Today’s News at a Glance
🔬 Tsinghua & Shanghai AI Lab propose GenPRM, small model reasoning may challenge giants
🇪🇺 EU releases “AI Continent Action Plan,” comprehensively planning for global leadership
🤝 Google DeepMind plans to merge Gemini & Veo to create an omnipotent AI assistant
⚙️ iFlytek Spark Agent Platform fully supports MCP, simplifying Agent development
01
🔬 Tsinghua & Shanghai AI Lab Propose GenPRM Small Model Reasoning May Challenge Giants
Tsinghua University and Shanghai AI Lab have proposed a novel AI supervision method, GenPRM, requiring models to first generate detailed Chain-of-Thought (CoT) explanations before evaluating reasoning steps, combined with code execution for verification. This “explain-then-verify” generative supervision mechanism significantly enhances the model’s reasoning ability and its detection of logical errors. Research shows that small models (1.5B/7B parameters) trained using this method, combined with test-time computation expansion (TTS), even surpass GPT-4o and specialized models with far more parameters on complex mathematical reasoning benchmarks. This indicates the potential for optimized small models to challenge large models on specific tasks and opens up a new data-efficient path for improving LLM reasoning capabilities.
Core Highlights
- Innovative Model: GenPRM transforms supervision from simple “scoring” to complex “explanation and verification,” leveraging the model’s generative capabilities for self-improvement.
- Generative Supervision: Models must first generate natural language CoT explanations of their thought process, making evaluation more transparent and interpretable.
- Code Verification: Introduces code generation and execution to verify the correctness of mathematical or logical steps, enhancing supervision reliability.
- Performance Breakthrough: Small models (1.5B/7B) trained with only a small amount of data, combined with TTS, surpass GPT-4o and larger specialized models in mathematical reasoning.
- Data Efficiency: The proposed RPE method combined with code verification can automatically synthesize high-quality supervision data, reducing reliance on manual labeling.
Researcher’s Thoughts
- Practitioner’s Thoughts:
- Improving LLM reasoning shouldn’t just rely on scaling up; sophisticated training, supervision, and optimization techniques (like GenPRM) are equally crucial.
- Optimized small models have huge potential and may replace large models in specific scenarios, lowering the barrier and cost for high-performance AI applications.
- Automated high-quality data synthesis methods (RPE + code verification) offer new ideas for solving the AI data bottleneck.
- GenPRM’s “generate-critique-reflect” mechanism provides a blueprint for achieving stronger model self-learning and self-improvement capabilities.
- Value for Ordinary People:
- AI performance in rigorous fields like mathematics and logic is expected to improve, providing more reliable learning support, professional knowledge, and decision assistance.
- If small model performance improves, advanced AI capabilities will be more easily integrated into everyday devices and services, accelerating AI adoption.
- Enhancing AI reasoning and self-correction capabilities is a key step towards more general and trustworthy AI, helping to build more responsible AI systems.
Recommended Reading
Link Name: AIbase Report: Tsinghua and Shanghai AI Lab Jointly Develop New Process Reward Model
02
🇪🇺 EU Releases “AI Continent Action Plan,” Comprehensively Planning for Global Leadership
The European Commission has unveiled the ambitious “AI Continent Action Plan,” aiming to establish the EU’s leadership in the global AI field. The plan revolves around five pillars: building large-scale AI data and computing infrastructure (deploying 13 AI Factories, planning 5 AI GigaFactories, proposing the ‘Cloud and AI Development Act’), increasing data access, developing advanced algorithms and promoting industrial applications, strengthening talent cultivation, and simplifying regulatory implementation. The EU plans to mobilize up to €200 billion in public and private investment, aiming to comprehensively enhance its competitiveness and strategic autonomy in AI R&D, application, and governance by integrating resources and optimizing policies, while ensuring AI development aligns with EU values.
Core Highlights
- Grand Vision: Aims to transform the EU into a leading global “AI Continent,” mobilizing up to €200 billion in public-private investment to support AI development.
