From Knowledge Query to Thought Forge: The Second Generation of Generative AI

The rapid development of generative AI has made impressive progress in recent years. The first generation, which we can retrospectively call "Act I" (2020-2023), was based on scaling parameters and data volumes. These models achieved astonishing performance, but at the same time encountered fundamental limitations. Problems such as knowledge actuality, superficial reasoning, and limited cognitive processes persisted. Prompt engineering developed into the primary interface during this phase for interacting with AI via natural language.

We are currently experiencing the beginning of "Act II" (2024-present). A transformation is taking place here: models are evolving from pure knowledge query systems (in latent space) to complex thought forges. This is made possible by so-called test-time scaling techniques. This new paradigm allows a connection with AI on an almost thought-based level, mediated by language-based thoughts.

Cognition Engineering: A New Field of AI Research

At the center of this development is cognition engineering. This approach aims to replicate and optimize cognitive processes in AI models. Test-time scaling techniques play a crucial role in this. They allow the capabilities of the models to be dynamically adapted and expanded during application, rather than being limited to a fixed set of parameters. This allows for the simulation of more complex thought processes and enables more flexible interactions.

The significance of this moment for the development of cognition engineering is immense. Through the new possibilities of test-time scaling, researchers and developers can explore the limits of what is feasible in AI development. Interaction with AI becomes more intuitive and efficient, opening up entirely new fields of application.

Democratizing Access to Cognition Engineering

To promote the further development of cognition engineering, it is crucial to democratize access to the necessary tools and resources. Initiatives like the GitHub repository "cognition-engineering" (https://github.com/GAIR-NLP/cognition-engineering) offer a central platform for researchers and developers. Here you can find comprehensive tutorials, optimized implementations, and a regularly updated collection of relevant publications. This allows anyone interested to actively participate in shaping "Act II" of generative AI and help shape the future of this technology.

Mindverse, as a German provider of all-in-one solutions for AI text, content, images, and research, is observing these developments with great interest. The company develops customized solutions such as chatbots, voicebots, AI search engines, and knowledge systems. The advances in cognition engineering also open up new possibilities for Mindverse to develop innovative and powerful AI solutions for the future.

Bibliography: - https://arxiv.org/abs/2504.13828 - https://github.com/GAIR-NLP/cognition-engineering - https://paperreading.club/page?id=300517 - https://arxiv.org/html/2504.13828v1 - https://chatpaper.com/chatpaper/?id=3&date=1745164800&page=1 - https://chatpaper.com/chatpaper/?id=2&date=1745164800&page=1 - https://www.sciencedirect.com/science/article/pii/S0268401223000233 - https://nips.cc/virtual/2024/papers.html - https://www.aisnakeoil.com/p/ai-as-normal-technology