AI Today: NVIDIA Nemotron and Gemini 4 Carbon

State of knowledge as of 2026-10-29. Source material collected on 2026-10-29.

October 28 brought intense updates to the artificial intelligence sector. NVIDIA released the open speech model Nemotron 3.5 Transcribe, while Google DeepMind introduced the cybersecurity-oriented Gemini 4 Carbon model. Furthermore, Google launched the Connected Apps platform for developers, XREAL presented AR glasses powered by Android XR, and OpenAI experts published an analysis regarding the impact of AI technology on the organizational structure of modern enterprises.

Key Takeaways (TL;DR)

• NVIDIA Nemotron 3.5 Transcribe: An open model for transcription and translation supporting 40 languages with diarization for up to eight speakers.
• Google Gemini 4 Carbon: A new model specializing in cybersecurity and engineering, demonstrating twice the effectiveness in secure code generation.
• Google Connected Apps Platform: A tool for developers integrating applications with Google Search and the Gemini system via MCP protocols.
• XREAL AURA Glasses: A wired XR device powered by Android XR, integrating the Gemini assistant with a spatial interface view.
• The Future of Organizations in the Age of AI: An OpenAI analysis outlining three scenarios for artificial intelligence deployment within corporate structures as employee support.

NVIDIA Nemotron 3.5 Transcribe

On October 28, NVIDIA unveiled Nemotron 3.5 Transcribe, an open speech-to-text model offering diarization, translation, and question-answering capabilities for audio files.

The model supports 40 languages, 31 of which can be translated to or from English, and in 13 languages, the system can assign utterances to specific individuals. Nemotron 3.5 Transcribe processes recordings up to one hour long and can distinguish up to eight speakers simultaneously. The tool ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for the English language (with an average error rate of 5.33%) and leads diarization benchmarks within David AI and VoiceArena systems.

Alongside the model, NVIDIA released Granary v2, a multilingual dataset and processing pipeline that allows for model adaptation to specific domain vocabulary. In a medical case study, model fine-tuning reduced the Word Error Rate (WER) by 45.8% compared to prompt-only usage. NVIDIA announced that quantized versions of the model will soon be released, enabling local execution on hardware such as RTX Spark, DGX Spark, and Jetson AGX Thor. Users can guide model outputs using natural language, defining jargon formatting or number notation styles, for example.

For AI builders: Teams working on meeting analysis or medical transcription can test the model for recognizing specific terminology using context prompts. The announcement does not specify whether the model offers full support for Polish, what the computational costs are using specific GPUs, or whether the license allows for commercial use of output data without restrictions. These details require verification in the technical documentation and model card before starting integration.

Google DeepMind Introduces Gemini 4 Carbon

Google DeepMind announced on October 28 the introduction of Gemini 4 Carbon, which represents the company's strongest unit specialized in cybersecurity tasks and long-term engineering processes. The model debuted just two weeks after the release of the Gemini 4 Argon version.

Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 leaderboard with a score of 59. In specialized benchmarks, the model achieved 88.2% effectiveness on SEC-bench Pro (vulnerability detection), 54.8% on ExploitGym (exploitation), and 71.0% on CWE-bench (patch generation). According to company data, compared to the Argon model, Carbon offers more than a double increase in effectiveness when generating secure code. In internal tests and environments such as the Chrome team, the model demonstrated higher efficiency in creating high-end patches than other frontier models.

The model is designed for continuous operation, enabling it to solve multi-step engineering problems that require days of work. In the area of web development, Carbon advanced to fourth place on the WebDev Arena leaderboard with a score of 1768 Elo points.

The company is conducting a phased rollout, collaborating with the US government as part of a voluntary pre-release testing process. Starting this week, a selected group of cybersecurity experts will receive access to the model, followed by broader access through the Fairwind program. The cost of using Gemini 4 Carbon is set at $2 per million input tokens and $10 per million output tokens, with cached tokens receiving a 95% discount relative to the input price.

For AI builders: Specialists dealing with security audits can test the model's performance in software vulnerability detection and remediation processes. The source does not provide information regarding the model's regional availability, support for languages other than English, or full technical documentation regarding API rate limits.

Google Connected Apps Platform

On October 28, Google released the Connected Apps platform, enabling developers to build and manage applications that integrate with Google Search and the Gemini app. The goal of the solution is to create a unified space where creators can deploy tools that allow users to handle everyday tasks, from ordering groceries to creating playlists.

