Zhipu AI Chief Scientist Tang Jie: Making Machines Think Like Humans
Summary
Zhipu AI's Chief Scientist Tang Jie discussed the company's journey and vision for AI, following the market impact of their GLM-5.2 large language model. GLM-5.2, an iterative refinement, reportedly matches Anthropic's Claude Opus 4.7 in coding at half the cost, significantly boosting Zhipu AI's market valuation. Tang Jie outlined the evolution of LLM intelligence from simple QA to complex reasoning and real-world programming, emphasizing a shift from "chat" to "doing" tasks by integrating thinking, agentic, and coding capabilities. He highlighted challenges in real-world environments, leading to innovations like the asynchronous reinforcement learning framework for GLM-4.7 and the open-sourced 9B AutoGLM model for device interaction. Tang Jie also acknowledged that despite China's growing open-source contributions, a gap with closed-source US models persists, advocating for future focus on multimodal adaptation, memory, continual learning, and self-awareness to achieve human-like cognition.
Key takeaway
For AI Engineers developing agentic systems, recognize that the "chat era" is maturing, and the next frontier involves enabling AI to "do" complex, multi-step tasks across diverse environments. You should prioritize integrating thinking, coding, and agentic capabilities, leveraging asynchronous reinforcement learning and hybrid API/GUI approaches to overcome real-world challenges like cold-start scenarios and ultra-long tasks. This shift demands focusing on robust, verifiable environments and exploring new architectures beyond simple scaling to achieve true generalization and device embodiment.
Key insights
The next AI paradigm shifts from chat to "doing" tasks, requiring integrated thinking, agentic, and coding capabilities.
Principles
- Long-term commitment is crucial for AGI development.
- Scaling data and compute alone is insufficient for advanced generalization.
- Human cognitive processes offer a blueprint for AI development.
Method
Zhipu AI developed a hybrid API/GUI approach for device interaction, integrating SFT and reinforcement learning with an asynchronous training framework to improve agent capabilities in cold-start scenarios.
In practice
- Explore RLVR for self-exploring AI feedback data.
- Implement hybrid API/GUI for robust agentic device control.
Topics
- Zhipu AI
- GLM-5.2
- Large Language Models
- AI Agents
- Reinforcement Learning
- Multimodal AI
- Knowledge Graphs
Best for: CTO, VP of Engineering/Data, Machine Learning Engineer, AI Scientist, Director of AI/ML, AI Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by LLM on Medium.