🧠 AlphaGenome 2 predicts how DNA variants affect specific cell types


AWS gives coding agents a local test cloud, United blocks unauthorized booking bots, and Google experiments with shared mini-apps.
The Merpati Post
Daily AI Briefing

Issue · November 12, 2026

AlphaGenome 2 brings DNA predictions down to individual cell types, while new agent infrastructure, access restrictions, and collaborative tools expose what practical AI still needs.

A faceted purple DNA helix connects an apricot-highlighted rung to the nucleus of a translucent cell, with other cell types behind it and a small gray pigeon at the platform edge.

AI in general

Frontier models, research and policy

2 stories

Only 2 strong recent items found.

Score · 97 / 100

AlphaGenome 2 predicts how DNA variants affect specific cell types

Source: The Verge — November 11, 2026

AI FOR SCIENCE · GENOMICS · GOOGLE DEEPMIND

Google DeepMind’s AlphaGenome 2 predicts how genetic changes affect particular cell types and biological states, rather than relying on broader tissue-level predictions. That matters because the same DNA variant can behave differently in different cells—and those differences can help researchers connect genetic associations to disease mechanisms. It is available for non-commercial research through the AlphaGenome API, with commercial access in a limited private preview.

What matters

  • Trained on more than 100 million individual cell profiles across 68 tissues.
  • DeepMind says it identifies nearly twice as many gene-expression-altering variants as its previous state-of-the-art model; this is a reported research result, not demonstrated clinical benefit.
  • It can make predictions for previously unseen cell types and states without retraining, but its greater computational cost means it does not simply replace the original AlphaGenome.

VerdictSKIM — The main advance is cell-specific resolution; open the article for researcher interviews and the trade-off between detail and compute.

Score · 96 / 100

United blocks unauthorized personal AI agents from booking flights

Source: CBS News — November 11, 2026

AGENTS · TRAVEL · PLATFORM ACCESS

United says it is restricting unauthorized third-party personal AI assistants from making reservations, both on its own digital platforms and indirectly through travel agencies and aggregators. The airline cites credential exposure, unintended purchases or account changes, and regulatory concerns around actions such as check-in attestations. It is a concrete reminder that an agent’s ability to complete a workflow does not mean the service permits it.

What matters

  • United specifically warns about sharing payment information and account credentials, including the risk of unintended MileagePlus-mile use.
  • This is not a ban on ordinary flight-comparison services: an analyst told CBS that United fares remain bookable through established aggregators.
  • For travel-agent builders, authorized integrations and explicit human handoffs are becoming product requirements, not optional polish.

VerdictSKIM — Useful evidence of platform-access friction; the core policy and implications are covered here.

Software engineering

Coding agents, developer tools and infrastructure

3 stories

Score · 99 / 100

AWS releases a local emulator built and tested by coding agents

Source: All Things Distributed / AWS — November 11, 2026

CODING AGENTS · TESTING · INFRASTRUCTURE

AWS Emulator gives developers and coding agents a free, offline environment for testing applications against emulated AWS services. AWS says agents wrote and tested all its code, using public API models and tests checked against real AWS to establish expected behavior. The more useful lesson than “agents wrote it” is the engineering architecture: external ground truth, fidelity tests, and strict merge gates give agents a way to check their work.

What matters

  • Built from more than 400 API models covering more than 200 AWS services; that scope should not be mistaken for complete behavioral parity.
  • Changes pass automated checks, adversarial security-agent reviews, and deterministic restrictions on what agents may modify.
  • Released under CC0, with known behavior mismatches documented in the repository’s “things-that-do-not-work” folder. Keep real-AWS integration tests for critical paths.

VerdictREAD FULL — A concrete account of how to structure agent-driven engineering around verifiable correctness rather than code volume.

Score · 97 / 100

LLM 0.37.2 fixes two paths that could send OpenAI keys to other providers

Source: Simon Willison — November 11, 2026

SECURITY · DEVELOPER TOOLS · API KEYS

Simon Willison released LLM 0.37.2 to fix two configuration-dependent cases where the CLI could accidentally send a configured OpenAI API key to an OpenAI-compatible third-party endpoint. One involved a missing stored key referenced by api_key_name; the other involved running a responses: true model with -o chat_completions 1. This is a narrow disclosure with an immediate action for affected users.

What matters

  • The affected feature is custom OpenAI-compatible endpoint configuration through extra-openai-models.yaml, also managed by llm openai endpoint add.
  • Upgrade to 0.37.2 if you use these configurations; inspect endpoint usage and consider rotating credentials if either condition applied.
  • The release describes potential accidental disclosure, not evidence that every installation—or any particular user—was compromised.

VerdictREAD FULL — Short and actionable if you use LLM with third-party endpoints; otherwise SKIP.

Score · 96 / 100

Google launches a dedicated API for fast, probabilistic multimodal decisions

Source: Google — November 11, 2026

APIS · MODEL ROUTING · MULTIMODAL

The Gemini Decisions API uses the new Gemini Decision 1 model to score bounded options instead of generating free-form answers. It supports yes/no checks, multiple-choice selection, and ordered rubric scoring over text, images, audio, video, and PDFs, with probabilities returned for application logic. This is a potentially useful control layer for routing, policy checks, and evaluation—not a replacement for a general-purpose reasoning model.

