AI Today: Gemini 3.8 Live and Claude AI

The current state of knowledge as of 2026-09-25. Source material gathered between 2026-09-24 and 2026-09-25.

Today's edition covers a wide spectrum of updates in the tech world. Google is introducing Gemini 3.8 Live with avatars and new external integrations. Anthropic used the Claude model to discover the ART enzymatic system, and NVIDIA together with Google DeepMind shared data on viral protein structures. Additionally, we look at the work of validation engineers at NVIDIA who ensure the reliability of server infrastructure supporting the development of artificial intelligence systems.

Key Takeaways Today (TL;DR)

• Gemini 3.8 Live: Google introduced interactive avatars supporting 97 languages with lip-sync and advanced tool calling.
• Claude and the ART system: The Claude model helped identify the ART enzymatic system in bacteriophages by analyzing 200,000 reverse transcriptases.
• Viral Protein Structures: NVIDIA and Google DeepMind published data on 2800 viral protein structures generated using the AlphaFold2 model.
• Gemini Integrations: Google enabled Gemini collaboration with external tools such as Airtable, Adobe, monday.com, and Squarespace.
• NVIDIA Validation Engineering: Sakeena Fiza explains the process of testing server hardware under real-world conditions before deploying GPUs to production.

Google Introduces Gemini 3.8 Live with Live Avatar

According to an announcement on September 24, Google presented Gemini 3.8 Live with the new Live Avatar feature, which introduces a visual presence in conversational model interactions. This solution combines a conversational model with low-latency video generation, allowing conversations with a dynamic visual character.

The Live Avatar feature supports 97 languages with native lip-sync and facial expression matching, adapting to spoken content. The system processes audio and video signals in real time to create a smoother communication flow. Gemini 3.8 Live also utilizes an asynchronous tool calling mechanism, enabling the model to perform background tasks such as data retrieval or process handling without interrupting the ongoing conversation.

Regarding personalization, enterprises can use a library of pre-built characters or create their own custom avatars based on a provided reference image. Access to creating custom characters is currently restricted via an allowlisting procedure within Gemini Enterprise. For security, all content generated by the model is automatically watermarked with an invisible SynthID watermark to facilitate the detection of AI-generated materials.

For AI builders: Customer support teams can check the possibility of integrating their own brand identity within the video interface. The announcement does not specify the costs of using the Live Avatar feature or detail API bandwidth limits, which should be verified in the technical documentation before production deployment.

Anthropic Discovers Novel Enzymatic System with Claude

The research team from Anthropic reported on September 23 that the Claude model played a key role in discovering a novel enzymatic system exhibiting CRISPR-like properties. This system, named ART (array-associated reverse transcriptases), was identified in bacteriophages by analyzing DNA sequence databases.

The research process took 21 hours, during which approximately 950 AI agents processed 210 million tokens. Claude searched a database of over 200,000 reverse transcriptases (RT), yielding 3,500 candidates for new systems, from which the 20 most promising were ultimately isolated. The model independently identified a repeating DNA sequence pattern occurring next to a previously unknown transcriptase gene. Scientists verified this discovery through laboratory analysis, demonstrating that the system possesses a set of characteristics typical of enzymes capable of cutting, copying, and pasting DNA. The research team emphasizes that Claude also assisted in interpreting biochemical and structural data obtained in the lab. The research results were shared as a pre-print publication.

For AI builders: Research teams dealing with biological data analysis can use an agent-based approach to screen large DNA sequence databases to identify unknown proteins. The announcement does not specify the operational costs of using such a large number of parallel agents or the licensing requirements for using tools like Claude Science and Claude Code in commercial projects—these issues require separate clarification before implementing a similar methodology.

NVIDIA and Google DeepMind Publish Viral Protein Structure Data

According to information from September 24, NVIDIA joined an international research coalition to share open 3D protein complex structure data for over 2,800 viruses. This initiative aims to support the scientific community's preparations for potential future pandemics by providing the knowledge necessary for rapid responses to biological threats.

The structures included in the shared dataset were generated using the AlphaFold2 model from Google DeepMind, with optimization provided by the NVIDIA BioNeMo Inference Runtime. Using this tool allowed scaling inference to thousands of viral proteomes. Approximately 30% of the protein interactions added to the database are entirely new to science and do not appear in the main Protein Data Bank repository.

