AI Austria monthly: September'25 Update

Autumn season is not just event season but also marks the kickoff for our WhatsApp Community.

Matthias Grabner, Ting Wasner-Lian, Clemens Wasner, and Ruben Hetfleisch

Sep 12, 2025

Our upcoming 2nd RL bootcamp, State of AI report 2025, AI use in federal administration and the launch of our latest community.

🚀 AI Austria news

💬 Launch of the AI Austria WhatsApp Community

We are celebrating a double premiere: with Fatemeh-Shirin Asgari as our new Community Manager and the launch of our very first WhatsApp Community. The AI AUSTRIA WhatsApp Community was recently created and already brings together more than 300 members. The goal is to establish a vibrant platform where experts and enthusiasts can exchange ideas on AI, innovation, and new technologies.

Currently, there are three groups: General Discussion, Jobs & Cooperation, and Events. » We warmly invite you to join the conversation

⌛️ 2nd International Reinforcement Learning Bootcamp

Our the 2nd International RL Bootcamp in Salzburg, September 17–19, 2025, is already fully booked. » If you want to watch the talks online head over to the YoutTube channel and join remotely.

🏰 Join us for the next AI Stammtisch in Salzburg

Together with aws Austria Wirtschaftsservice AI Adoption, we’re hosting the next event in our AI Stammtisch series in Salzburg at Privatuniversität Schloss Seeburg. This is a great opportunity to engage with Austria’s AI community, exchange ideas, and explore collaboration opportunities in an open and informal setting. Whether you’re an AI professional, researcher, or simply curious about the field, this event is the perfect chance to connect with like-minded individuals and contribute to the conversation. » interested? registrations are open

🇦🇹 Latest developments

🔎 Report: Artificial Intelligence, Market Power and India in a Multipolar World

A joint report by extensity AI, the Indic Pacific Legal Research, the Indian Society of Artificial Intelligence and Law (ISAIL), and Schematise LDA dives deep into the critical issues of our time: vendor lock-ins, the real human cost of AI, and the appropriation of the digital commons. It was a fascinating process, blending legal informatics, technology law, and my own perspective from the world of neuro-symbolic AI.

💡 Study: Are AI Agents interacting with Online Ads?

How do AI agents make decisions, and what signals guide them? A joint study by the University of Applied Sciences Upper Austria and 506.ai reveals that chatbots and agents are influenced by advertising signals, demonstrating a clear preference for structured text over images.
A recent deepdive by the FT highlighted why this matters: Tech giants such as OpenAI, Google, Microsoft and Perplexity are competing to make agent-driven shopping the norm, at the same time traditional SEO is giving way to AEO (Agent Experience Optimization) and GEO (Generative Engine Optimization) as new battlegrounds for visibility. » Read the full study here

✨ Generative AI News ✨

👉 Apple Puts a Video Expert on Your Device
Apple released FastVLM and MobileCLIP2, new on-device AI models that enable real-time video captioning directly in a browser, enhancing privacy by processing video locally without sending data to the cloud.

👉 OpenAI May Report User Chats to Police
OpenAI confirmed it scans ChatGPT conversations for threats of imminent harm and may report flagged content to law enforcement after human review, sparking significant privacy concerns.

👉 AI Music Videos Becomes a Reality
A new creative pipeline combining Google's "Nano Banana" image generator and the Kling 2.1 video model allows creators to produce fully automated music videos from simple text prompts.

👉 OpenAI's Bumpy GPT-5 Rollout
OpenAI faced a user backlash after its GPT-5 launch removed access to older models, prompting the company to quickly reverse course by restoring model selection and increasing usage limits.

👉 A Deep Dive into High-Throughput LLM Inference
A widely praised, comprehensive guide to vLLM internals has been released by Aleksa Gordić, detailing advanced techniques for high-throughput LLM inference such as paged attention, continuous batching, and various scaling topologies.

👉 Meituan's Massive 560B Open-Source Model
Meituan open-sourced LongCat-Flash-Chat, a massive 560B parameter Mixture-of-Experts (MoE) model detailed in a candid tech report that highlights novel architectural elements like a "Zero-Compute sink expert" and load balancing without an auxiliary loss.

🌏 Asia spotlight

Xiaohongshu (RED) unveiled its first social LLM: RedOne designed to balance social-context understanding with platform governance rules while delivering deep insights into user intent.

Baidu AI Cloud introduced seven digital employees, each with clearly defined business skills, spanning core roles such as recruiting consultant, marketing manager, collections assistant, course advisor, promotions specialist, automotive telesales, and product manager.

ByteDance open-sourced M3-Agent, a multimodal agent with long-term memory that processes audio and visual inputs for human-like perception.

China’s State Council officially released the “Opinions on Deepening the Implementation of the ‘AI+’ Initiative,” setting phased targets and priority tasks through 2027 and 2030, with implementation focused on science and technology, industry, public services, and governance.

China has set the national direction for “AI+,” releasing a ten-year, three-stage roadmap along with a strategic interpretation that clarifies priorities and implementation paths.

Tencent’s Hunyuan open-sourced four lightweight language models. Designed to run on consumer-grade GPUs—including on laptops and even smartphones—they target low-power scenarios such as in-vehicle smart cockpits and smart-home devices, and support low-cost fine-tuning for vertical applications.

An at-a-glance infographic explains Shanghai’s Implementation Plan for the Development of the Embodied Intelligence Industry.