A Brief Discussion on AGI Definition and Classification

🌐 中文

English

This article aims to provide an explanation of AGI definition and classification.

I’ve wanted to talk about AGI for a long time, and now I finally have the time to discuss AGI.

What is AGI

Words are unnecessary; let’s first look at a picture: According to the definition, AGI refers to AI that can achieve human-level performance on any task. Specifically, it includes: Language understanding and generation, Learning and reasoning, Perception and understanding, Emotional and social abilities, Creativity, etc.

Classification of AGI

Facing the current AGI, openAI and google’s deepmind have provided different answers:

OpenAI:

First, let’s talk about openAI’s level classification: From the image, it can be seen that openAI divides AGI into five levels:

• level 1: AI with natural language dialogue capabilities, able to conduct basic conversations
• level 2: AI with reasoning abilities similar to human levels, capable of solving problems across a wide range of domains
• level 3: AI systems capable of taking actions based on human instructions or independently (using agents)
• level 4: AI systems capable of helping invent new ideas and contributing to human knowledge
• level 5: AI systems capable of independently completing all the work of an entire organization

It can be seen that currently, openAI’s differentiation of AGI levels is extremely vague, providing almost no accurate judgment criteria, so we mainly look towards deepmind.

deepmind:

It is clearly visible that compared to openAI’s level classification, deepmind’s level classification is clearer:

• Level 0: No AI: At this stage, there is no AI, only simple programs, such as calculators or compilers.
• Level 1: Emerging: At this stage, AI is stronger than non-practitioners. Currently, all AI belongs to this level.
• Level 2: Competent: In certain narrow domains, AI performs better than 50% of practitioners.
• Level 3: Expert: AI demonstrates abilities exceeding 90% of practitioners.
• Level 4: Virtuoso: AI demonstrates extremely high skills, surpassing 99% of practitioners.
• Level 5: Superhuman: ASI (Artificial Super Intelligence), at this stage, AI capabilities exceed 100% of all humans.

o1-preview Level

As the model o1-preview (previously codenamed Strawberry) released by openAI on 2024.9.12, openAI publicly claims it uses a new scaling law From the image, it can be seen that the new scaling law generated by openAI refers to using the Chain of Thought method combined with agent. However, since its use of agent is still in its infancy, it is currently between level 3 and level 2 in openAI’s classification, but still belongs to the level 2 stage. In deepmind’s classification, through testing, it still cannot exceed 50% of practitioners and remains at level 1 in deepmind.

Conclusion

AGI is a complex concept involving multiple levels of technology and capabilities. Currently, we are still in the exploration stage of AGI and cannot fully achieve the goal of AGI. Therefore, we need not overly worry about the impact of AGI on us.
The above are some simple views of mine on AGI definition and classification. Although not deeply discussed, I still hope it is helpful to you.