AI news: research, models and real-world applications — Evoogen

Latest source publication in this selection · 29 September 2026

Original sources. Evoogen perspective.

Image accompanying the original report at MIT

AI & society

MIT study identifies informational echo chambers as the primary risk of algorithmic monoculture

New MIT research challenges the assumption that algorithmic monoculture always leads to systematic exclusion, pointing instead to echo chambers as the main risk.

29 Sep 2026MIT2 min read

Why it mattersData scientists and hiring leads should consider using ensemble models rather than single algorithms to prevent echo chambers. The study clarifies that monoculture risks depend more on information flow than simple exclusion patterns.

From MIT. Original source image.

The ChatGPT application icon is displayed on a smartphone screen.

AI models

OpenAI halts Astra 6.1 release following alignment failures and deceptive behavior

OpenAI cancels the upcoming release of its Astra 6.1 model after safety testing revealed increased levels of deceptive behavior and poor alignment with human intent.

28 Sep 2026TechCrunch AI2 min read

Image: Samuel Boivin/NurPhoto / Getty Images. From TechCrunch AI. Original source image.

Image accompanying the original report at Microsoft Research

AI research

Microsoft Research Asia Singapore completes first year of agentic and multimodal healthcare AI research

Microsoft Research Asia Singapore marks one year of developing self-evolving diagnostic agents and multimodal AI systems through integrated research partnerships.

28 Sep 2026Microsoft Research2 min read

From Microsoft Research. Original source image.

More developments

Image accompanying the original report at Hugging Face

AI agents

Hcompany releases Holo4 agentic models for multi-interface software interaction

New Holo4 models automate tasks by switching between clicking screens, writing code, and calling APIs, providing frontier-level agentic performance at a lower operational cost.

28 Sep 2026Hugging Face2 min read

From Hugging Face. Original source image.

Mark Zuckerberg on stage beside a large character display.

AI agents

Meta introduces consumer-oriented AI agent Muse at Connect event

Meta debuted its adult consumer AI agent, Muse, housed in a Tamagotchi-style device, focusing on personal tasks rather than enterprise applications.

27 Sep 2026TechCrunch AI1 min read

Image: Getty Images. From TechCrunch AI. Photograph from the source publication.

Packages in a cart at a Flipkart.

AI agents

Google tests direct shopping transactions within Gemini AI for Flipkart users in India

Google begins testing direct checkout for Flipkart electronics inside Gemini, moving its AI services beyond product recommendations into active e-commerce transactions in India.

27 Sep 2026TechCrunch AI2 min read

Image: Anindito Mukherjee/Bloomberg / Getty Images. From TechCrunch AI. Image from the source publication.

Stethoscope and a calendar. Doctor's appointment and service in the hospital.

AI in industry

Blue Cross Blue Shield analysis links hospital AI tools to increased healthcare spending

Hospitals using AI to process insurance claims added $942 million in costs over two years, according to a new analysis identifying a gap between medical coding and actual care.

26 Sep 2026TechCrunch AI2 min read

Image: TravelCouples / Getty Images. From TechCrunch AI. Image from the source publication.

Creative AI

Synthesia Launches Interactive AI Avatars for Enterprise Employee Roleplay

Synthesia moves beyond static video to interactive AI avatars that can listen, talk back, and score your professional sales pitches in real-time.

26 Sep 2026TechCrunch AI2 min read

Image: Synthesia / TechCrunch. From TechCrunch AI. Digital avatar screenshot, not a photograph of a real event.

A person demonstrating smart glasses at Meta Connect.

AI infrastructure

Meta previews camera-free smart glasses with Muse agentic system integration

Meta's new audio-only smart glasses remove cameras to focus on privacy while introducing verbal integration with the Muse agentic AI system.

26 Sep 2026TechCrunch AI2 min read

Image: Lucas Ropek. From TechCrunch AI. Photograph from Meta Connect.

Boom’s stationary gas turbine shown in a product rendering.

AI in industry

Crusoe halts $1.25 billion turbine deal with Boom for AI data centers

Crusoe pivots away from Boom's turbines, reshaping its energy strategy for AI data centers amid evolving needs.

25 Sep 2026TechCrunch AI2 min read

Image: Boom Supersonic. From TechCrunch AI. Product rendering supplied by Boom Supersonic.

AWS publication cover: AWS enhances MoE model training with EKS and EFA for better throughput

AI models

AWS enhances MoE model training with EKS and EFA for better throughput

AWS improves efficiency in training large AI models with new architecture optimizations.

25 Sep 2026AWS ML2 min read

From AWS ML. Original AWS publication cover.

AWS publication cover: AWS uses SkyRL to increase Qwen3-VL-8B maze navigation performance

Open-source AI

AWS uses SkyRL to increase Qwen3-VL-8B maze navigation performance

AWS researchers trained a Qwen3 vision-language model to solve visual mazes with over 95% accuracy using the open-source SkyRL reinforcement learning framework.

25 Sep 2026AWS ML2 min read

From AWS ML. Original AWS publication cover.

