The terminologies surrounding the industry are expanding rapidly as the generation AI enters the product, financing and development landscape. TechCrunch has updated an AI terminology guide, which focuses on the most common sets of concepts at present and helps readers understand model capabilities, training methods and application patterns.

This guide does not confine itself to chat robots, but extends to model training, reasoning, computing infrastructure and automation tools. It was mentioned that, although many terms are widely used, there is still no fully harmonized definition within the industry.

AGI still lacks a uniform definition

In this comb, AGI remains one of the most divisive words. It is usually used to refer to AI systems that are close to or above the human average for most tasks, but the standards given by different agencies are not consistent.

For example, institutions such as OpenAI and Google DeepMind each focus on AGI. Some emphasis was placed on replacing human labour in most economic activities, others on reaching human levels in most cognitive tasks. According to TechCrunch, this difference in itself suggests that AGI is now more like an industry target than a clear technical definition of boundaries.

AI-Agent and Encoding-Agent-I Wen.

The article defines AI proxy as a system that can continue to perform multi-step missions on behalf of users, with capabilities beyond normal chat robots. Typical scenarios include claims, booking, calling for third-party services, or even partial software development with less manual intervention.

On this basis, coding agents are considered a more specific type of tool. It not only generates codes, but also carries out testing, quelling errors in the code library and submits them for repair. The article described such tools as automated assistants for development processes, but noted that manual review remained indispensable.

Thinking chains, fine tunes and distillation are more common.

At the modelling capacity level, TechCrunch focused on explaining the concept of the “think chain”. At its core, it breaks complex issues into intermediate steps to improve the accuracy of logical reasoning or programming tasks. This approach is often slower than the direct answer, but results are often more stable.

The article also mentioned that fine-tuning and distillation has become a common method for AI companies. The fine-tuning is based on the continuing training model for specific data to make it more appropriate for a particular industry or mission; distillation allows smaller models to learn the output behaviour of larger models and thus retain some of their capabilities while reducing costs.

  • Align: make the generic model more appropriate for vertical scenarios
  • Distillation: Train lighter models with large model output
  • Crucial chain of thinking: improving the quality of reasoning through intermediate steps

The algorithm and the generation model are still fundamental.

In addition to the applied layers of terminology, the article goes back to the bottom structure of AI. The so-called compute refers to the computing resources that support model training and deployment, which correspond to hardware infrastructure such as GPU, CPU, TPU etc. Calculus remains one of the core resources for the expansion of AI.

In the generation of models, the guidelines refer to such concepts as in-depth learning, diffusion models and PAN. In-depth learning relies on a multilayered neural network, extracting features from a large amount of data; diffusion models generate images, music or text by learning the "noise" process; and GAN, an earlier generation framework, has contributed to the development of veritable content generation and deep forgery tools.

Overall, the glossary does not attempt to give a uniform standard, but rather provides a set of “common language” that is most common in the current AI industry. The meaning of these terms may also continue to change as models, products and infrastructure continue to evolve.