AI News — Afternoon — 2026-11-11

Test cloud-dependent code without deploying it

• Amazon releases a local tool for testing cloud applications — Developers and coding agents can run AWS-facing tests on a laptop before using a real cloud account.
  • The released source covers APIs for more than 200 AWS services. An agent is a model running a multi-step task with tools; AWS says agents wrote the emulator’s code from public API descriptions and tests checked against real AWS.
  • Linux users can install an RPM or Debian package. macOS and Windows users build a Docker image; AWS also documents an embedded option for Rust tests. The emulator has no additional service fee.
  • It runs at a local HTTPS address with generated credentials. AWS publishes per-service fidelity tests, including failed cases, so teams can check whether their own operations are covered.
  • AWS’s limitations include in-memory state that disappears on restart and uneven support for service features. Anonymous usage and error telemetry is on by default; the repository documents how to turn it off.
  • Caveat: The emulator skips identity and access policies, leaving production security and compatibility unproven by local tests. (HN 447 points · 138 comments · AWS launch, documentation and source)

A viral AI archaeology claim gets corrected

• AI-assisted Maya site claim was a rediscovery, author says — Checking the literature matters as much as finding a promising pattern in images.
  • Pietro Miotti says he had 48 agents examine publicly available laser-mapped terrain. They flagged a hilltop with what looked like a plaza and pyramid, and Mexico’s archaeology agency found no site registered at the supplied location.
  • An update added November 11 says the site had already been mapped in an Estrada-Belli and colleagues paper from 2025. Miotti says the agents missed that paper, so their claim of an unreported discovery was wrong.
  • The agency’s registry check and a literature search answer different questions. The original account describes a proposed field check, not completed archaeological verification.
  • Caveat: The author’s correction identifies prior published mapping; it leaves the agents’ interpretation of individual site features unverified. Standalone summary and HN discussion → (HN 871 points · 317 comments · collected Bluesky correction · author update)

Route bounded decisions without generating answer text

• Google launches fast, fixed-choice answers for model-powered apps — A developer can request routing or classification scores directly instead of generating prose and parsing it.
  • Gemini Decision 1 takes a shared text, image, audio, video or PDF input and scores yes/no, choice or ordered-rubric questions. The API guide documents up to 1,000 independent questions in one request and a 262,144-token input-and-questions limit.
  • Google quotes 150–250 milliseconds per request end to end and $0.10 per million input tokens, with no output-token charge. Those timings and comparisons are Google’s own measurements.
  • Each answer carries a probability or score and a confidence value. The guide recommends an explicit catch-all choice or a separate scope check: choosing the best listed option does not mean any option fits.
  • Access is through Google AI Studio’s Gemini Developer API in beta; the guide says the endpoint does not accept tool-call blocks or conversation history as such.
  • Caveat: Google’s calibration and speed claims are vendor-reported, and a score still needs task-specific thresholds and a human fallback for consequential decisions. (HN 108 points · 35 comments · Techmeme · Google announcement and API guide)