policy here, and what does it give up by never updating weights?
    Try it30 min · 5 steps
    You need: Python 3.10 or newer and uv; the mock quickstart needs no simulator, GPU or model, so no API key is needed for it. RoboTwin 2.0 runs need a separate simulator install.

    Steps

    1. Clone the repo with git clone https://github.com/ArcheOne-AI/RoboScript.git and enter the folder.
    2. Install dependencies with uv sync --all-packages.
    3. Change into examples/tutorials and run uv run python -m quickstart_mock.run to play one mock episode.
    4. Read the printed episode output and find where feedback is written into the Task Memory.
    5. Open docs/README.md and trace how a frozen memory is loaded for a new scene.

    Small angles to try

    • Run the mock episode twice and diff the stored memory between runs.
    • Read the RoboTwin 2.0 benchmark README to estimate the hardware a full evaluation needs.
    • Sketch how the same rehearse and freeze loop would work for a browser or coding agent.
  12. #12Research

    WikiTrace: agent memory that links each lesson to its source run

    KAIST AI researchers argue that distilled lessons in agent memory can hurt an agent when it cannot check where a lesson came from. WikiTrace stores each insight with a summary of why its source run succeeded or failed plus the raw trajectory, and the agent follows those links with terminal tools when a lesson looks incomplete, conflicting or came from a failed task.

    Heat · HF Papers, fourth on the Nov 11 daily list with 52 upvotes; no HN thread found
    What people foundThe README reports WikiTrace as the best method on all eight benchmark suites, 8.5 points above the best prior memory method on average, and an HF comment by author Yumin Choi puts the gain at 3.2 points over using the same lessons alone. For engineers keeping wiki or CLAUDE.md style memory, storing a link back to the original transcript is a cheap safeguard.
    Learn it15 min · 4 steps

    Key ideas

    Memory induced degradation
    When stored lessons make an agent fail tasks it would have solved without any memory.
    Evidence linked memory
    Each lesson keeps pointers to the run summary and raw trajectory it came from, so the agent can verify it.
    Insights only baseline
    The same lessons given to the agent without any links, used to isolate the value of the provenance links.

    Steps

    1. Read the abstract on the HF Papers page for Don't Just Trust Your Memory.
    2. Read the WikiTrace README results table and the list of five benchmarks: ALFWorld, ScienceWorld, PlanCraft, τ³-bench Retail and AppWorld.
    3. Compare the build step, which runs the agent without memory on training tasks, with the run step that uses the memory.
    4. Self check: in which three situations does the agent leave the lesson and open the raw trajectory?
    Try it30 min · 6 steps
    You need: Python 3.11 or newer, uv, Node 20, Docker and an OpenRouter API key set as OPENROUTER_API_KEY in .env; model calls cost money.

    Steps

    1. Clone the WikiTrace repo, run uv sync --extra dev, then copy .env.example to .env and add your key.
    2. Run scripts/setup/pi.sh and scripts/setup/sandbox_image.sh to install the terminal agent and its offline sandbox.
    3. Set up one benchmark with DATA_ROOT=data scripts/setup/alfworld.sh.
    4. Build memory with uv run wikitrace build --config configs/alfworld.yaml --benchmark alfworld --suite valid-seen --trial 0.
    5. Run the same suite with --method no-memory, then --method kinj, then --method wikitrace using uv run wikitrace run and the same trial number.
    6. Compare results with uv run wikitrace analyze all --runs outputs/ --out analysis/ after uv sync --extra analysis.

    Small angles to try

    • Repeat the no memory run twice to see how much success flips from noise alone.
    • Try the unseen suite, valid-unseen, to test whether linked lessons transfer.
    • Add source links to an agent memory file in your own project and log when the agent opens them.
  13. #13Tools & open source

    Pomerado, YC F26, turns websites into MCP servers for agents

    Pomerado has an agent explore a website once, record the steps and turn them into a deterministic hosted API or MCP server, with a separate Guardian model checking actions against guardrails. When a site changes, a maintenance job repairs the tool, and the founders say sign in runs through a deterministic flow so models never read customer credentials. An MIT licensed open source version and a free hosted tier are available.

