Thursday, November 12, 2026

Google's New Decisions API

Google's New Decisions API

Today’s Overview

Good morning, Google is turning a familiar AI task into a fast choice among fixed options, while DeepMind is aiming its genetics model at the cells where disease unfolds. Meanwhile, a leadership question is hanging over Nvidia, and researchers are finding ways to make language models learn from textbooks rather than rely on preloaded knowledge. Let's dive in.

Top Stories

Google's Gemini Decisions API turns questions into choices

Google has launched Gemini Decisions, a Gemini API endpoint for bounded judgments such as routing, policy checks, and rubric scoring. Its gemini-decision-1 model scores candidate answers in one pass, returning confidence values rather than generated prose, and is available in Google AI Studio at $0.10 per million input tokens with no output charge.

  • A single request can cover up to 1,000 questions about a shared input.
  • The endpoint accepts text, images, audio, video, and PDFs with inputs up to 256K tokens.
  • Google says Gemini Notebook's routing cut everyday-query latency by up to 40% while reserving deeper reasoning models for complex tasks.

AlphaGenome 2 sharpens DNA predictions by cell type

AlphaGenome 2 predicts how DNA variants affect specific human cell types and states, helping researchers investigate the cellular drivers of disease. It identifies nearly twice as many variants that change gene expression as its predecessor and can analyze new cell states without retraining; it is available for non-commercial research through the AlphaGenome API.

  • Its training data includes more than 107 million individual cell profiles across 68 tissues.
  • On the TenK10K immune-cell dataset, it identified 30% of causal variants, versus 16% for the original model fine-tuned on single-cell data and other specialized tools.
  • In data from a COVID-19 vaccine study, it found a severe-COVID-linked variant had a stronger effect in activated immune T cells than in resting cells.

Nvidia weighs Huang's future after new allegation

Nvidia directors discussed whether to remove CEO Jensen Huang after a second alleged relationship, involving health care vice president Kimberly Powell, who reports to him. Shares fell nearly 7% in after-hours trading as the leadership question intensified.

  • An independent panel is investigating both relationships with Wachtell, Lipton, Rosen & Katz handling the probe.
  • Nvidia has no obvious successor to its longtime CEO and public face.
  • The investigation could raise questions about what the board knew and when it knew it.

Research & Analysis

Teaching language models to learn from textbooks

A new reinforcement-learning setup asks models to study textbooks as they solve proof problems, rather than relying on a fixed context or textbook-free reasoning. With its BookNavigator interface and fine-grained rewards, the method raised Intern-S2-Preview-35B's proof accuracy from 51.5% to 79.5% and also improved performance on other benchmarks.

  • The researchers assembled 5,183 proof problems from 158 textbooks for the study task.
  • BookNavigator lets the model inspect a table of contents, search sections, and read them as its reasoning develops.
  • Evaluation included books and subjects held out from training to test whether the study strategy transfers.

A structured textbook corpus boosts proof training

TOME-Math converts textbooks into structured proof supervision, retaining chapter organization and links between propositions and proofs instead of flattening everything into isolated question-answer pairs. Fine-tuning on the corpus raised the unweighted mean across six evaluations by 16.7 percentage points and beat all three public instruction-data baselines in aggregate.

  • The dataset contains 380,292 records from 2,118 textbooks spanning 29 mathematical subjects.
  • Its records include 174,431 Gap QA pairs alongside 205,861 Proof QA pairs.
  • Structure-guided retrieval raised proof-generation pass@1 from 13.25% to 17.38% under a matched context budget.

Sparse reading helps agents avoid content overload

SparseRead is a reading-control layer for web-facing agents that limits how much external content reaches the model without changing the model itself. Across six models and five scenarios, it reduced token use and wall time while preserving or improving task quality in 97% of settings.

  • In long-document QA, supporting evidence covered a median of 3.17% of source tokens per question in the LooGLE analysis.
  • In paired high-sparsity runs, full-read agents used up to 17.3 times more tokens without improving task scores.
  • Forced sparse reading increased token use by 309% for global statistics and 210% for a rule-coding task, illustrating why the system keeps a full-view fallback.

SplatAnything normalizes 3D Gaussian spacing

SplatAnything addresses a scaling mismatch in feed-forward 3D Gaussian Splatting: normalizing overall radial distance does not stabilize spacing between neighboring Gaussian centers. Its Gaussian Spacing Normalization method uses a robust scene-level spacing statistic, and the paper's benchmark reports strong results under large changes in scene scale and viewpoint.

  • Across scene categories, the prior normalization left median Gaussian spacing varying by up to 4.5 times while GSNorm canonicalizes it.
  • GSNorm derives its scale from the median of depth divided by the geometric mean of the camera focal lengths across valid observations.
  • The model was trained for 20,000 iterations on 64 NVIDIA H20 GPUs using between 4 and 32 input views.

Trending AI Tools

  • ElevenLabs Scribe v3 Speech-to-text models for 102 languages, with real-time speaker labeling in Scribe v3 Live.

  • OpenAI MCP extensions TypeScript and Python packages add ChatGPT-specific features to MCP servers and apps.

  • Grok Bot 1.0 Adds shareable Team Bots, Slack and Teams support, and participation in Meet and Zoom calls.

Quick Hits

  • ElevenAI desktop app is designed to help users amplify their work with agents.

  • Grok 4.8 is set for a soft rollout this week after a rebuild replaced JAX with C++.

  • LongCat-2.5 has open MIT-licensed weights and is designed for long-horizon agent work across text, images, and video.

  • Gemini Robotics 2.5 reportedly sorted car parts with 96% accuracy.

  • Runway Praxis-1 is a 27-billion-parameter open-weight model under Apache 2.0, with a reported 46.2% success rate across 1,200 RoboLab-120 episodes.

  • OpenAI's third-quarter figures reportedly show about $10 billion in revenue and a $10.9 billion adjusted operating loss, based on leaked documents.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.