- Infrastructure First: Focuses investment on AI infrastructure, including deploying 13 AI Factories based on supercomputers (approx. €2 billion investment), planning 5 AI GigaFactories with greater computing power (mobilizing €20 billion investment), and proposing legislation to promote cloud and data center construction (aiming to double capacity).
- Data Strategy: Creates an internal data market through the “Data Alliance Strategy,” establishes “Data Labs” in AI Factories to gather high-quality data, and enhances data accessibility.
- Talent Development: Strengthens AI literacy, professional training (AI Skills Academies), simplifies the attraction of high-skilled talent, and incentivizes the return of local talent.
- Regulatory Simplification: While implementing the ‘AI Act,’ simplifies its application, reduces burdens on SMEs, sets up helpdesks for guidance, and launches public consultations on related strategies.
Researcher’s Thoughts
- Practitioner’s Thoughts:
- The EU sends a strong signal supporting the local AI industry, bringing opportunities and policy direction for AI companies in or planning to enter Europe.
- Large-scale infrastructure investment benefits hardware vendors, cloud service providers, data centers, etc., and provides stronger local computing power.
- The data strategy aims to break down silos but needs to balance data sharing, privacy, and commercial secrets in practice.
- The implementation details and simplification measures of the ‘AI Act’ are key focuses, as compliance costs and clarity affect innovation and market access.
- EU actions impact global AI geopolitics and may promote the formation of a European-style AI development model and governance standards.
- Value for Ordinary People:
- The plan aims to enhance economic competitiveness, potentially bringing jobs and growth in the long term, but short-term job impacts and supporting policies need attention.
- Emphasizes “trustworthy and human-centric” AI, relying on the ‘AI Act’ to protect citizens’ rights, privacy, and safety.
- Promotes AI application in key services like healthcare and public administration, potentially improving service efficiency, quality, and accessibility.
- The EU plays an active role in global AI governance, helping to promote more responsible, ethical international norms.
Recommended Reading
Link Name: European Commission: AI Continent Action Plan Official Page
03
🤝 Google DeepMind Plans to Merge Gemini & Veo, Creating an Omnipotent AI Assistant
Google DeepMind CEO Demis Hassabis revealed plans to eventually integrate its multimodal model Gemini with its video generation model Veo. The goal is to enhance Gemini’s understanding of the physical world to create a “universal digital assistant” capable of providing help in the real world. Hassabis noted that the AI industry is moving towards “omnipotent” models that can understand and synthesize multiple media types and hinted that Veo learns physical laws by analyzing massive amounts of YouTube videos. This move suggests that the next generation of AI assistants will possess stronger environmental perception and interaction capabilities.
Core Highlights
- Model Fusion: Combines the strengths of Gemini (language, image, audio) and Veo (video generation and understanding) to create more powerful AI.
- Physical Understanding: The main purpose is to enable Gemini to better understand real-world rules and dynamics by learning physical knowledge extracted by Veo from videos.
- Omnipotent Assistant: The ultimate goal is to create a universal assistant that can provide practical help in the physical world, requiring a deep understanding of the environment; video integration is key.
- Multimodal Trend: Aligns with the industry trend towards “omnipotent” or “any-to-any” models capable of understanding and generating across multiple modalities.
- Data Source: Explicitly states YouTube as the primary source of Veo’s training data, learning physical laws by “watching” videos.
Researcher’s Thoughts
- Practitioner’s Thoughts:
- Multimodal fusion (especially video) is a key frontier for enhancing AI’s ability to understand the real world.
- The value of video data is highlighted; how to efficiently and compliantly utilize large-scale video for model training is a core competitive factor.
- Universal assistants with physical world understanding will open up new applications in robotics, embodied intelligence, AR/VR, etc.
- AI model architectures are evolving towards “omnipotence,” requiring attention to how models processing multiple information streams are designed.
- Value for Ordinary People:
- Future AI assistants will be smarter, better understanding the user’s situation and environment to provide more reality-aligned help.