    Heat · Launch HN on Nov 11; HN API views disagree on points, from 11 to 55
    What people foundThe founders say calls typically return in 5 to 30 seconds and cost less than agent driven browsing; these are vendor claims. The free tier allows 100 tool calls a day and 10 new integrations a month, and logins can be sent per request or stored with 2FA.
    Learn it15 min · 4 steps

    Key ideas

    MCP server
    A service that exposes tools in the Model Context Protocol format so any compatible agent can call them.
    Deterministic replay
    Recorded steps run the same way each time, without a model reasoning through every page.
    Self repair
    A background job notices when a site change breaks a tool and patches the recorded steps.

    Steps

    1. Read the Launch HN post by the founders first.
    2. Skim the Pomerado site sections on self repair, pricing and credential handling.
    3. Read the comment thread for open questions, such as where session cookies are stored.
    4. Self check: why is replaying recorded steps faster and cheaper than a computer use agent, and when would it break?
    Try it30 min · 5 steps
    You need: A free Pomerado account or the open source repo, plus a small website you own or run locally; no third party sites.

    Steps

    1. Start a simple site you control, such as a local form that saves a note, or use a test page on your own domain.
    2. Open the Pomerado docs and the GitHub README, and follow their setup for the hosted tier or the open source version.
    3. Create an integration for your own site's form action and inspect the generated input and output schema.
    4. Call the tool from an MCP client and time several calls.
    5. Rename a button on your site and call the tool again to see whether the repair job fixes it.

    Small angles to try

    • Compare call time against a Playwright script for the same action on your site.
    • Add a login to your test site and check how credentials are passed per request versus stored.
    • Write a guardrail rule and try a call that should be blocked.
  14. #14AI in the world

    United Airlines bans third party personal AI agents from booking

    United said on Nov 11 that it will prohibit unauthorized third party personal AI agents or assistants from booking flights, both on its own platforms and indirectly through travel agencies and aggregators, and is updating its booking and ticketing policies. It cites exposure of account and payment credentials and federal rules that require travelers to confirm in person that bags hold no banned items.

    Heat · HN front page, 543 points and about 530 comments
    What people foundAnalyst Henry Harteveldt of Atmosphere Research Group says the move keeps United in control of booking and avoids the cost of heavy AI bot searches, and notes fares stay bookable through Kayak, Hopper and Google Flights. Southwest is going the other way by exposing its inventory to ChatGPT users. CBS reports no enforcement method, so agent builders should treat consent and site terms as product requirements.
    Learn it15 min · 4 steps

    Key ideas

    Personal AI agent
    An assistant that acts for a user on outside sites, such as searching and paying for a flight.
    Aggregator
    A site like Kayak or Google Flights that collects fares from many airlines in one place.
    Agent authorization
    Proof that a site and a user both allow an agent to act, beyond simply holding the user's password.

    Steps

    1. Read the CBS News article first.
    2. List each reason United gives and decide which is technical and which is commercial.
    3. Compare with the Southwest approach of opening inventory to ChatGPT.
    4. Self check: what would an agent need to show a site to count as authorized rather than unauthorized?
    Try it30 min · 4 steps
    You need: Any agent project or script you own; no API key needed for the exercise; no testing against airline or other third party sites.

    Steps

    1. Write down every external site your own agent can buy from or log in to.
    2. For each, note whether its terms you have read allow automated purchases, and mark unknown ones as blocked.
    3. Add a domain allowlist and a confirmation prompt before any payment step in your agent, using a mock checkout page you run locally.
    4. Run the agent against the mock checkout and confirm it stops for consent and refuses blocked domains.

