Skip to content
Lambda Brief

Thu, Nov 5, 2026

Siri AI opens app integrations to all developers in preview

Siri AI opens app integrations to all developers in preview, Anthropic releases FlashProve for provable inference, and Go stabilizes its new JSON packages. FlashProve brings receipts for a model’s computation, though its GPU overhead suggests “trust, but verify” should have come with a compute budget.

Siri AI opens app integrations to all developers in preview

Apple opens Siri AI app integrations to all developers in the iOS, iPadOS, macOS, and visionOS 27.3 betas. App Intents expose actions; Skills combine them for complex requests. The Developer Preview requires an Apple Intelligence-enabled device set to a supported language, with App Store submissions opening in spring.

Anthropic releases FlashProve inference-proof prototype

Anthropic’s open-source FlashProve prototype generates proofs for 4-bit Llama 4 Maverick inference at about 110 tokens/s on one H100. It uses about 72 times vLLM’s GPU time per GPU for the same 24-request batch. The proofs cover exact integer computation under greedy decoding; floating-point or sampled outputs may differ.

Also

Go’s new JSON packages reach GA — Go 1.27 makes encoding/json/v2 and encoding/json/jsontext generally available. Benchmarks show up to 10× faster unmarshaling than v1; marshaling stays roughly at parity.

Next.js patches seven vulnerabilities — Next.js 16.4.5 and 15.5.29 fix seven vulnerabilities, including a high-severity cache disclosure that can expose another tenant’s data in certain Cache Components configurations.

StateBench tests long-context state tracking — AI2 introduces StateBench for tracking state changes across long documents. Tested Gemini, Claude, and GPT models scored 13% or lower on the longest inputs for harder tasks.

Compose outlines adaptive layouts and alpha AppState — Compose’s upcoming 1.13 stable release adds Grid and Flexbox layouts. Jetpack AppState, now in alpha, shares observable state across activities, services, and widgets.

Anthropic maps the first prokaryotic vault structure — Anthropic’s biology team reports a prokaryotic vault structure, with Claude producing a first full structural model in 34 hours. The particles’ biological function remains unknown.

Pando cuts OpenAI’s Git push latency — OpenAI’s Pando backend cuts median push latency nearly 80% while preserving Git client compatibility. It delivers nearly a petabyte of Git data daily.

Vercel Security Review enters public beta — Vercel Security Review’s Pro and Enterprise public beta checks every GitHub push once enabled, including force pushes to production. It analyzes source code, not runtime behavior.

Workflow SDK closes duplicate-start race — Workflow SDK’s start() now claims a hook atomically with run admission. Its token prevents duplicate runs, replacing a race-prone check-then-start pattern.

Safari 27.3 beta adds Adaptive Dark Mode — Safari’s beta can darken light websites without exposing the transformation to JavaScript. Apple says almost 40% of page loads come from dark-mode users, but only 10% support it.

Transformers proposes model integration contracts — Transformers proposes pinned model fixtures and reviewed usage baselines. A historical replay flagged a UMT5 regression, fixed in September, that let its decoder attend to future tokens during training.

Cloudflare trims DNS rebuilds — Cloudflare’s targeted DNS builds cut peak Zone Builder memory from 79 GiB to 12 GiB.

CWE-bench v2 opens security-fixing benchmark — CWE-bench v2 opens 100 vulnerability-fixing tasks in real repositories, with deterministic scoring for reproducible agent evaluations.

Palette tops two PresentBench domains — IBM’s Palette fine-tunes a 20B open-weight model for slide generation, ranking first in PresentBench’s Economics and Education domains.

OpenAI announces Pierre acquisition — OpenAI says it will acquire Pierre Computer Company, Inc.

Releases

pydantic-ai 2.66.0 fixes a high-severity flaw in its web-fetch tools. Deeply nested JSON could expand before content limits applied, consuming memory or blocking the event loop.

  • OpenSandbox and RenderSandbox workspaces, Pixeltable tools and memory, Azure Durable Task durability
  • Factory intake routes and skill editing, Knowledge explorer, deploy picker defaults to creating projects
  • allocation tracking for linter rules, allow and allowPattern for react/capitalized-calls
  • WebSocket transport for OpenAI Responses, ConversationStore with revision checks, native mid-conversation system prompts
  • --last-failed test runs, import-map module mocking, interactive snapshot review, failure-only trace view
  • extension update changelogs, faster streaming, custom self-update commands, GPT-6.1 Terra

That’s all for Thursday. Reply with any missed points or tools we should cover. Share this edition with a colleague who’d find it useful.

Share this edition: X · LinkedIn · Email

Reducing entropy

Lambda Brief is a free newsletter on AI and software development.

Sent every Monday and Thursday. Unsubscribe anytime. By subscribing, you agree to our Terms and Privacy Policy.