- Integrating video understanding promises more natural and intuitive human-computer interaction.
- AI will play a greater role in scenarios involving interaction with the physical world, such as autonomous driving, smart homes, and robotic services.
- Using platform data to train AI again raises privacy and copyright discussions; attention must be paid to how tech companies use data.
Recommended Reading
Link Name: TechCrunch Report: Demis Hassabis says Google will eventually merge Gemini and Veo models
04
⚙️ iFlytek Spark Agent Platform Fully Supports MCP, Simplifying Agent Development
iFlytek announced that its Spark Agent development platform now fully supports MCP (Meta-protocol for Capabilities), aiming to help developers easily and efficiently build Agent applications. Developers can conveniently configure and call industry-leading MCP Servers on the platform, or publish custom MCP Servers with one click, achieving “plug-and-play” Agent capabilities. Initially, the platform supports over 20 MCP Servers covering fields like AI capabilities, lifestyle services, and content generation, and allows developers to host self-developed Servers and publish Agentic Workflows as Servers with one click, effectively expanding the boundaries of tool capabilities. iFlytek believes MCP is becoming the de facto standard for AI application middleware and will continue to participate in it, building more refined services and a high-quality tool market.
Core Highlights
- Full Support: Spark platform integrates MCP, simplifying the Agent application development process.
- Convenient Calling: Developers can easily configure and use MCP Servers provided by the platform or custom ones.
- One-Click Publishing: Supports rapid publishing of custom capabilities or Agentic Workflows as MCP Servers.
- Ecosystem Integration: Initially supports 20+ MCP Servers and allows hosting of self-developed Servers, integrating into the mainstream MCP ecosystem.
- Standard Trend: iFlytek recognizes and follows the trend of MCP becoming the de facto standard for AI application middleware, planning continued investment.
Researcher’s Thoughts
- Practitioner’s Thoughts:
- The advancement of MCP standardization lowers the barrier for integrating AI capabilities, allowing developers to more easily combine tools and services from different sources to build complex Agents.
- iFlytek platform’s participation enriches the MCP ecosystem, providing developers with more tool choices and hosting solutions.
- Publishing Agentic Workflows as MCP Servers offers a new way to encapsulate and reuse complex business logic.
- Low-code/no-code combined with MCP support further reduces the difficulty of Agent application development, benefiting rapid innovation and prototype validation.
- Value for Ordinary People:
- The “plug-and-play” nature of Agent capabilities and the lowered development barrier are expected to accelerate the emergence of more intelligent, personalized AI applications, enriching the digital life experience.
- With different service providers adhering to a unified standard (MCP), future AI applications might work together more conveniently, offering more coherent, cross-platform services.
- A thriving ecosystem could ultimately lead to more high-quality, easy-to-use AI tools, benefiting ordinary users.
Recommended Reading
Link Name: MCP One-Click Hosting Entry (iFlytek Star Agent Platform)
Today’s Summary
- Model Technology Continues to Evolve: Tsinghua & Shanghai AI Lab’s GenPRM proves optimized small models can challenge large ones on specific reasoning tasks; Google plans to merge Gemini & Veo, moving towards an “omnipotent” AI assistant with better physical world understanding; xAI opens Grok 3 API, intensifying competition in the large model market.
- Strategic Layout & Global Competition: The EU launches its “AI Continent Action Plan,” investing heavily in infrastructure, talent, and simplified regulation to secure leadership in the global AI race.
Overall, today’s news reflects the rapid development and fierce competition in the AI field across multiple dimensions, including capital investment, technological innovation (especially in Agents and model reasoning), ecosystem building, and national-level strategic planning.
About Us
🚀 Leading enterprise digital transformation, shaping the future of the industry together. We specialize in creating customized digital systems integrated with AI, achieving intelligent upgrades and deep integration of business processes. Relying on a core team with top tech backgrounds from MIT, Microsoft, etc., we help you build powerful AI-driven infrastructure, enhance efficiency, drive innovation, and achieve industry leadership.