    Small angles to try

    • Log every credential the agent touches and see which could be replaced by a scoped token.
    • Write a short user facing message explaining why the agent will not book on a blocked site.
    • Compare against an aggregator path that the site still allows.
  15. #15Models & launches

    Tether's BotPsy-0.8B: small Apache 2.0 tool caller for local agents

    Tether's QVAC team released BotPsy-0.8B on Nov 11, built on Qwen3.5-0.8B with supervised fine tuning, multi teacher on policy distillation from three Qwen3.5-4B specialists and two rounds of DAPO reinforcement learning. It targets function calling for search, travel, files, calendar, email and device control, with GGUF files in BF16, Q8_0 and Q4_K_M for llama.cpp.

    Heat · Not measured beyond 5 likes and 39 downloads last month on Hugging Face; no HN thread found
    What people foundAgainst its base, the model goes from 23.4 to 44.0 on BFCL v4 and from 23.2 to 51.5 on τ-bench, where that figure is the mean of retail, telecom and banking subsets and banking stays at 4.1. The authors say it suits low to medium difficulty tasks, needs thinking mode turned on with enable_thinking=True, and describe the Apache 2.0 weights as for research and educational use.
    Learn it15 min · 4 steps

    Key ideas

    On policy distillation
    The small model generates its own answers and larger teacher models score or correct them, so it learns from its own mistakes.
    BFCL
    The Berkeley Function Calling Leaderboard, which checks whether a model calls the right function with the right arguments.
    GGUF quantization
    Compressed weight files, such as Q4_K_M, that let llama.cpp run a model on a laptop or phone.

    Steps

    1. Read the Hugging Face blog post on BotPsy first.
    2. Open the model card and read the usage section and the note on thinking mode.
    3. Look at the τ-bench subset scores to see where the average hides weak spots.
    4. Self check: why does a 0.8B model need tools for facts, and which tasks would you not trust it with?
    Try it30 min · 5 steps
    You need: A laptop with a recent llama.cpp build, about 1 GB of disk for the GGUF; no API key needed.

    Steps

    1. Download botpsy-0.8b-q8_0.gguf from the qvac/BotPsy-0.8B-GGUFs repo on Hugging Face.
    2. Start the server with llama-server -m botpsy-0.8b-q8_0.gguf --jinja -c 32768 --temp 0.3 --top-p 0.95 --top-k 20 --presence-penalty 1.5 --chat-template-kwargs '{"enable_thinking": true}' --port 8080.
    3. Send an OpenAI style chat request to port 8080 with one tool, such as the create_event calendar tool from the model card, and the prompt to add the dentist on Friday at 3 pm.
    4. Check that tool_calls holds the right function and arguments and that reasoning appears in reasoning_content.
    5. Repeat with ten prompts of your own and count correct calls.

    Small angles to try

    • Try the Q4_K_M file and compare accuracy and speed.
    • Turn thinking off and see how often calls break.
    • Give it three similar tools and test whether it picks the right one.
  16. #16Models & launches

    AlphaGenome 2: DeepMind's DNA model now predicts per cell type

    Google DeepMind released AlphaGenome 2, a conditional model that reads a 1 million letter DNA sequence together with a target tissue, cell type or state and predicts how that sequence behaves in that context. It was trained on bulk and single cell data from human and mouse, with over 107 million single cell profiles across 68 tissues, and it can take a condition built from a lab's own single cell data without retraining.

    Heat · HN, 13 points and 0 comments; niche, but carried by the DeepMind name
    What people foundOn TenK10K, a single cell map of 28 immune cell types, it found causal variants 30% of the time versus 16% for the original AlphaGenome fine tuned on single cell data and other specialized tools. Access is through the AlphaGenome API, free for non commercial research, with commercial research in private preview; the post says nothing about releasing weights.
    Learn it15 min · 5 steps

    Key ideas

    Conditional model
    A model that takes an extra input describing the context, here a tissue or cell type, and changes its prediction to match that context.
    Variant effect prediction
    Comparing the model's output for a normal DNA sequence and the same sequence with one letter changed, to estimate what the change does.
    Single cell profile
    A measurement of which genes are active in one individual cell, so models can learn differences between cell types instead of tissue averages.