What matters

  • Supports up to 1,000 independent questions against shared input, with a maximum context of 256K tokens.
  • Google lists pricing at $0.10 per million input tokens, with no output-token charge.
  • Reported 150–250 ms end-to-end latency applies to the described tests with contexts up to 16K tokens; do not assume it holds at maximum context and question count.
  • Confidence scores enable abstention and human review, but should be validated on your own data before they drive consequential actions.

VerdictREAD FULL — Worth evaluating if you currently use generative calls for classification, routing, or repeated rubric checks.

Design & creative

Creative workflows and user experience

2 stories

Only 2 strong recent items found.

Score · 94 / 100

Google’s Tabletop turns group plans into shared interactive mini-apps

Source: Google Labs — November 11, 2026

COLLABORATIVE UX · GENERATIVE INTERFACES · PROTOTYPING

Tabletop is a Google Labs experiment that generates interactive group spaces from links, files, ideas, and prompts. Examples include trip itineraries, dinner polls, expense splitters, and tournament brackets, which participants can use together in real time through an invitation link. The interesting design direction is moving beyond a private chat response toward a shared, task-specific interface.

What matters

  • Users can prompt @tabletop or begin with a starter template.
  • Access is rolling out through a U.S. waitlist for adults aged 18 and older—not general global availability.
  • The announcement establishes the collaboration model, but does not demonstrate reliability or suitability for sensitive group data.

VerdictSKIM — A short announcement with a useful product pattern; watch for hands-on evidence before adopting it for important workflows.

Score · 94 / 100

Good AI product design starts with the task—and keeps an obvious way to take control

Source: IEEE Spectrum — November 11, 2026

PRODUCT DESIGN · HUMAN CONTROL · UX

In a new interview, IDEO’s Savannah Kunovsky argues that AI should earn its place by making a real task easier, not by supplying a company with an “AI story.” Her examples explain why replacing familiar controls with interpretation-heavy interfaces can make products slower or less trustworthy. She also frames evaluations and behavioral boundaries as core product-definition work, rather than an engineering check added at the end.

What matters

  • During testing of an AI-controlled cooktop, a smoking pan exposed the cost of removing an immediate physical control; the team kept the knobs.
  • Products should distinguish where behavior must be consistent from where generative exploration is acceptable.
  • Undo, editing, manual controls, and feature disablement should be obvious and available when needed—not buried in settings.

VerdictREAD FULL — Practical review criteria for anyone deciding where AI belongs in a product, especially where errors or delays have consequences.

Open-source watch

Projects gaining meaningful traction

3 stories

These are current trending discoveries, not necessarily newly launched projects. Traction figures below come from today’s GitHub Trending inventory.

Project · 01

Improve — shadcn/improve

AI · DEVTOOL · CODING AGENTS

Improve proposes a split-model coding workflow: use your strongest model to audit a codebase and write plans, then have cheaper models execute them. It is interesting as a lightweight way to separate expensive judgment from more routine implementation, rather than paying the same model rate for every step.

What matters

  • The inventory reports 333 stars today.
  • Most relevant to teams already using multiple model tiers; whether it saves money depends on plan quality, rework, and verification.
  • Treat the workflow as an approach to test, not evidence that cheaper execution preserves quality automatically.

VerdictTRY — Run it on a small, well-tested change and compare total cost and review effort with your usual workflow.

Project · 02

Home Assistant MCP — homeassistant-ai/ha-mcp

AGENT · MCP · HOME AUTOMATION

This Python project connects Home Assistant configuration and automation to assistants such as Claude and ChatGPT. It is a practical example of agents interacting with an existing automation platform rather than requiring users to rebuild their environment around a new assistant.

What matters

  • The inventory reports 640 stars today.
  • Relevant if you already run Home Assistant and want assistant-mediated setup or automation.
  • Because actions can affect physical devices, review permissions and preserve manual controls before exposing consequential capabilities.

VerdictWATCH — Useful integration territory, but inspect its permission model before granting access to a live home.

Project · 03

Stitch Skills — google-labs-code/stitch-skills

DESIGN · AGENT SKILLS · MCP

Stitch Skills is a library of Agent Skills designed to work with the Stitch MCP server. It packages design-tool workflows in the Agent Skills format, aiming to make them usable across coding assistants rather than tying them to a single client.

What matters

  • The inventory reports 15 stars today; this is a relevance pick, not a popularity leader.
  • The description names compatibility with assistants including Antigravity, Gemini CLI, Claude Code, and Cursor through the skills standard.
  • Most useful to existing Stitch users; inspect each skill and its tool requirements before installation.

VerdictWATCH — A focused bridge between design tooling and coding agents, worth tracking as the workflows mature.

Editor’s note

Fresh, non-duplicative developments with concrete scientific value, actionable engineering implications, or useful lessons about access, collaboration, and human control.

30-second feedback

How useful was today’s digest?

★★★ Very useful ★★ Somewhat useful ★ Not useful

Your choice opens the short form with your rating filled in. Or share a quick note.


Prefer reading on Telegram? Subscribe to The Merpati Post →

The Merpati Post · Daily AI Briefing

background

Subscribe to The Merpati Post