If you’re one of those who eagerly follows the latest developments in the world of artificial intelligence (AI), you’ve probably heard about Elon Musk’s recent interview on Lex Fridman’s podcast. Musk, the founder of xAI, recently discussed concepts like AGI (Artificial General Intelligence) and ASI (Artificial Superintelligence), emphasizing how data is becoming the biggest bottleneck in AI’s progression.
The Current State of AI Development
Since last year, there’s been a growing chorus suggesting that the primary obstacle to AI advancement isn’t computational power anymore, but rather the quality and quantity of data available. In the Lex Fridman podcast, Musk pointed out that the internet is increasingly flooded with content generated by AI, which poses a serious issue.
Elon Musk: War, AI, Aliens, Politics, Physics, Video Games, and Humanity | Lex Fridman Podcast #400
Just think about it: if we train an AI model on data created by another AI, we could end up facing “model collapse.” According to at least two studies published in Nature in 2024, this scenario leads to AI producing outputs that become increasingly homogeneous and flawed, losing the diversity that real-world data provides. Research indicates that this could result in a dead-end: the AI “forgets” rare events and the richness derived from genuine human data.
But there’s hope! It looks like we might be able to tackle this problem. We’re getting smarter about how we train models, developing better algorithms (like reasoning models), and leveraging more computational power for “deeper” models with additional layers. Even with the mixed reception of GPT-5, let’s face it—overclocking the engine isn’t the answer. Fortunately, there are no roadblocks; new algorithms and chips are on the way. A potential solution could involve AI learning about the world through new modalities—meaning new forms of input. For example, through videos, sensory data, or even the senses of robots. With the “Chain-of-Modality” approach, AI learns manipulation programs from human videos, bringing us closer to a true understanding of the world.
Human Intelligence
Children learn in this way too. They grow up, and their knowledge from what they see and experience merges with what they hear and read… leading us to the level of HI (human intelligence).
Intelligence isn’t one-dimensional. According to Howard Gardner’s theory of multiple intelligences, human intelligence can be broken down into eight main types: musical, visual-spatial, linguistic, logical-mathematical, bodily-kinesthetic, interpersonal, intrapersonal, and naturalistic. This framework can help compare AI to human capabilities—AI excels in logical-mathematical skills but lags behind in emotional (interpersonal) intelligence.
Main Types of Machine Intelligence
To understand the concepts of AGI and ASI, it’s helpful to review the spectrum of machine intelligence, categorized in a slightly different but ultimately converging way:
- ANI (Artificial Narrow Intelligence): This is the type of AI we see today, like chess programs or voice recognition systems. These are optimized for specific tasks but lack generality. Deep learning has revolutionized this level, but it’s still limited. They can be abstractly aligned with the dimensions of multiple intelligences.
- AGI (Artificial General Intelligence): Here, AI reaches human-level intelligence across multiple domains. It can learn, adapt, and solve problems without being explicitly programmed. According to Musk, xAI aims to build AGI, and we may soon reach this level with models like OpenAI’s O3 (possibly a successor to O1), Anthropic’s Opus 4, GPT-5, or Grok 4.
- ASI (Artificial Superintelligence): This is an advanced form of AGI, where AI far surpasses human intelligence across all domains. It’s both daunting and exciting: what happens if the “student surpasses the teacher”? Existential dilemmas arise—what will we, as humans, do? Are we mentally prepared for that moment?
- Augmented/Hybrid Intelligence: This involves human-AI collaboration, where humans remain “in the loop.” For instance, Neuralink aims to extend human intelligence (Human Extended Intelligence).
What Does the Future Hold?
Are we ready for AI to surpass us? What will we do if AI gets a physical body (as discussed in our sister blog) and robots take over our jobs? These are existential questions: AI presents risks but also opportunities for the expansion of humanity.