    Steps

    1. Read the DeepMind blog post first, focusing on how the condition is fed into every stage of the model.
    2. Skim the AlphaGenome 2 paper linked from the post, looking only at the architecture figure and the TenK10K evaluation.
    3. Open the AlphaGenome docs site and read the notebook titled Choosing a condition to see how a tissue or cell type is picked from the reference catalog.
    4. Note the difference between Atlas lookups of precomputed results and live model predictions, which have different rate limits.
    5. Self check: why would a variant that looks harmless in liver predictions still matter in an immune cell?
    Try it30 min · 6 steps
    You need: Python, a free AlphaGenome API key for non commercial use requested from the AlphaGenome site, a browser or Colab; no GPU needed

    Steps

    1. Request a free non commercial API key from the Get an API key link on the AlphaGenome docs site and read the terms.
    2. Follow the docs Installation page to install the official Python client in a fresh virtual environment or Colab.
    3. Run the v2 notebook Choosing a condition and pick two contrasting conditions, for example a liver tissue and an immune cell type.
    4. Run the notebook Predicting tracks for an interval or sequence on the same well known gene region for both conditions.
    5. Plot the two predicted tracks side by side and write down where they differ most.
    6. Change one letter in the input sequence and compare the prediction against the original for each condition.

    Small angles to try

    • Compare an Atlas lookup with a live prediction for the same region and note speed and rate limits.
    • Try a condition built from a public single cell dataset instead of the reference catalog.
    • Repeat with the legacy AlphaGenome 1 quick start notebook to see what the conditioning adds.
  17. #17Models & launches

    Caramel-1: a 413M speech to speech model trained only on audio

    Researcher naklecha released caramel-1, a 413M parameter conversation model trained on 11,555 hours of speech, about half two person conversations and half single speaker audio, with no text at all. Training took about 2 hours 15 minutes, roughly 4 GPU hours. Weights are on Hugging Face and the code and training recipe are on GitHub.

    Heat · Show HN, 4 points and 0 comments
    What people foundIt generates 1 second of audio in about 0.09 seconds on a GPU, and two copies talking to each other produce 1 second of conversation in about 0.25 seconds on an M3 Pro MacBook Pro. The author admits a generated speech detector still tells it apart from real people easily. The write up names no license, but the GitHub repo is marked Apache 2.0.
    Learn it15 min · 5 steps

    Key ideas

    Audio codec tokens
    A neural codec turns sound into a short stream of discrete codes per frame, so a transformer can predict audio the way a text model predicts words.
    Full duplex dialogue
    Both speakers are modeled at once on separate channels, so the model can overlap, pause and interrupt like people do.
    Text free training
    Learning only from raw speech, with no transcripts, which keeps tone and timing but gives weaker meaning.

    Steps

    1. Read the caramel-1 write up on naklecha.com first, end to end.
    2. Open the GitHub README and study the training data format: tokens.bin holds 16 codebooks per frame for two speakers or 8 for one.
    3. Look at the three stage training commands to see how mono speech is used first and stereo conversation is mixed in later.
    4. Note the 80 ms per frame real time budget that the realtime flag measures against.
    5. Self check: why does a model trained with no text sound natural but often say little that makes sense?
    Try it30 min · 5 steps
    You need: Python 3.12, an NVIDIA GPU, an Apple Silicon Mac or just a CPU; PyTorch and transformers are installed by the setup script; No API key needed

    Steps

    1. Clone the naklecha/caramel repo from GitHub on your own machine.
    2. Run ./setup.sh ./caramel-1 to create the venv and download the weights, then source ./.venv/bin/activate.
    3. Run python generate.py --model ./caramel-1 --seconds 30 --realtime and watch the milliseconds per frame against the 80 ms budget.
    4. Listen to the stereo output and the separate _a, _b and _mix files to hear each speaker.
    5. Record a short stereo or mono clip of your own voice and continue it with python generate.py --model ./caramel-1 --prompt opening.wav --seconds 30 --out talk.wav.

