• AI KATANA
  • Posts
  • China's DeepSeek releases AI model upgrade

China's DeepSeek releases AI model upgrade

Also: The Generative Beings community is hosting their first AI hackathon event, powered by Nebius

Hello!

Today’s AI digest reveals a global AI race gaining momentum, with major developments across China, India, and the U.S. China’s DeepSeek has rolled out a major upgrade to its language model, intensifying its pursuit of U.S. AI leaders, while Tencent counters with its T1 model, promising faster reasoning and better long-text comprehension. Meanwhile, OpenAI and Meta are eyeing India as their next big market, entering advanced talks with Reliance Industries to integrate AI tools like ChatGPT and host models in what could be the world’s largest data center. Back in the enterprise world, a staggering 88% of AI pilots are failing to reach production due to poor planning, lack of readiness, and mounting executive pressure—an insight that adds nuance to the generative AI hype. Microsoft is tackling a different challenge: securing AI infrastructure. Its latest Security Copilot agents aim to fortify defenses against increasingly complex cyber threats, while also rolling out protections for generative AI applications. Each of these stories reflects a maturing ecosystem grappling with competition, localization, integration, and trust as the AI revolution scales globally.

Sliced just for you:

  • 🤖 China’s DeepSeek releases AI model upgrade, intensifies rivalry with OpenAI

  • 🇮🇳 OpenAI, Meta in talks with Reliance for AI partnerships

  • 📉 88% of AI pilots fail to reach production — but that’s not all on IT

  • 🧠 Tencent launches T1 reasoning model amid growing AI competition in China

  • 🛡️ Microsoft unveils Microsoft Security Copilot agents and new protections for AI

​​The Generative Beings (TGB) community is hosting their first AI hackathon event, powered by Nebius.

🗓️ Friday 4-7 April

​Top AI builders globally can build AI Agents to solve industry problems on Nebius' GPUaaS infra and compete alongside other developers in the AI space.

​You can find more details on the problem spaces on the TGB blog.

Winning prize/benefits

​Nebius Cloud Credits for Top 3 Hackathon Winners:

  • ​1st Place: $10,000 in GPU cloud credits

  • ​2nd Place: $5,000 in GPU cloud credits

  • ​3rd Place: $3,000 in GPU cloud credits

Register here: https://lu.ma/gc7f2w8t

China-based DeepSeek has unveiled a significant upgrade to its large language model, DeepSeek-V3-0324, reinforcing its ambitions to rival leading U.S. players like OpenAI and Anthropic. The model, released via Hugging Face, exhibits marked improvements in reasoning and coding capabilities, as demonstrated by enhanced benchmark results. This latest development follows a rapid series of model launches, with the V3 introduced in December and the R1 in January, signaling DeepSeek’s fast-paced growth and cost-efficient alternatives in the global AI race. The upgrade further narrows the competitive gap between Chinese and U.S. AI advancements.

OpenAI and Meta are reportedly in separate discussions with India’s Reliance Industries to explore AI partnerships aimed at deepening their presence in the country. Talks include possibilities such as integrating ChatGPT with Reliance Jio, reducing subscription pricing, and selling OpenAI’s models to enterprise clients via APIs. Reliance is also considering hosting OpenAI and Meta models within a planned three-gigawatt data center in Jamnagar, Gujarat, which is expected to be the world’s largest. These discussions reflect Reliance’s ambition to become a central player in India’s AI infrastructure while enabling global tech giants to localize operations and comply with data sovereignty requirements.

Despite AI being a top priority across the enterprise landscape, a staggering 88% of AI pilots still fail to make it into production, according to new IDC research. Key contributors to this high attrition rate include unclear objectives, poor data readiness, a lack of in-house expertise, and hasty approval of pilot projects under executive pressure. Generative AI pilots, in particular, are being greenlit with minimal scrutiny as CEOs and boards push aggressively for experimentation, often without solid business cases or sufficient funding. While some view these failures as valuable learning experiences, others warn that corporate urgency and inflated expectations are distorting ROI calculations and driving unsustainable practices. The disconnect between strategic ambition and operational readiness continues to be a major hurdle for organizations seeking real AI impact.

Tencent has officially launched its T1 reasoning model, intensifying the competition in China’s rapidly expanding AI sector. The T1 model, now powered by Tencent’s Turbo S foundational language model, boasts faster response times and improved handling of long-form text, with a particularly low hallucination rate. Early benchmarks suggest it outperforms DeepSeek’s R1 on certain reasoning and knowledge tasks. This move follows Tencent’s broader AI investment surge and reflects the company’s ambition to assert dominance as rivals like DeepSeek gain traction with cost-efficient, high-performing models. The T1 rollout signals a continued escalation in China’s race to develop advanced AI infrastructure.

Microsoft has expanded its AI-powered Security Copilot platform with a suite of new autonomous agents designed to combat escalating cyber threats, including phishing, insider risk, and identity management. Six new in-house agents and five developed by partners will begin rolling out in April 2025, each tailored to automate and enhance specific security functions while aligning with Microsoft’s Zero Trust framework. Alongside these, the company introduced AI-powered data security investigations, new protections for generative AI applications, and capabilities to govern risky access and prevent sensitive data leaks into shadow AI tools. With multicloud and multimodel support now extended across platforms like Google VertexAI and Meta Llama, Microsoft Defender aims to provide full-spectrum AI security posture management. These advances reflect Microsoft’s commitment to fusing AI with cybersecurity to help organizations scale defenses and adapt to the evolving threat landscape.

🛠️ AI tools updates

Adobe has launched a new wave of AI agents aimed at transforming online marketing tools, enabling brands to deliver personalized experiences and streamline website management. These agents help websites tailor content based on user activity—such as differentiating between traffic from TikTok ads versus search queries—and enhance interactions through chatbot-like experiences that respond intelligently using context from user behavior. Adobe also introduced AI-powered tools that allow marketers to set goals and receive recommendations for web updates, which the agents can implement directly, drastically reducing timelines for digital optimization. This move underscores Adobe’s commitment to accelerating digital marketing innovation through AI integration.

💵 Venture Capital updates

Silverstream AI has emerged from stealth with $1.2 million in pre-seed funding led by Google’s Gradient Ventures, aiming to accelerate the development of autonomous web agents. Tackling a core challenge in the space—insufficient high-quality training data—Silverstream is building open-source infrastructure and tools to help developers create reliable, safe, and effective agents. Its platform converts web pages into markdown formats optimized for large language models, integrates privacy controls and permission systems, and includes a decision-making engine that enables agents to perform complex browser-based tasks. Central to its mission is the forthcoming release of Pasta-1T, a freely available dataset of web trajectories intended to democratize access and fuel innovation. Silverstream promotes open collaboration and high-precision agent deployment, with an emphasis on incremental rollouts and behavioral learning, positioning itself as a foundational player in the agentic AI ecosystem.

🫡 Meme of the day

⭐️ Generative AI image of the day