AiPulse — Issue No. 40

More than 100 companies, including OpenAI, Anthropic, and Google, have called for public and private sectors to collaborate on establishing new defense standards, warning of the proliferation of AI-enabled cyberattacks.

IBM has released the Granite 4.2 models, which are licensed under Apache 2.0 and feature agentic capabilities, enabling companies to host them on their own infrastructure without sending their data externally.

Arduino has opened pre-orders for the VENTUNO Q platform, powered by a Qualcomm processor, enabling "Physical AI" applications that combine local large language models and real-time control without relying on the cloud.

The government announced that it has allocated 2.3 trillion won to the humanoid robot ecosystem and 2.8 trillion won to physical artificial intelligence applications by 2030, increasing public procurement and research and development support to ensure domestic 'full-stack' technological independence and reduce dependence on Chinese-made products.

Tencent's WeChat team has released the WeMM-Embedding models, which align text, images, and video, as open source, including code and weights, so that developers can use them in multimodal search and recommendation applications.

TrueFoundry's integration of Alibaba's Qwen model family into its open-source agent framework TrueForge provides access to all model sizes through a single API, reducing operating costs by approximately 50% compared to single-vendor platforms.

This new speech recognition model announced by Google supports more than 85 languages, filters background noise and hesitations, distinguishes up to three different speakers, and reduces transcription time by up to 70% compared to the previous generation, offering faster and more accurate speech-to-text conversion for both developers and general users.

DeepSeek's achievement of reaching 970 billion won by surpassing last year's total revenue tenfold in the first seven months of this year demonstrates that the company is growing rapidly thanks to its cost efficiency and is strengthening its potential initial public offering process with a new 10 trillion won investment round.

What is Prompt Tuning?

Prompt tuning is an energy-efficient method for steering large language models toward narrow tasks with limited data.