    Small angles to try

    • Compare speed on GPU, Apple GPU and CPU.
    • Feed a mono prompt and hear how speaker B responds to you.
    • Try several prompt lengths and note when the reply starts to drift.
  18. #18Tools & open source

    llm 0.37.2 fixes two ways your OpenAI key could leak to other endpoints

    Simon Willison's LLM command line tool and Python library shipped 0.37.2 with fixes for two security advisories reported by Daniel de Araujo. Custom models in extra-openai-models.yaml that set api_base and api_key_name could send OPENAI_API_KEY to that endpoint when the named key was not stored. Models with responses mode and an api_base could send the stored openai key or OPENAI_API_KEY there when run with the chat completions option.

    Heat · Not measured; no HN story found, only the release notes and Willison's blog
    What people foundIf llm points at any OpenAI compatible endpoint that is not OpenAI, upgrade now. For the first advisory, GHSA-v9h7-jj4g-f862, such models now fail with a No key found error instead of falling back to your OpenAI key; the notes for the second, GHSA-r639-rr7r-x72w, do not state the new error. The general lesson: a silent fallback to a default credential is a leak waiting to happen.
    Learn it15 min · 4 steps

    Key ideas

    api_base
    The URL a model entry sends requests to, which lets one tool talk to many OpenAI compatible providers.
    Credential fallback
    When a tool cannot find the key you named, it quietly uses a default key, which may belong to a different provider.
    GHSA
    A GitHub Security Advisory id that tracks one vulnerability and the versions that fix it.

    Steps

    1. Read the GitHub release notes for 0.37.2 first, then Simon Willison's short note.
    2. Open both advisories by id on GitHub and read which configurations were affected.
    3. Read the LLM setup docs section on keys, which explains that OpenAI models read OPENAI_API_KEY when no key is passed or stored.
    4. Self check: in your own setup, which endpoints would have received your OpenAI key before this release?
    Try it30 min · 5 steps
    You need: Python 3 and the llm package on your own machine; a fake dummy key value, no real API key needed

    Steps

    1. Upgrade with pip install -U llm, or pipx upgrade llm, uv tool upgrade llm or brew upgrade llm depending on how you installed it.
    2. Run llm keys path and llm keys to see which keys are stored and where.
    3. Open extra-openai-models.yaml in the same folder and list every entry that has an api_base, noting whether each has an api_key_name stored.
    4. Start a tiny local HTTP server on localhost that only prints request headers, and add a test model entry whose api_base points at it.
    5. Set OPENAI_API_KEY to an obviously fake string in a throwaway shell, run a prompt against the test model, and confirm the fake key does not show up in the server log and llm reports No key found.

    Small angles to try

    • Repeat on an old version in a separate virtual environment to see the fallback for yourself.
    • Audit other AI CLIs you use for the same default key fallback.
    • Store a dedicated key per provider with llm keys set so no entry relies on a fallback.
  19. #19Trends × tech

    Smart glasses × AI vision: covert filming fears put Meta glasses on notice

    ITmedia looks at AI glasses as regulators abroad react to covert filming. Norway announced on October 5 that it plans a bill restricting their use in public places such as parks and beaches, and the Dutch chain Hans Anders paused sales of Meta's glasses after a consumer group called for a ban. In Japan, Meta Glasses went on sale August 26 from 42,680 yen, while Meta sold more than 7 million AI glasses worldwide in 2025.

    Heat · Not measured; a claimed ITmedia ranking spot could not be confirmed
    What people foundThe only bystander protection the article describes is an LED plus a sound when recording, and ITmedia notes the LED is hard to see in bright places. It also reports AI glasses being used to cheat on TOEIC and university entrance exams in South Korea. For builders of camera wearables, a capture signal that bystanders can actually notice is becoming a regulatory issue, not a nice extra.
    Learn it15 min · 5 steps

    Key ideas

    On device vision
    Running a small model such as BlazeFace on the device itself, so frames can be analyzed without being uploaded.
    Capture indicator
    A light, sound or on screen signal that tells people nearby a camera is recording; it only helps if they can perceive it.
    Privacy filter
    Detecting faces or other people in a frame and blurring them before anything is stored or shared.