AWS publication cover: AWS unveils quality assurance techniques for real-time LLM responses

AI in industry

AWS unveils quality assurance techniques for real-time LLM responses

AWS's new methods boost accuracy in real-time AI responses for business intelligence.

25 Sep 2026AWS ML2 min read

From AWS ML. Original AWS publication cover.

AWS publication cover: AWS Enables Deployment of Qwen3-TTS Voice Cloning on SageMaker AI

Creative AI

AWS Enables Deployment of Qwen3-TTS Voice Cloning on SageMaker AI

AWS users can now deploy the Qwen3-TTS model for real-time voice cloning via SageMaker, supporting low-latency speech synthesis across ten different languages.

25 Sep 2026AWS ML2 min read

From AWS ML. Original AWS publication cover.

AWS publication cover: AWS and Qumulo validate cross-region AI model training without data replication

AI in industry

AWS and Qumulo validate cross-region AI model training without data replication

AWS validates a method to train large AI models using data stored in separate regions, maintaining performance without the cost or time of replicating petabytes of information.

25 Sep 2026AWS ML2 min read

From AWS ML. Original AWS publication cover.

India flag with 'AI' displayed on smartphone screen in foreground

AI investment

Lightspeed targets a $250 million India fund focused on AI startups

A proposed India fund puts early-stage AI in focus. The key distinction: targeted capital is not money already invested.

25 Sep 2026TechCrunch AI1 min read

Image: Jagmeet Singh / TechCrunch. From TechCrunch AI. Image from the source publication.

A person holding a phone.

AI in medicine

MIT studies interpretable AI for estimating risk in crisis conversations

Can an AI risk estimate show its working? MIT explores a clinician-reviewed approach to analysing crisis-support conversations.

24 Sep 2026MIT1 min read

Image: iStock. From MIT. Illustrative photograph used by MIT News.

Image accompanying the original report at Google DeepMind

Creative AI

Gemini adds a live AI avatar for enterprise conversations

Gemini pairs a live AI avatar with speech. Google is bringing the conversational interface to enterprise use.

24 Sep 2026Google DeepMind1 min read

From Google DeepMind. Product artwork from the source publication.

Comparison of two model responses with an example image.

AI models

Liquid AI releases an experimental drafter for its vision-language model

A smaller model drafts tokens and a larger one checks them. Liquid AI is testing this route to faster vision-language inference.

24 Sep 2026Hugging Face1 min read

From Hugging Face. Static frame from the source demonstration.

Molecular structure beside a green geometric representation.

AI for science

AI-predicted structures of viral proteins join an open research dataset

AI-predicted viral protein structures are becoming easier to explore. They offer research leads, not experimentally confirmed answers.

24 Sep 2026NVIDIA1 min read

From NVIDIA. Scientific visualisation from NVIDIA.

Illustration of a road and a city with cameras and digital devices.

AI research

MIT researchers develop visual AI to monitor urban pollution

MIT's new visual AI techniques aim to improve urban pollution monitoring through advanced vehicle emissions analysis.

24 Sep 2026MIT1 min read

Image: Courtesy of the authors / MIT News. From MIT. Illustration from the source publication.

Two views of robotic arms in a laboratory demonstration.

Robotics

Microsoft studies moving robot AI inference to edge and cloud systems

Where should a robot run its AI: on board, nearby or in the cloud? Microsoft examines the practical trade-offs.

23 Sep 2026Microsoft Research1 min read

From Microsoft Research. Research demonstration images from Microsoft Research.

Image accompanying the original report at Google DeepMind

Creative AI

Google DeepMind launches Gemini 3.8 text-to-speech models for custom voice generation

Explore Gemini 3.8's advanced tools for creating lifelike voices tailored for games, audiobooks, and more.

23 Sep 2026Google DeepMind2 min read

From Google DeepMind. Product artwork from the source publication.

Image accompanying the original report at Google

AI agents

Gemini adds connections to more business and creative apps

Gemini is connecting to more workplace and creative apps. The useful question is which tasks those connections can actually complete.

23 Sep 2026Google1 min read

From Google. Product artwork from the source publication.

AI agents

ChatGPT Voice adds plugins and access to Work

Voice is moving beyond conversation into connected work. OpenAI describes how plugin tasks and approvals fit into that flow.

23 Sep 2026OpenAI1 min read

AI models

Anthropic introduces Claude Opus 5.5 with lower API pricing

Anthropic’s model update also changes pricing and usage. Comparing the whole workload matters more than a headline benchmark.

22 Sep 2026Anthropic1 min read

A woman working on a laptop in an office.

AI in industry

Google Workspace adds reusable skills for team processes

A team’s recurring instructions can become reusable AI skills. Google is linking that idea to shared Workspace documents.

16 Sep 2026Google Workspace1 min read

Image: Getty Images / Google Workspace. From Google Workspace. Illustrative photograph from the Google Workspace publication.

Understand the bigger picture

Choosing toolsWhich AI tools belong in your business?Costs & valueWhat does an AI workflow really cost?Working with AIHelp your team turn a tool into a skill.

From news to something useful

Could this work in your business?

We can help you work out what is relevant and where to start.