Within the system, developers integrate with Model Context Protocol (MCP) servers or Android MCP using AppFunctions APIs. Next, using the Google Connected Console, they can register, test, and submit their apps for review. During registration, creators specify whether the solution should be available in Search, the Gemini app, or both. The platform also enables tracking traffic data, latency, and error metrics, alongside managing brand identity, including logo verification and account authentication.

Currently, platform access is opening in phases. Google has already launched over 50 apps in areas such as productivity, event tickets, health, and fitness. Partners such as Adobe, Experian, Localiza, and Peloton use these solutions. In the coming weeks, expansion of these integrations to Search is planned, along with the introduction of a new Calm app. Interested developers can register for the preview and explore documentation at developers.google.com/connected-apps.

For AI builders: Developers of apps integrated with Search or Gemini can register their tools in the console and check technical deployment parameters. The announcement does not specify whether the platform imposes API usage limits, what the maintenance costs for MCP servers are, or if there are regional restrictions regarding app publishing in various markets — this must be verified in the technical documentation before starting work.

XREAL AURA Glasses with Android XR

On October 28, XREAL, in collaboration with Google, launched XREAL AURA, the first wired extended reality (XR) glasses powered by Android XR and the Gemini assistant. The device allows users to display applications in space while maintaining visibility of their surroundings through transparent lenses.

The glasses weigh under 95 grams because computation is handled by an external puck connected via a cable, which houses the battery and the Snapdragon Reality Elite platform. The device offers a 70-degree field of view. Users can open multiple app windows simultaneously, manage them via a Bluetooth keyboard and mouse, or use the touchpad on the computing module. The Wired Mirroring feature allows connecting a laptop via a USB-C cable and displaying its screen as a movable window. The Gemini assistant integrated into the system understands the video feed captured by the device and screen contents, enabling conversation regarding displayed content or surroundings. In situations requiring focus, the lenses can dim the image to isolate the user from the outside world. Sales have started in the United States, Canada, and Korea, while the Japan launch is scheduled for November 2.

For AI builders: Teams working on spatial interfaces can test Wired Mirroring support and how Gemini interprets the shared view in real time. The material does not specify device availability in European Union countries, licensing details for XR app developers, or exact pricing for business customers.

The Future of the Organization in the Age of AI

OpenAI analysts Hemanth Asirvatham and Elliott Mokski published an article on October 28 analyzing the impact of artificial intelligence on corporate organizational structures and team management methods. The authors point out that the traditional corporate ladder is undergoing transformation because AI systems are beginning to function as digital subordinates performing routine tasks.

The article outlines three main scenarios for AI deployment in enterprises. In the first model, referred to as work "below us," employees keep their jobs but gain support through an AI layer, meaning every worker becomes a manager overseeing their own agents. The second scenario assumes a "beside us" situation, where artificial intelligence functions as a teammate directly collaborating with human managers and performing an increasingly broad range of tasks. The third variant, "between us," positions AI as an intermediary between people at various hierarchical levels, responsible for coordinating actions and monitoring work.

The authors emphasize that none of these visions involve full automation, but rather a transformation of the economy toward hybrid firms. They note that technology adoption is not an instantaneous process; initially, AI takes over entry-level tasks, eventually expanding into managerial areas as model capabilities grow. Currently, employees already use systems capable of writing notes, preparing memes, or offering technical advice, though they still require human supervision.

For AI builders: Team managers can consider mapping processes in their departments to isolate routine tasks that can now be delegated to AI agents. The publication does not specify the maintenance costs of such infrastructure or the security requirements for data processed by systems within the corporate structure, which requires verification prior to deployment.

Frequently Asked Questions (FAQ)

What capabilities does the NVIDIA Nemotron 3.5 Transcribe model offer?

The model enables speech-to-text conversion, utterance diarization, translation, and question-answering for audio files. It supports 40 languages, processes recordings up to an hour long, and can distinguish up to eight speakers simultaneously, holding top positions on open-weight model leaderboards.

What are XREAL AURA glasses used for and how do they work?

XREAL AURA are AR glasses powered by Android XR that display applications in space while keeping the surroundings visible. The device uses an external computing module with Snapdragon Reality Elite and the Gemini assistant, which understands the equipment's video feed and the contents of displayed screens.

Sources

• NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt — 2026-10-28
• Gemini 4 Carbon: Our next frontier model — 2026-10-28
• Build and manage Connected Apps in one place — 2026-10-28
• XREAL AURA: Wired XR glasses to watch, work, and play — 2026-10-28
• The future of the firm — 2026-10-28

--- Content automatically prepared by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Pre-human-editing version.