    Steps

    1. Read the ITmedia article first to see what regulators are reacting to.
    2. Read the MediaPipe Face Detector guide from Google to see how a small face model runs on phones and in the browser.
    3. Note the three BlazeFace variants, short range, full range and full range sparse, and when each fits glasses versus a selfie camera.
    4. Think through where a blur step must sit in a capture pipeline so raw faces never reach storage.
    5. Self check: why does an LED alone fail as consent, and what could an on device detector add?
    Try it30 min · 5 steps
    You need: A laptop or phone with a camera and a modern browser, your own photos; No API key needed

    Steps

    1. Open the MediaPipe web demo for the face detector linked from the guide, and allow camera access only for this test.
    2. Point it at yourself and watch the boxes and confidence scores update live, then check in your browser's network panel that video frames are not being uploaded.
    3. Switch to image mode and run it on a few of your own photos with only you or consenting friends in them.
    4. Try bright backlight, low light and partly covered faces, and note when detection fails, since a privacy blur would fail at the same moments.
    5. Look at your phone or laptop's own camera indicator in bright and dim light and judge whether a stranger across a room would notice it.

    Small angles to try

    • Sketch a capture flow that blurs every detected face before saving.
    • Compare short range and full range models on a distant face in one of your photos.
    • Write a one paragraph bystander notice your app would show when recording.
  20. #20Trends × tech

    Phones × agents: November Pixel Drop lets Gemini order takeout and rides

    Google's November Pixel Drop began rolling out on November 11 and arrives in stages over several weeks. On the Pixel 10 and Pixel 11 series, Gemini can carry out everyday tasks inside certain apps, such as shopping, ride hailing, takeout orders and restaurant reservations, and users can watch a task's progress and stop it partway. The drop also adds search across photos, email, messages, calendar events and settings on Pixel 6 and later.

    Heat · Not measured
    What people foundLetting the user watch and stop the agent mid task is the human in control pattern phone agents are converging on. Impress Watch does not name the supported apps or say whether the task feature works in Japan. Mobile teams should assume an assistant may soon drive their checkout or booking flow and make each step clearly labeled and easy to confirm.
    Learn it15 min · 5 steps

    Key ideas

    GUI agent
    A model that reads what is on screen, picks a button or field, acts, then reads the screen again until the task is done.
    Accessibility tree
    The structured list of on screen elements with roles and labels that assistive tools, and many agents, read instead of raw pixels.
    Human in the loop checkpoint
    A point, such as before paying, where the agent shows progress and waits so the person can stop or confirm.

    Steps

    1. Read the Impress Watch article first for what the feature does and on which devices.
    2. Read the Android UI Automator testing guide, noting that its waitForStable call watches the accessibility tree, the same structure many agents rely on.
    3. Map a simple food order into screen, action, new screen steps the way an agent would see it.
    4. Mark the step where a careful agent should pause for the user.
    5. Self check: what happens to an agent if your Pay button has no text label?
    Try it30 min · 6 steps
    You need: A desktop Chrome browser and a site or app you own or a local demo page; No API key needed

    Steps

    1. Open your own web app's checkout or booking page in Chrome, or a simple local page you write for this test.
    2. Open DevTools, go to the Accessibility pane and enable the full accessibility tree view.
    3. Walk the flow and write down each control's role and name as an agent would read them.
    4. Flag any icon only buttons, unlabeled inputs or controls whose names change unpredictably.
    5. Fix one label in your own code and confirm the new name appears in the tree.
    6. Write down where your flow should require an explicit user confirmation before money or a booking is committed.

    Small angles to try

    • Repeat on an Android build of your own app using the Layout Inspector in Android Studio.
    • Ask a chat model to plan the clicks from your written tree and see where it guesses wrong.
    • Compare a native app screen and a web view screen of the same flow.

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.