Thursday, October 29, 2026

Gemini Carbon Raises The Stakes

Gemini Carbon Raises The Stakes

Today’s Overview

Good morning, Google is pushing Gemini toward lower-carbon AI, Meta is testing a gentler kind of AI companion with a pixel pet, and OpenAI is making unusually bold scientific claims that researchers still need to unpack. There is also a fresh wave of agent infrastructure, benchmarks, and on-device models. Let's dive in.

Top Stories

Google introduces Gemini 4 Carbon

Google introduced Gemini 4 Carbon, a new model framed around stronger AI performance with improved energy efficiency. The company says the model advances natural language processing while reducing its carbon footprint, positioning it as part of a broader push toward more sustainable AI development.

  • Google describes Carbon as its strongest model with state-of-the-art cybersecurity and long-horizon capabilities.
  • The model ranks first on Artificial Analysis Intelligence Index v5.0 with a score of 59 and leads several cybersecurity benchmarks.
  • Carbon is being rolled out first to cyber defenders and trusted partners before broader developer, business, and consumer availability.

Meta launches Waddle World AI pet app

Meta launched Waddle World, a virtual AI pet app in the Philippines for iOS and Android. The app blends pet interaction, mini-games, and journaling while avoiding the tone of a productivity tool or traditional chatbot.

  • The app lets users pick, grow, feed, and play with a virtual pet.
  • Meta's description says the app is for users older than 18 and is not intended to be a habit-building experience.
  • Journal entries can be created by voice or text then rewritten by the pet in its own voice with a doodle.

NVIDIA releases Nemotron 3.5 Transcribe and Granary v2

NVIDIA released Nemotron 3.5 Transcribe, an open-weight multilingual model for transcription, translation, speaker attribution, and audio Q&A. Granary v2 adds reusable data pipelines and fine-tuning recipes so teams can adapt the system to specialized vocabularies, accents, and domains.

  • Nemotron 3.5 Transcribe supports transcription across 40 languages and translation between English and 31 languages.
  • The model can handle recordings up to one hour and label as many as eight speakers.
  • NVIDIA says quantized versions are coming for local deployment on RTX Spark, DGX Spark, and Jetson AGX Thor as well as larger Blackwell GPU setups.

Research & Analysis

OmniShowcase improves multi-view video consistency

OmniShowcase targets subject-consistent video generation from multiple reference images. The method combines geometry-aware conditioning with synthetic training videos, helping the model preserve identity and 3D structure as cameras move around a subject.

  • The paper was listed as Hugging Face's number two paper of the day after being submitted on October 29.
  • Its training setup includes full 360-degree orbits to expose models to large viewpoint changes.
  • The authors provide links to an arXiv page, project page, and GitHub repository from the paper listing.

OpenAI claims major AI-generated scientific proofs

OpenAI says its models produced rigorous existence proofs for interacting quantum field theories in four dimensions, including quantum chromodynamics. The claim is potentially enormous, but the physics materials have not been publicly released yet, leaving outside researchers unable to verify the result for now.

  • OpenAI describes the outputs as raw research data that need expert communities to turn them into scientific understanding.
  • The math result is being released with both a natural language writeup and Lean proof while the physics artifacts are being held back for now.
  • OpenAI says the collaborations follow principles of transparency, no academic credit for OpenAI, recognition for expert writeups, and independent participation by working-group members.

Artificial Analysis updates Intelligence Index v5

Artificial Analysis released Intelligence Index v5, a benchmark update designed to make model rankings harder to game. The update moves 60% of the index to private test sets and adds more realistic coding and science agent tasks.

  • The index is split evenly across knowledge work, coding, science, and general knowledge with each category weighted at 25%.
  • FrontierCode 2.0 uses 132 open-source repository tasks with code changes evaluated through tests and rubric-based judging.
  • Terminal-Bench-Science includes 70 terminal-based research workflow tasks across five domains, scored by pass or fail test results.

TAO schedules tool-using AI agents more efficiently

TAO is a scheduling framework for agentic services that alternate between LLM inference on clustered GPUs and external tools on distributed CPUs. It accounts for tool readiness, cache value, and token footprint to reduce latency and improve throughput across the evaluated workloads.

  • TAO uses tool-boundary interruption to pause sessions during external tool calls and free GPU capacity.
  • Its session valuation prioritizes time-decayed KV-cache credit alongside tool-return readiness and token footprint.
  • The paper reports gains across 12 model-workload-concurrency settings against the best evaluated state-of-the-art baseline for each metric.

Trending AI Tools

  • GPT-6.1 Terra A mid-tier OpenAI model for visual reasoning and computer use, available in ChatGPT Work, Codex, and the API.

  • Microsoft Copilot Harness A managed Microsoft Foundry system for building multi-model agents that reason, use tools, and complete long-running tasks.

  • Liquid AI LFM3 Six open-weight 2B-class models built for local deployment across laptops, phones, Raspberry Pi devices, and multimodal workflows.

Quick Hits

  • Anthropic spending tools give enterprise customers reporting on productivity, costs, and recommendations for Claude and related products.

  • OpenAI election war room is pulling hundreds of staffers into monitoring ChatGPT misuse ahead of next week's U.S. midterm elections.

  • Hunger beats tenure argues that AI is helping junior engineers move faster by making execution cheaper and bold prototypes easier to ship.

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.