The project involves broad collaboration among institutions, including EMBL-EBI, the Coalition for Epidemic Preparedness Innovations, and several universities. The data is publicly available via the AlphaFold Database, which after the update contains over 260 million protein predictions. NVIDIA also released the BioNeMo Structure Prediction Pipeline—a GPU-accelerated workflow that allows researchers to independently transform protein sequences into predicted 3D structures. This method is intended as an alternative to traditional, expensive, and time-consuming experimental studies, allowing results to be obtained in minutes.

For AI builders: Researchers working in computational biology can use the BioNeMo Structure Prediction Pipeline to generate their own protein structure predictions. The source does not specify the maintenance costs of the GPU-based computing infrastructure needed to run the pipeline, model precision requirements in the context of specific clinical studies, or licensing restrictions for commercial use of the results.

Google Integrates External Applications with Gemini

Google announced on September 23 that it is introducing a new group of connected tools to the Gemini app. Users can now manage projects, design visual assets, and plan activities without switching browser tabs.

The new integrations are divided into three categories. In the productivity area, collaboration was introduced with Airtable, Linear, monday.com, PandaDoc, Wispr AI, and Zoho. In the creative section, Adobe, Picsart, Squarespace, and Webflow were added. Meanwhile, the lifestyle category includes apartments.com, Experian, Peloton, and SeatGeek. To use the new features, users must connect selected apps in Gemini settings or invoke them in the chat using an at-sign or a direct query.

For AI builders: Teams using external project management tools can test invoking the functions of these platforms directly from the chat interface. The announcement does not specify whether these integrations work in all languages, what the API availability scope is across regions, or whether paid premium accounts in external services are required—these details must be verified before production deployment.

Sakeena Fiza and Validation Engineering at NVIDIA

According to an article published on September 23, Sakeena Fiza serves as a validation engineer at NVIDIA, focusing on bug detection and ensuring hardware reliability before mass production. Her work involves testing systems under near-real-world conditions, often before a product becomes widely known. Fiza describes the validation process as being the "first customer" who checks equipment durability, ranging from individual motherboards to entire server racks in clusters.

The engineering team in the data center systems lab investigates a wide spectrum of issues—from microscopic flaws to rack-scale challenges such as high-frequency signal integrity, thermal margins, and power integrity. This process involves collaboration among software, firmware, mechanical, and thermal design specialists. Fiza emphasizes that a single system can consist of hundreds of thousands of components, and each must operate correctly within a complex "AI factory" configuration. The work of a validation engineer requires diagnosing failure causes, replicating the conditions of their occurrence, and eliminating mechanical or electrical variables to ensure the stable operation of GPUs, such as the Rubin GPU, in target customer infrastructures.

For AI builders: Systems engineers designing server infrastructure can deploy intensive component testing procedures prior to full integration to identify hardware bottlenecks. The article does not provide detailed technical parameters of the tested systems, sample selection methodologies, or environmental requirements for external data centers—these details must be obtained directly from the manufacturer's technical documentation.

Frequently Asked Questions (FAQ)

What features does the new Live Avatar offer in Gemini 3.8 Live?

Live Avatar enables interaction with a dynamic visual character that supports 97 languages with lip-sync. The system processes audio and video signals in real time. Enterprises can use pre-built character libraries or create their own avatars based on reference images using accounts within Gemini Enterprise.

How did Claude assist in the research on the ART system?

Claude searched a database of over 200,000 reverse transcriptases in 21 hours. The model independently identified a repeating DNA sequence pattern, which allowed the isolation of the 20 most promising candidates for new enzymatic systems. Ultimately, scientists confirmed in the lab that the system possesses features capable of cutting, copying, and pasting DNA.

Sources

• Introducing Gemini 3.8 Live with Live Avatar — 2026-09-24
• Claude discovers a novel enzyme system with CRISPR-like repeats — 2026-09-23
• How Open Science Can Help Researchers Prepare for the Next Pandemic — 2026-09-24
• A new wave of Connected Apps is rolling out to Gemini. — 2026-09-23
• Sakeena Fiza Helps NVIDIA Hardware Succeed at Scale — 2026-09-23

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